Methods for Neural Ensemble Recordings,

METHODS for NEURAL ENSEMBLE RECORDINGS SECOND EDITION © 2008 by Taylor & Francis Group, LLC FRONTIERS IN NEUROSCIENC...

Author: Miguel A. L. Nicolelis

139 downloads 1079 Views 15MB Size Report

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form

DOWNLOAD PDF

METHODS for NEURAL ENSEMBLE RECORDINGS SECOND EDITION

© 2008 by Taylor & Francis Group, LLC

FRONTIERS IN NEUROSCIENCE Series Editors Sidney A. Simon, Ph.D. Miguel A.L. Nicolelis, M.D., Ph.D.

Published Titles Apoptosis in Neurobiology Yusuf A. Hannun, M.D., Professor of Biomedical Research and Chairman/Department of Biochemistry and Molecular Biology, Medical University of South Carolina, Charleston, South Carolina Rose-Mary Boustany, M.D., tenured Associate Professor of Pediatrics and Neurobiology, Duke University Medical Center, Durham, North Carolina Methods for Neural Ensemble Recordings Miguel A.L. Nicolelis, M.D., Ph.D., Professor of Neurobiology and Biomedical Engineering, Duke University Medical Center, Durham, North Carolina Methods of Behavioral Analysis in Neuroscience Jerry J. Buccafusco, Ph.D., Alzheimer’s Research Center, Professor of Pharmacology and Toxicology, Professor of Psychiatry and Health Behavior, Medical College of Georgia, Augusta, Georgia Neural Prostheses for Restoration of Sensory and Motor Function John K. Chapin, Ph.D., Professor of Physiology and Pharmacology, State University of New York Health Science Center, Brooklyn, New York Karen A. Moxon, Ph.D., Assistant Professor/School of Biomedical Engineering, Science, and Health Systems, Drexel University, Philadelphia, Pennsylvania Computational Neuroscience: Realistic Modeling for Experimentalists Eric DeSchutter, M.D., Ph.D., Professor/Department of Medicine, University of Antwerp, Antwerp, Belgium Methods in Pain Research Lawrence Kruger, Ph.D., Professor of Neurobiology (Emeritus), UCLA School of Medicine and Brain Research Institute, Los Angeles, California Motor Neurobiology of the Spinal Cord Timothy C. Cope, Ph.D., Professor of Physiology, Wright State University, Dayton, Ohio Nicotinic Receptors in the Nervous System Edward D. Levin, Ph.D., Associate Professor/Department of Psychiatry and Pharmacology and Molecular Cancer Biology and Department of Psychiatry and Behavioral Sciences, Duke University School of Medicine, Durham, North Carolina Methods in Genomic Neuroscience Helmin R. Chin, Ph.D., Genetics Research Branch, NIMH, NIH, Bethesda, Maryland Steven O. Moldin, Ph.D, University of Southern California, Washington, D.C.


Methods in Chemosensory Research Sidney A. Simon, Ph.D., Professor of Neurobiology, Biomedical Engineering, and Anesthesiology, Duke University, Durham, North Carolina Miguel A.L. Nicolelis, M.D., Ph.D., Professor of Neurobiology and Biomedical Engineering, Duke University, Durham, North Carolina The Somatosensory System: Deciphering the Brain’s Own Body Image Randall J. Nelson, Ph.D., Professor of Anatomy and Neurobiology, University of Tennessee Health Sciences Center, Memphis, Tennessee The Superior Colliculus: New Approaches for Studying Sensorimotor Integration William C. Hall, Ph.D., Department of Neuroscience, Duke University, Durham, North Carolina Adonis Moschovakis, Ph.D., Department of Basic Sciences, University of Crete, Heraklion, Greece New Concepts in Cerebral Ischemia Rick C. S. Lin, Ph.D., Professor of Anatomy, University of Mississippi Medical Center, Jackson, Mississippi DNA Arrays: Technologies and Experimental Strategies Elena Grigorenko, Ph.D., Technology Development Group, Millennium Pharmaceuticals, Cambridge, Massachusetts Methods for Alcohol-Related Neuroscience Research Yuan Liu, Ph.D., National Institute of Neurological Disorders and Stroke, National Institutes of Health, Bethesda, Maryland David M. Lovinger, Ph.D., Laboratory of Integrative Neuroscience, NIAAA, Nashville, Tennessee In Vivo Optical Imaging of Brain Function Ron Frostig, Ph.D., Associate Professor/Department of Psychobiology, University of California, Irvine, California Primate Audition: Behavior and Neurobiology Asif A. Ghazanfar, Ph.D., Princeton University, Princeton, New Jersey Methods in Drug Abuse Research: Cellular and Circuit Level Analyses Dr. Barry D. Waterhouse, Ph.D., MCP-Hahnemann University, Philadelphia, Pennsylvania Functional and Neural Mechanisms of Interval Timing Warren H. Meck, Ph.D., Professor of Psychology, Duke University, Durham, North Carolina Biomedical Imaging in Experimental Neuroscience Nick Van Bruggen, Ph.D., Department of Neuroscience Genentech, Inc. Timothy P.L. Roberts, Ph.D., Associate Professor, University of Toronto, Canada The Primate Visual System John H. Kaas, Department of Psychology, Vanderbilt University Christine Collins, Department of Psychology, Vanderbilt University, Nashville, Tennessee Neurosteroid Effects in the Central Nervous System Sheryl S. Smith, Ph.D., Department of Physiology, SUNY Health Science Center, Brooklyn, New York


Modern Neurosurgery: Clinical Translation of Neuroscience Advances Dennis A. Turner, Department of Surgery, Division of Neurosurgery, Duke University Medical Center, Durham, North Carolina Sleep: Circuits and Functions Pierre-Hervé Luoou, Université Claude Bernard Lyon, France Methods in Insect Sensory Neuroscience Thomas A. Christensen, Arizona Research Laboratories, Division of Neurobiology, University of Arizona, Tuscon, Arizona Motor Cortex in Voluntary Movements Alexa Riehle, INCM-CNRS, Marseille, France Eilon Vaadia, The Hebrew University, Jerusalem, Israel Neural Plasticity in Adult Somatic Sensory-Motor Systems Ford F. Ebner, Vanderbilt University, Nashville, Tennessee Advances in Vagal Afferent Neurobiology Bradley J. Undem, Johns Hopkins Asthma Center, Baltimore, Maryland Daniel Weinreich, University of Maryland, Baltimore, Maryland The Dynamic Synapse: Molecular Methods in Ionotropic Receptor Biology Josef T. Kittler, University College, London, England Stephen J. Moss, University College, London, England Animal Models of Cognitive Impairment Edward D. Levin, Duke University Medical Center, Durham, North Carolina Jerry J. Buccafusco, Medical College of Georgia, Augusta, Georgia The Role of the Nucleus of the Solitary Tract in Gustatory Processing Robert M. Bradley, University of Michigan, Ann Arbor, Michigan Brain Aging: Models, Methods, and Mechanisms David R. Riddle, Wake Forest University, Winston Salem, North Carolina Neural Plasticity and Memory: From Genes to Brain Imaging Frederico Bermudez-Rattoni, National University of Mexico, Mexico City, Mexico Serotonin Receptors in Neurobiology Amitabha Chattopadhyay, Center for Cellular and Molecular Biology, Hyderabad, India Methods for Neural Ensemble Recordings, Second Edition Miguel A.L. Nicolelis, M.D., Ph.D., Professor of Neurobiology and Biomedical Engineering, Duke University Medical Center, Durham, North Carolina


METHODS for NEURAL ENSEMBLE RECORDINGS SECOND EDITION

Edited by

Miguel A. L. Nicolelis Duke University Medical Center Durham, NC

Boca Raton London New York

CRC Press is an imprint of the Taylor & Francis Group, an informa business


CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2008 by Taylor & Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Printed in the United States of America on acid-free paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number-13: 978-0-8493-7046-5 (Hardcover) This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. No part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www. copyright.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC) 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Library of Congress Cataloging-in-Publication Data Methods for neural ensemble recordings / [edited by] Miguel A.L. Nicolelis. -- 2nd ed. p. ; cm. -- (Frontiers in neuroscience) Includes bibliographical references and index. ISBN 978-0-8493-7046-5 (alk. paper) 1. Electroencephalography. 2. Microelectrodes. 3. Neurons. I. Nicolelis, Miguel A. L. II. Series: Frontiers in neuroscience (Boca Raton, Fla.) [DNLM: 1. Neurons--physiology. 2. Brain--physiology. 3. Electrophysiology--methods. 4. Microelectrodes. WL 102.5 M592 2008] QP376.5.M47 2008 616.8’047547--dc22 Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com


2007027468

To Pedro, Rafael, and Daniel, May your voyages be full of lore and fun


Contents Series Preface............................................................................................................xi Preface.................................................................................................................... xiii Editor .....................................................................................................................xvii Contributors ............................................................................................................xix Chapter 1

State-of-the-Art Microwire Array Design for Chronic Neural Recordings in Behaving Animals ........................................................ 1 Gary Lehew and Miguel A. L. Nicolelis

Chapter 2

Surgical Techniques for Chronic Implantation of Microwire Arrays in Rodents and Primates......................................................... 21 Laura M. O. Oliveira and Dragan Dimitrov

Chapter 3

Technology for Multielectrode MicroStimulation of Brain Tissue.................................................................................................. 47 Timothy Hanson, Nathan Fitzsimmons, and Joseph E. O’Doherty

Chapter 4

Strategies for Neural Ensemble Data Analysis for Brain–Machine Interface (BMI) Applications............................................................. 57 Miriam Zacksenhouse and Simona Nemets

Chapter 5

Chronic Recordings in Transgenic Mice............................................ 83 Kafui Dzirasa

Chapter 6

Multielectrode Recordings in the Somatosensory System.................97 Michael Wiest, Eric Thomson, and Jim Meloy

Chapter 7

Chronic Recording During Learning............................................... 125 Aaron J. Sandler

Chapter 8

Defining Global Brain States Using Multielectrode Field Potential Recordings ........................................................................ 145 Shih-Chieh Lin and Damien Gervasoni

ix © 2008 by Taylor & Francis Group, LLC

x

Chapter 9

Methods for Neural Ensemble Recordings, Second Edition

Multielectrode Recording in Behaving Monkeys ............................ 169 R. E. Crist and M. A. Lebedev

Chapter 10 Neural Ensemble Recordings from Central Gustatory-Reward Pathways in Awake and Behaving Animals..................................... 189 Albino J. Oliveira-Maia, Sidney A. Simon, and Miguel A. L. Nicolelis Chapter 11 Building Brain–Machine Interfaces to Restore Neurological Functions .......................................................................................... 219 Mikhail A. Lebedev, Roy E. Crist, and Miguel A. L. Nicolelis Chapter 12 Conceptual and Technical Approaches to Human Neural Ensemble Recordings ....................................................................... 241 Dennis A. Turner, Parag G. Patil, and Miguel A.L. Nicolelis


Series Preface Our goal in creating the Frontiers in Neuroscience Series is to present the insights of experts on emerging fields and theoretical concepts that are, or will be, in the vanguard of neuroscience. Books in the series cover genetics, ion channels, apoptosis, electrodes, neural ensemble recordings in behaving animals, and even robotics. The series also covers new and exciting multidisciplinary areas of brain research, such as computational neuroscience and neuroengineering, and describes breakthroughs in classical fields like behavioral neuroscience. We hope every neuroscientist will use these books in order to get acquainted with new ideas and frontiers in brain research. These books can be given to graduate students and postdoctoral fellows when they are looking for guidance to start a new line of research. Each book is edited by an expert and consists of chapters written by the leaders in a particular field. Books are richly illustrated and contain comprehensive bibliographies. Chapters provide substantial background material relevant to the particular subject. We hope that as the volumes become available, the effort put in by us, the publisher, the book editors, and individual authors will contribute to the further development of brain research. The extent to which we achieve this goal will be determined by the utility of these books. Sidney A. Simon, Ph.D. Miguel A.L. Nicolelis, M.D.,Ph.D. Series Editors

xi © 2008 by Taylor & Francis Group, LLC

Preface Almost 10 years ago, as the fresh-off-the-press volumes of the first edition of Methods for Neural Ensemble Recordings reached the CRC booth just in time for the inaugural day of the Society for Neuroscience Meeting, there were already several laboratories around the world applying new approaches and technologies to chronically record the simultaneous extracellular activity of small populations of single neurons in behaving animals. Yet, in those days a considerable number of neurophysiologists were still not convinced that such a technique would bring significant benefits to the field. Those opinions were bluntly articulated in two peculiar encounters that I experienced during that time. In the first, one of the leading neurophysiologists of our time stopped by my SFN poster display to, according to him, kindly remind me that there was “no future and no career in neural ensemble recordings.” Instead, he suggested that I should simply give up this “foolish stuff” and come back to the “church of single unit recording.” Just a couple of years later, this grandfatherly and well-intended advice was pointedly reinforced by a NIH reviewer who, in response to a grant that John Chapin and I had submitted, wondered in despair “…why one needs to use space-age technology to study the brain?…” What a difference 10 years make! Today, a whole generation of young graduate students and postdocs dive into the great adventure of systems neurophysiology by learning how to record from tens or even hundreds of single neurons simultaneously as soon as they enter a lab. Without a doubt, this young and innovative branch of neuroscience has been totally embraced by the next generation of neuroscientists. Those closely following the development of the field have also seen that, whereas a decade ago, most if not all studies employing chronic multi-electrode recording were performed only on rats, today the same basic technique is employed for studies in different species of non-human primates, such as owls, squirrels and Rhesus monkeys, as well as mice (wild and transgenic strains). In the last few years, the method has found its way to neurosurgical suites where it is now used during intra-operative recordings in Parkinsonian patients whose symptoms can only be improved through a deep brain stimulator. If these dramatic changes were not enough, multiple technological breakthroughs in the fabrication of high-density microelectrode arrays, surgical implantation techniques, and multi-channel signal processing have significantly expanded the yield and longevity of chronic neural ensemble recordings in behaving animals. These advances have allowed the establishment of direct real-time brain–machine interfaces, a new experimental paradigm that holds a lot of promise, not only in terms of basic neurophysiology research, but also as the potential core for the development of a new generation of neuroprosthetic devices aimed at restoring mobility and communication skills in severely disabled patients. xiii © 2008 by Taylor & Francis Group, LLC

xiv


Over the last 10 years, thanks to the generosity of some of the field’s pioneers who agreed to write a series of detailed methods-oriented chapters for the book, the first edition of Methods for Neural Ensemble Recordings has become a technical reference book for those who have taken the plunge into the field. As such, it has been a most rewarding experience for those of us who participated in the original volume to spot copies of the book, not nicely stored on book shelves, but worn and torn next to neurophysiological setups, spread all over the world. The second edition of this book comes to light during a very different time and environment in systems neuroscience. Partially liberated from having to maintain at any cost some of its classic dogmas, systems neurophysiology is once again thriving in both theory and practice. Furthermore, the ever growing consensus that distributed populations of neurons, rather than single neurons, define the true functional unit of the central nervous system creates continuous demand for new technologies that can push multi-electrode recordings methods and neural ensemble physiology to new limits. Since 1994, when I joined the Department of Neurobiology at Duke, I have had the privilege to work and collaborate with a unique group of highly talented people who arrived at our laboratory with the goal of achieving just that: pushing the limits of our blossoming field and, during this process, shedding new light on the neurophysiological principles that make ensembles of neurons perform the “business of the brain.” This book has been written by a group of these young collaborators, whose collective groundbreaking work covers most of the current areas of basic and clinical research that utilize multi-electrode recordings as the method of choice to probe brain circuits. The final product of this two year project, the second edition of Methods in Neural Ensemble Physiology, is dedicated to all the high-school, undergraduate and graduate students, along with the postdoctoral fellows, technicians, research associates, collaborators, and visitors who, by joining our laboratory, made this journey through distributed and dynamic brain circuits quite a thrill. I would like to thank Duke University, the Brain and Mind Institute at the Ecole Polytechnic Federale de Lausanne, the Edmond and Lily Safra International Institute of Neuroscience of Natal, the National Institutes of Health, DARPA, and the Anne W. Deane Professorship Endowment for providing me with the time, space, support and resources for completing this and other projects that have considerably expanded the horizons of my life as a scientist in the last two years. I would like to sincerely thank our CRC editor, Barbara Norwitz, for waiting patiently for the completion of this project. I would also like to acknowledge the continuous support of my “older brother” and series co-editor, Sidney Simon, who invited me to edit the first edition of this book, 13 years ago, as my first assignment as a Duke assistant professor. We never stopped getting into trouble together after that. I also would like to thank my dear friend and collaborator, Susan Halkiotis, for dedicating 2 long years of her life to this project, and many more years to make sure that everything that crosses her desk, no matter how crazy or challenging the task, is taken care of with utmost care, perfection, and class. With fun on top!


Preface

xv

Finally, I would like to thank Laura Oliveira for thirty years of trusting and supporting the pursuit of my dreams, far from home, wherever they may take us, no matter the odds or personal costs. I can barely wait to see what the next ten years will bring! Miguel A. L. Nicolelis Natal, Brazil


Editor Miguel A. L. Nicolelis, M.D., Ph.D., is the Anne W. Deane Professor of Neuroscience and professor in the departments of neurobiology, biomedical engineering, and psychology at Duke University, where he also serves as co-director of the Center for Neuroengineering. In addition, Dr. Nicolelis is scientific coordinator at the Edmond and Lily Safra International Institute of Neuroscience of Natal (ELS-IINN) in Natal, Brazil and taught neuroscience during 2006–2007 at École Polytechique Féderale de Lausanne in Lausanne, Switzerland. Dr. Nicolelis is a native of Sao Paulo, Brazil, where he received his M.D. and Ph.D. in neurophysiology from the University of Sao Paulo. He graduated from the University of Sao Paulo School of Medicine and was awarded the Oswaldo Cruz Prize for research, the highest honor awarded to a Brazilian medical student. After postdoctoral work at Hahnemann University, Dr. Nicolelis joined the faculty at Duke University in 1994. Dr. Nicolelis is interested in understanding the general computational principles underlying the dynamic interactions between populations of cortical and subcortical neurons involved in motor control and tactile perception. Although Dr. Nicolelis is best known for his study of brain–machine interfaces (BMI) for neuroprosthetics in human patients and nonhuman primates, he is also developing an integrative approach to studying neurological and psychiatric disorders by recording neuronal ensemble activity across different brain areas in genetically modified mice. Dr. Nicolelis believes that this approach will allow the integration of molecular, cellular, systems, and behavioral data in the same animal to produce a more complete understanding of the nature of the alterations associated with these disorders. Neuroscience laboratories in the U.S. and Europe have incorporated Dr. Nicolelis’ experimental paradigm to study a variety of mammalian neuronal systems. His research has influenced basic and applied research in computer science, robotics, and biomedical engineering. This multidisciplinary approach to research has become widely recognized in the neuroscience community. Dr. Nicolelis was named as one of Scientific American’s Top 50 Technology Leaders in America in 2004 and has received a number of honors and awards including the Whitehead Scholar Award, DARPA Award for Sustained Excellence by a Performer; Ruth and A. Morris Williams, Jr., Faculty Research Prize; Whitehall Foundation Award; McDonnell-Pew Foundation Award; Duke University Thomas Langford Lectureship Award; the Ramon y Cajal Chair at the University of Mexico; and the Santiago Grisolia Chair at Catedra Santiago Grisolia. He has authored more than 120 manuscripts in scientific journals, and edited numerous books and special journal issues. He is frequently an invited speaker at scientific conferences and meetings throughout the world.

xvii © 2008 by Taylor & Francis Group, LLC

Contributors Dr. Roy Crist Department of Neurobiology Duke University Durham, North Carolina

Mr. Jim Meloy Department of Neurobiology Duke University Durham, North Carolina

Dr. Dragan Dimitrov Monterey, California

Dr. Simona Nemets Faculty of Mechanical Engineering Technion–Israel Institute of Technology Haifa, Israel

Dr. Kafui Dzirasa Department of Neurobiology Duke University Durham, North Carolina Mr. Nathan Fitzsimmons Department of Neurobiology Duke University Durham, North Carolina Dr. Damien Gervasoni Faculté de Médecine Laënnec University Claude Bernard Lyon, France Mr. Timothy Hanson Duke University Durham, North Carolina Dr. Mikhail Lebedev Department of Neurobiology Duke University Durham, North Carolina

Dr. Miguel A. L. Nicolelis Department of Neurobiology Duke University Durham, North Carolina Mr. Joseph E. O’Doherty Department of Biomedical Engineering Duke University Durham, North Carolina Dr. Laura M. Oliveira Department of Neurobiology Duke University Durham, North Carolina Dr. Albino J. Oliveira-Maia Department of Neurobiology Duke University Durham, North Carolina

Mr. Gary Lehew Department of Neurobiology Duke University Durham, North Carolina

Dr. Parag Patil Department of Neurosurgery University of Michigan Hospitals Ann Arbor, Michgan

Dr. Shih-Chieh Lin Department of Neurobiology Duke University Durham, North Carolina

Dr. Aaron Sandler Department of Neurobiology Duke University Durham, North Carolina

xix © 2008 by Taylor & Francis Group, LLC

xx


Dr. Sidney A. Simon Department of Neurobiology Duke University Durham, North Carolina

Dr. Dennis Turner Division of Neurosurgery Duke Medical Center Durham, North Carolina

Dr. Eric Thomson Department of Neurobiology Duke University Durham, North Carolina

Dr. Michael Wiest Department of Neurobiology Duke University Durham, North Carolina

Dr. Miriam Zacksenhouse Faculty of Mechanical Engineering Technion–Israel Institute of Technology Haifa, Israel


1

State-of-the-Art Microwire Array Design for Chronic Neural Recordings in Behaving Animals Gary Lehew and Miguel A. L. Nicolelis

CONTENTS Introduction................................................................................................................ 1 The DUCN Design and Fabrication Approach for Multisite, Chronic Neural Ensemble Recordings...................................................................................... 2 The Layered Approach .............................................................................................. 4 Discretely Wired Approach ..................................................................................... 10 Building a Layered PCB Microwire Array.............................................................. 15 Building a Discretely Wired Array.......................................................................... 18 Conclusions ..............................................................................................................20 References................................................................................................................20

INTRODUCTION Over the last two decades, many laboratories around the world have started to rely on microelectrode arrays formed by fine microwires, organized in different geometrical configurations, to chronically record the extracellular activity of populations of individual neurons in both anesthetized and behaving animals (Nicolelis et al. 1997, 2003; Lebedev et al. 2006; Verloop and Holsheimer 1984; Williams et al. 1999). As the field of chronic multielectrode recordings evolved, so did the designs of such microwire-based arrays. Indeed, during the last 13 years, our laboratory at the Duke University Center for Neuroengineering (DUCN) has specialized in producing a large variety of microwire array configurations that can now be utilized in a large variety of species (e.g., mice, rats, monkeys, and intraoperative human recordings). In particular, our efforts have been directed at producing arrays that can be utilized in experimental protocols demanding simultaneous recordings from large samples of single neurons 1 © 2008 by Taylor & Francis Group, LLC

2


(e.g., 50 to 500), distributed across multiple cortical and subcortical brain sites in fully awake and behaving animals over long periods of time (months to years). The goal of this chapter, therefore, is to review the DUCN accumulated experience and describe its current state-of-the-art design and fabrication approach for producing high-quality microwire-based arrays for chronic, multisite, neural ensemble recordings.

THE DUCN DESIGN AND FABRICATION APPROACH FOR MULTISITE, CHRONIC NEURAL ENSEMBLE RECORDINGS Over the years, the DUCN technical staff has abided by the principle that microelectrode arrays for chronic single-neuron recordings have to be designed according to the main anatomical characteristics and contours of the specific brain area targeted in a particular experimental project. Thus, by collaborating with neurophysiologists specialized in working with different animal species, our technicians developed a large variety of microwire array (bundle) configurations for targeting different cortical and subcortical brain areas. This initiative proved to be essential for achieving optimal configurations that significantly increase the neuronal yield and longevity of our chronic recordings. In this design process, both 2-D and 3-D contour matching can be performed. In addition, different strategies have been implemented to fabricate adjustable microelectrodes. Such techniques allow microwire tips to be repeatedly and selectively repositioned after surgery. Using this option, users can resample the neural area of interest once the physiological properties of a given populations of neurons have been accomplished. Among other advantages, such a data collection strategy increases considerably the overall neuronal sample obtained per recording site over the course of several weeks or months. In recent years, we have added stimulation electrodes, local ground reference electrodes, and cannula for drug injection to the basic configuration of such microwire arrays, increasing significantly the range of experimental manipulations that can be carried out once such multiple devices are chronically implanted in the animal’s brain. The basic configuration of the DUCN microelectrode arrays consist of insulated metallic conductors with an overall diameter in the range of 25 to 50 μm. The metallic conductor is selected based on corrosion resistance and degree of stiffness, because the wire is driven lengthwise. Tungsten or a stainless steel alloy is often used with an insulation coating of polyimide, Formvar, Isonel, or Teflon. The micro wires are processed to be as straight and stiff as possible. Using a variety of cutting devices, the tips of the wires can be cut blunt or at any desired angle (see Figure 1.1). The angle cut exposes more of the metallic conductor, which lowers the impedance of the electrode. This can reduce the ability to record from well-isolated single neurons. However, the sharply raked tip of the angle cut makes penetration through the brain tissues much easier and less traumatic, while producing less resistance as the length of the microwire travels through the tissue. At this level of design, multiple factors have to be balanced in order to produce an electrode that will yield viable long-term single unit recordings. For instance, not only must optimal impedance for single unit recordings be a consideration during electrode production, but care must be taken to produce an array that will penetrate


State-of-the-Art Microwire Array Design for Chronic Neural Recordings

FIGURE 1.1

3

Two 35-μm electrode tips, top is angle cut, bottom is blunt cut.

tissue easily during the surgical insertion process. Conversely, relying solely on a fine tip microelectrode array (Schwartz 2004) to facilitate tissue penetration may increase the chance of encapsulation of the electrode tips by glia and extracellular protein deposits and thus diminish the ability to record neural signals beyond a few days or weeks. The lesson here is simple: Finding the right compromise among a large number of variables (electrode material, electrode tip shape, cut angle, etc.) is the single most relevant challenge facing developers of microelectrode arrays for chronic recordings. In the DUCN fabrication routine, electrode tips can be left bare, or they can be electroplated, either selectively or as a group, to reduce impedance. An electrode impedance-measuring device such as the A-M systems model 2900 is a useful tool to quantify the variables in electrode impedance, such as the effect of electroplating on impedance. The spacing or separation of the recording electrode tips are in the range of 200 to 1000 μm, center to center, depending on the brain area targeted, the animal species utilized, and the experimental protocol. As a design principle, our electrode arrays are fabricated as part of a complete investigative system. Moreover, a series of features are incorporated into the design of the array to improve the efficiency of the manufacturing, testing, and surgical implantation processes. For instance, reinforced breakaway tabs are used to provide a standard clamp point for electrode holders of the stereotaxic equipment used during surgery as well as assembly and testing operations. At the end of the implantation surgery, the tabs are easily removed. Another set of small breakaway tabs are used to hold alignment in assembly jigs as the microwires are positioned and bonded in place on the printed circuit board (PCB). These same tabs are also used later in the manufacturing process as a palette for conductive paint as it is applied at the microwire to printed circuit trace junction. Guide holes in the PCB are also used to accept alignment pins when multiple arrays are combined or stacked.


4


THE LAYERED APPROACH The various methods employed to manage and organize microwires into arrays can be classified as either layered or discretely wired. In some cases both approaches are used to combine microelectrode arrays on to a single-output connector to streamline the interface to the subject under investigation. The layered approach involves designing PCBs for the purpose of adapting the spacing of the electrode array to the spacing of the chosen connector and providing a platform on which the array components can be mounted. As shown in Figure 1.2, the plated through-hole pads facilitate the mounting of a surface mount connector, and the traces connected to each pad are positioned to mechanically and electrically bond to microwires in an array. The PCB designs are very specific and dedicated to a limited number of applications because the thickness of the board determines row spacing, and the printed circuit trace determines the spacing of electrodes within each row. Alignment jigs, specific to each array pattern and variation, are fabricated for use in the assembly process. The layered approach requires a significant initial investment, and minor changes can be costly because the boards and jigs are specifically dedicated to a given application. The benefit of this approach is that higher production volumes are achievable, meeting the demand with repeatable high quality devices. The majority of layered designs are arrays with two row groups, combined with other assemblies to form multiple rows, or used singularly as a two-row array. Variations of this design exist in many forms. Figure 1.2 illustrates a design in which two separate arrays are built onto a single board and combined onto a single connector. This design allows for implanting into two regions of the brain in a single step. Combining the signals in one output connector helps to streamline the interface to the subject. A further variation of this layered design incorporates remote arrays or satellite arrays. These satellite arrays are connected to the central array via flexible cable assemblies. With this design, the electrode arrays are positioned independently but utilize a single output connector.

FIGURE 1.2 Two 4 × 4 electrode arrays share a common 32-channel connector.



FIGURE 1.3

5

A 2 × 8 array and a movable 2 × 4 array.

Figure 1.3 shows a layered array which contains the additional feature of being movable in depth after installation. Several millimeters of travel allow researchers to fine tune the electrode position and also extend the usefulness of the array. A further variation of this layered design incorporates remote arrays or satellite arrays. These satellite arrays are connected to the central array via flexible cable assemblies. With this design, the electrode arrays are positioned independently, but utilize a singleoutput connector. As shown in Figure 1.3, taken in midfabrication, a 2 × 8 electrode array is combined with a moveable 2 × 4 array and a remote or satellite 2 × 4 array (not pictured). All electrode wires in these arrays are 35 μm tungsten spaced at 250 μm. The spacing between the 2 × 8 and moveable 2 × 4 arrays is 300 μm. A 32-channel preamplifier commonly referred to as a headstage, can be seen connected to the output connector of the electrode array. The headstage provides a buffer between the electrode and the recording equipment in terms of amplification, impedance matching, and filtering. It is sometimes necessary to fabricate electrodes with the headstage in place to allow for a working clearance with electrode holders, stimulation connectors, moveable components, injection cannula, or other ancillary features. Figure 1.4 shows a custom-built electrode holder attached at the clamp points of a 2 × 16 array PCB. This board adapts the spacing of a 0.025 in. surface mount connector to the array pattern of 250 μm on centers. The board is FR4 composite material. The overall thickness dimension of the bare FR4, as seen at the right edge of the board, is a major determining factor in the spacing between rows of the array. The dark green area is a solder-masking agent for insulation.


6


FIGURE 1.4 Electrode holder clamping a 2 × 16 PCB prior to the installation of the surfacemount connector.

The PCB shown in Figure 1.5 illustrates the manner in which microwires are mechanically bonded at the edge of the board and electrically bonded in a staggered formation along the gold-plated traces. The black area is a solder-mask agent used to provide electrical insulation. The tab to the right was used as a palette during the application of conductive paint and can be removed along with the tab on the left later in the fabrication process. The hole in each tab facilitates alignment if multiple assemblies are combined. The wires are 35 μm diameter tungsten with polyimide insulation. The insulation is removed at the silver paint locations to provide the best possible conductivity.

FIGURE 1.5

A close-up view of microwire electrodes attached to a printed circuit board.



FIGURE 1.6 fabrication.

7

96-channel, 6 × 16 array with 35 μm electrodes spaced at 250 μm in mid-

With the layered approach, the assemblies, which form two rows of electrodes, are combined and connectors are installed on each PCB or bridged between boards. As shown in Figure 1.6, the assemblies can be stacked to combine high-density arrays, or used singularly as a two-row array. Fixtures or jigs are designed and constructed for the purpose of achieving alignment of multiple PCBs or, in many cases, soft moldable clay can be used to temporarily hold alignment while a permanent bond is made with adhesive. The example shown in Figure 1.6 is a 96-channel, 6 × 16 electrode array with 35 μm wire spaced at 250 μm on center. This array consists of three 2 × 16 assemblies aligned and bonded with a coiled common ground wire in the foreground. A single set of electrode holder tabs are left in place, which will be removed in the final stages of surgery. The electrode holder tabs are reinforced by plated copper cladding, which provides a robust mechanical and electrical ground connection between the electrode and stereotaxic equipment during surgery. The 16-channel 4 × 4 array with 50 μm wires spaced at 250 μm shown in Figure 1.7, is formed by joining two 2 × 4 assemblies. A double row surface-mount nano connector spans the pads of both boards. The two boards are separated using a shim material to obtain the correct spacing. The staggered microwire termination pattern, used to promote ease of assembly, is visible through the clear epoxy coating. The coating is applied for insulation and additional mechanical integrity. A single set of holder tabs is left in place, and will be removed after installation. Multiple array designs can sometimes be included on the same PCB in the interest of promoting economy by design. Shown in Figure 1.8, two spacing variations


8

FIGURE 1.7


A 4 × 4 array completed and ready for testing.

FIGURE 1.8 Multiuse layered array PCB design being fitted with microwires in an assembly jig.

of layered 4 × 4 arrays occupy a common electrode-printed circuit assembly, joined at the holder tabs. Guide plates, visible in the center, are also included in the design and will be removed for use in fabrication jigs or included in arrays with ultralong



FIGURE 1.9

9

Dual 4 × 4 layered array with drug-injection cannula guides.

electrode wires that require floating guide plate supports. The board is shown mounted in an assembly jig and partially completed with one set of 35 μm electrode wires attached at corresponding traces of the board. Two sets of pads per trace are included to provide a universal mounting for the connector or for remote wiring of a satellite array. The jig is fabricated from a vast collection of the salvaged remains of routed FR4 panels in conjunction with the custom-drilled wire guides. The critical dimensional requirement of the bare FR4 thickness of a layered array is easily met by using the scrap materials from the production run. The supporting deep green-colored framework material in this case is FR4 material salvaged from other PCB production runs. The electrode featured in Figure 1.9 is a 32-channel, dual 4 × 4 layered array with electrodes spaced at 250 μm. The arrays are spaced at 2.5 mm. Guide tubes for drug injection cannula are fabricated from 23-gauge stainless steel tubing. The guide tubes are shown with inner pins fabricated from 0.013 in. stainless steel wire. The inner pins maintain a clear tract and are replaced with injection cannula when needed. The black material is a rubberized cyanoacrylate adhesive, which is used to provide mechanical strength to the assembly, bonding the connector and guide tubes firmly to the PCB. The layered 128-channel array shown in Figure 1.10 illustrates how the separate assemblies are combined to provide a high density array. This array features 35 μm tungsten electrodes spaced at 250 μm and organized in eight rows of 16 electrodes. The electrode holder tabs have been removed on all assemblies except one, which will provide a clamp point for the entire assembly. The coiled ground wire is a common ground connection to all four assemblies and is attached after the individual assemblies are combined. The connectors have been spaced to provide clearance to allow the four 32-channel headstages to simultaneously interface. This design features an offset between the array and connector to allow the array to be in close proximity to other arrays. This offset shifts the array footprint to coincide with the


10


connector footprint along two planes in order to allow clearance for adjacent arrays or to shift the connector location to a more suitable location on the subject.

DISCRETELY WIRED APPROACH The discretely wired approach is to form the array with straight microwires, mechanically bonding the microwires in the desired pattern and then routing the free ends of microwires to the connector. PCBs designed for this application are a tremendous benefit with regard to conformity in manufacturing, especially when small-scale, surface-mount nano-size connectors are used. This approach can be easily applied to achieve complex spacing and shapes, more so than with the layered approach. A single PCB design can be utilized for a multitude of array designs, because the printed circuit design is no longer directly related to the array pattern. In this case, the PCB provides two functions: to provide a structural component used to hold and position the assembly during manufacture and implantation, and to provide a mechanical and electrical bonding of the wire to the connector pin. With moveable designs, the PCB can also be used as the structural foundation for the moveable implements of the array. These boards also include features similar to boards used with layered arrays, such as tabs for clamping and holding the assembly, tabs for use as a palette, guide holes for alignment, and array patterns. Assembly jigs are designed and built specifically for each array pattern and can be partially or selectively loaded with microwire to yield variations within each

FIGURE 1.10 128-channel 8 × 16 layered array.



11

FIGURE 1.11 (See color insert following page 140.) 8 × 8 electrode array fabricated using the discrete-wired method.

pattern. These arrays can be used in various combinations of satellite configurations or as a singular device. The 64-channel electrode array shown in Figure 1.11 was built using a discrete method and sacrificial wire guide. The wire guide in this case is used as a part of the assembly jig and is incorporated as part of the electrode assembly. The microwires are 35 μm diameter, spaced on a 1000 μm grid. Three separate PCBs are used in this example, one for each of the two connectors, and the third is to establish and maintain the spacing of the electrodes forming the wire guide. Epoxy is used as a conformal coating to provide insulation and mechanical integrity. A 32-channel independently positionable array is shown in Figure 1.12, during the assembly process attached to the jig. The pins on the left are 0.010 in. diameter tungsten pins; the wires on the right are 35 μm tungsten electrodes with a spacing of 1000 μm on center. In this moveable design, the electrode wire is attached at the tip of the pin, and a coil is formed around the pin to store the additional wire needed in travel. The pin is friction fit into a molded grid and transfers motion to the electrode wire. Silicon gel is molded at the electrode-to-pin interface to provide protection and allow for motion within. The electrode signal wires can be seen at the top of the picture routed to a 32-channel dual row connector. The electrodes can be positioned independently for peak efficiency. Several methods can be used to move the electrodes. A simple manual pin-pushing device can be built with an outer guide tube and inner adjustable depth stop. This can be adjusted with the aid of a micrometer for exact depth settings. A variation on approach is to modify a micrometer, adding the outer guide tube and inner push rod directly to the micrometer as shown in Figure 1.13. A more complex variation of this method is illustrated in Figure 1.14 and Figure 1.15.


12


FIGURE 1.12 A discrete-wired array still attached to the assembly jig.

FIGURE 1.13 A micrometer modified to provide movements to a friction pin moveable electrode.

The device shown in Figure 1.14 is a miniature linear actuator, motorized and encoded to provide accurate movements of the electrode friction pin-drive system. A touch-down sensor in the drive mechanism allows the device to be used in a handheld manner. The operator positions the guide tube over the electrode drive pin, bottoming the guide tube in the electrode shell.



FIGURE 1.14

A miniature linear actuator equipped to advance moveable electrodes.

FIGURE 1.15

Hand-held electrode positioning system.

13

Using the controller shown in Figure 1.15, the linear actuator is advanced until the inner drive rod contacts the electrode drive pin. Upon contact, the linear actuator is automatically stopped. The operator then enters the amount of travel and speed of travel desired at the electrode. A switch mounted on the actuator enables the controller to begin advancing the electrode at the discretion of the operator. The controller provides readout of the motion in microns, and has the capacity to store and retrieve position information, and communicate via RS-232. The electrode in Figure 1.16 is an example of a completed array that incorporates friction fit pins to independently position the electrodes. The main body of the array is fabricated by modifying a type F connector shell and forms the enclosure for the moveable components. The protective cap, threaded on the top of the body, has a


14


FIGURE 1.16 An independently positionable 32-channel array and stereotaxic-mounted cap.

FIGURE 1.17

A 16-channel moveable array used with dual 8-channel headstages.

shaft that provides for stereotaxic attachment. This cap is replaced with a flat top cap after the array is firmly affixed in surgery. The remote-wired connector can be strategically positioned in surgery as needed to provide clearance. The electrode featured in Figure 1.17 is a 16-channel 2 × 8 moveable array using 15 μm tungsten wire. The electrode wires are drawn to a bundle and driven through a cast alignment block with a maximum deployment of 2.0 mm. Single-row connectors are used to connect to two 8-channel headstages. The electrode is shown prior



FIGURE 1.18

15

A 16-channel moveable array used with a 16-channel headstage.

to completion without a conformal coating. Adjustments are made by rotating the slotted head screw at the top. The 16-channel electrode in Figure 1.18 has the moveable drive assembly built into the body of the connector. The drive mechanism is a 0–80 threaded screw, which advances the electrodes at the rate of 317.5 μm/r. The maximum travel for this electrode is 2.0 mm.

BUILDING A LAYERED PCB MICROWIRE ARRAY The first step in building the DUCN microwire array is to design and fabricate the PCBs. With the design parameters such as the array pattern and connector pad layout established, the PCB design will be largely governed by these decisions. A PCB software tool is used to create the artwork files that are then used to manufacture the printed circuits. The files can be e-mailed to a PCB manufacturing facility for use in automated processing machines. Boards can be created, tested, and shipped within a matter of days. It is also possible to have the connectors installed by automated process, if desired. With the PCBs in hand, the next step is to fabricate a jig to assist in aligning the microwires to the board for bonding. The basic concept is to have two identical panels, drilled on centers to the array pattern with a center-mounted slotted beam in place to hold the PCB in the correct position relative to the panels. The microwires are loaded into the holes, a PCB is positioned into the slot, and the process of bonding the wires to the board can begin. Several assemblies can be produced before the jig will require reloading. The panels are drilled with slightly larger size holes than the outside diameter of the insulated microwire. A typical hole size is 90 μm for a 50 μm outside diameter (OD) microwire. The hole size can be critical; if the hole does not provide enough clearance for the OD variations of the wire, the insulation could be damaged and the jig will be more difficult to operate. The panels are usually


16

FIGURE 1.19


Microwires are held in alignment with a jig.

thin sheets of Delrin, FR4 composite material, or paper. The jig components can be included in the layout of the PCB or be built by hand. If the panels will be produced by hand, a precision drill press with X, Y, and Z feed will be needed. A Servo model 7060 drill press and a Newport model 200 XY table (shown in Figure 1.6) are capable of producing good quality jig guide plates. It is helpful to specify that the scrap materials from the routed panels be shipped with the PCBs because the exact thickness of the boards is needed to be matched on the jig for optimum alignment. This scrap material can be used to fabricate the entire jig, or combined with other production runs as needed. With the microwires attached, the assemblies can be tested, and the connector and ground wire installed. A conformal coating is applied as an insulation barrier. In Figure 1.19, a 17 × 2 array is being assembled with a jig. The microwire array spacing is established by drilling the pattern in Delrin. A split beam holds the printed circuit directly over the array pattern as the microwires are attached to the board in a staggered length formation. The jig is then flipped and the process is repeated on the other side of the PCB. With both sides connected the board is removed from the jig slightly and the microwires cut to free the assembly. Several boards can be processed before the jig must be reloaded with wire. Micro drill bits in sizes ranging from 50 μm to well over 100 μm can be used in the fabrication of the jig. These drill bits are used with the drill press and translation stage shown in Figure 1.20. The drill press has a standard Albrecht keystone chuck, but for precision drilling a 1-mm-diameter collet will provide the best accuracy and ease of use with micro drill bits. The flute length versus diameter of commercially



FIGURE 1.20

17

Precision drilling equipment.

available bits shown below indicates the maximum material thickness for a throughhole in the jig plate. The close-up view of a 90 μm drill bit in Figure 1.21 illustrates the flute length of 500 μm versus diameter. Due to their fragility, micro drill bits require very close tolerance drilling equipment to function.


Drill Diameter (in microns)

Flute Length (in microns)

50 60 70 80 90 100 110

400 400 400 500 500 700 700

18

FIGURE 1.21 for jigs.


A 90 μm diameter drill bit typically used to drill microwire-guide arrays

BUILDING A DISCRETELY WIRED ARRAY The first step in building this type of array is to fabricate a jig, which will be used to hold the microwires in alignment for a mechanical bonding process. The jig consists of two thin plates drilled to match the pattern of the array, mounted to a frame with approximately 25 mm space between. The microwires are loaded into the jig and mechanically bonded as an assembly. If the spacing is greater than 300 μm center to center, it can be helpful to use a third sacrificial plate to act as a carrier for the adhesive. This sacrificial plate helps to maintain the integrity of the pattern. The assembly can then be removed from the jig and mechanically bonded to a connector printed circuit assembly. The free ends of the microwires are then routed through plated through-holes, which are electrically connected to the conductor pins of the connector and mechanically bonded. The board is flipped to expose the microwireplated hole junction. The microwires are trimmed to length, insulation is removed, and conductive paint or epoxy is applied at this junction. A conformal coating is applied to achieve an insulation barrier. The discrete wired jig shown in Figure 1.22 is designed to make a 4 × 8 array with electrode spacing of 500 μm and a row spacing of 800 μm. The jig is loaded with 35 μm wires, fixing to be bonded to form the array. Delrin blocks are used to help the adhesive span the distances between rows in order to maintain spacing integrity. The 16-channel adjustable electrode shown in Figure 1.23 features a 0.050 in. hex socket exposed at the top of the assembly, which is used to deploy the microwire electrodes from the cannula after being implanted over the target area. The adjustable mechanism is installed in an extruded brass square tube. The 0–80 threads of the mechanism provide incremental travel of 317.5 μm/r of the hex



19

FIGURE 1.22 Discrete-type jig loaded with 35 μm wires.

FIGURE 1.23

A 16-channel moveable array shown with electrodes partially deployed.

socket. A single reinforced tab provides a clamp point for the stereotaxic electrode holder. The signal wires are terminated remotely to the connector PCB adapter of another array.


20


CONCLUSIONS The design of a large variety of microwire array (bundle) configurations has enabled our laboratory to perform chronic, multielectrode recordings in a variety of animal species, including wild type and transgenic mice, rats, and owl and rhesus monkeys. More recently, the same approach has been translated into a new methodology for monitoring brain activity in patients subjected to neurosurgical procedures. This new generation of microelectrode arrays has pushed the limits of systems neurophysiology and allowed, for the first time, the simultaneous monitoring of the activity of hundreds of individual neurons, distributed across multiple, interconnected cortical and subcortical structures that define particular neural circuits (e.g., somatosensory, motor, gustatory, etc.) for long periods of time (weeks to years, depending on the animal species and experimental protocol) in behaving animals.

REFERENCES Lebedev, M.A. and Nicolelis, M.A.L. (2006). Brain machine interfaces: Past, present and future. Trends Neurosci 29: 536–546. Nicolelis, M.A.L., Ghazanfar, A.A., Faggin, B., Votaw, S., and Oliveira, L.M.O. (1997). Reconstructing the engram: simultaneous, multiple site, many single neuron recordings. Neuron 18: 529–537. Schwartz, A.B. (2004). Cortical Neural Prosthetics. Annu Rev Neurosci 27: 487–507. Verloop, A.J. and Holsheimer, J. (1984). A simple method for the construction of electrode arrays. J Neurosci Meth 11: 173–178. Williams, J.C., Renmaker, R.L., and Kipke, D.R. (1999). Long-term neural recording characteristics of wire microelectrode arrays implanted in cerebral cortex. Brain Res Protocols 4: 303–313.


2

Surgical Techniques for Chronic Implantation of Microwire Arrays in Rodents and Primates Laura M. O. Oliveira and Dragan Dimitrov

And still we could never suppose that fortune were to be so friendly to us, such as to allow us to be perhaps the first in handling, as it were, the electricity concealed in nerves, in extracting it from nerves, and, in some way, in putting it under everyone’s eyes. Luigi Galvani, 1791

CONTENTS Introduction.............................................................................................................. 22 Differences Between Rodents and Primates Pertinent to Surgical Technique........ 23 Surgical Techniques for Rodents .............................................................................24 Preoperative Supplies and Room Preparation ..............................................24 Preoperative Animal Preparation .................................................................25 Anesthesia Techniques and Intraoperative Monitoring................................26 Electrode Specific for Rodents ..................................................................... 27 Implantation Techniques...............................................................................28 Brain Electrode Arrays ...................................................................... 28 EMG Electrode Surgery .................................................................... 31 Postoperative Care ........................................................................................ 32 Surgical Techniques for Primates ............................................................................ 32 Preoperative Supplies and Room Preparation .............................................. 32 Preoperative Animal Preparation .................................................................34 Anesthesia Techniques and Intraoperative Monitoring................................34 Electrode Specific for Primates .................................................................... 35 Implantation Techniques............................................................................... 35 Exposure ............................................................................................ 35 Cortical Localization ......................................................................... 36 Drilling of Craniotomies.................................................................... 36 Opening of the Brain Coverings ........................................................40

21 © 2008 by Taylor & Francis Group, LLC

22


Electrode Lowering............................................................................ 41 Skull and Wound Closure .................................................................. 41 Postoperative Care ........................................................................................ 43 Conclusions and Future Directions..........................................................................44 Acknowledgments.................................................................................................... 45 References................................................................................................................ 45

INTRODUCTION The study of electrophysiology started with the work of Luigi Galvani (1737–1798), who was the first to provide evidence for the electrical nature of “the mysterious fluid” (at the time referred to as “animal spirits”). Galvani’s nephew, Giovanni Aldini (1762-1834), continued this line of inquiry in 1803, using Galvani’s and Alessandro Volta’s (bimetallic electricity) principles together, despite the fact that Volta did not believe in animal electricity. Carlo Mateucci (1818–1868) in Bologna and Emil Du Bois-Reymond (1818–1896) in Berlin described the phenomenon called “negative variation” when a galvanometer showed an unexpected decrease in current intensity during muscle contraction. The study of the electrophysiology of the nervous system began when Julius Berstein (1839–1917) proposed his theory of the nerve impulse as a wave of negativity (membrane theory of the nerve tissue). Later, using a galvanometer with one electrode in the gray matter and one on the skull surface (or electrodes in different points of the external surface of the brain), Richard Caton (1842–1926) recorded a feeble current in the brain. In 1870, for the first time, Gustav Fritsch (1838–1927) and Eduard Hitzig (1838–1907) inserted an electrode in the dura of a dog brain and stimulated the motor area, generating movement in the contralateral side of the animal’s body (Niedermeyer 1993; Piccolino 1998). The work of these and many other scientists marked the beginning of the study of the electrophysiology of the nervous system, opening doors to the possibility of stimulating different areas of the brain through electrical current and subsequently recording the brain electrical activity. Improvements in electrode manufacturing, the advent of modern acquisition equipment, and better surgical and asepsis techniques have provided us the ability to chronically implant multiple electrodes simultaneously in several areas of the brain in the same animal (Nicolelis, Baccala et al., 1995) and to study the interactions of populations of neurons (Nicolelis, Fanselow et al., 1997; Ghazanfar, Stambaugh et al., 2000). Upon the animal’s recovery from surgery, we have been able to record simultaneously from different brain areas of mice (Costa, Cohen et al., 2004), rats (Faggin, Nguyen et al., 1997; Ghazanfar and Nicolelis 1997; Nicolelis, Ghazanfar et al., 1997) and nonhuman primate brains (Nicolelis, Stambaugh et al., 1999; Nicolelis, Dimitrov et al., 2003) for long periods of time, from a couple of months in rodents (Ghazanfar and Nicolelis 1997; Nicolelis, Ghazanfar et al., 1997) to up to years in non-human primates, such as owl monkeys (Nicolelis, Ghazanfar et al., 1998) and Rhesus monkeys (Nicolelis, Dimitrov et al., 2003). These recordings are carried out under several different experimental conditions and behavior tasks (Kralik, Dimitrov et al., 2001; Nicolelis and Ribeiro 2002). With chronically implanted multiple electrodes, it is also possible to record different


Surgical Techniques for Chronic Implantation of Microwire Arrays

23

layers of the same area of the brain (Chapin and Lin 1984) and study spatiotemporal response of many neurons (Nicolelis and Chapin 1994). Microcannulae can be attached to the electrode arrays and are used to inject drugs in the areas of the implant during chronic experimental recordings (Shuler, Krupa et al., 2002). Chronically implanted electrodes offer unparalleled advantages for correlating neuronal activity and animal behavior. In our lab, these techniques were developed in rodents and later adapted to primates. Over the last several years, there have been significant strides in making rodent implantations more reliable, faster, and easier. We have identified and resolved many of the issues that now permit larger neuronal yields that last longer. As a consequence of continuous improvement in techniques, the length of time required for surgery has been reduced. At the same time, over the last 14 years, we have developed a surgical technique adapted to the unique features of primates. This has made primate implantations routine and reproducible. Here, we will describe detailed technical aspects of the current surgical implantation approach used in our laboratory at the Duke University Center for Neuroengineering (DUCN). Such a surgical protocol has evolved and benefited from almost two decades of accumulated experience on chronic multielectrode neural recordings (Nicolelis, Stambaugh et al., 1999; Nicolelis, Dimitrov et al., 2003, Kralik, Dimitrov et al., 2001; Nicolelis and Ribeiro 2002).

DIFFERENCES BETWEEN RODENTS AND PRIMATES PERTINENT TO SURGICAL TECHNIQUE The success in obtaining recordings from chronically implanted electrodes and how long they last depends fundamentally on the quality of the surgical implantation technique. The ability to open small craniotomies, only large enough to fit the electrode array with minimal bleeding from the bone and from the meninges in very small animals such as mice and rats, requires practicing the techniques several times before attempting to gather good data using the implant of electrode arrays. This is especially true in cortex areas, which are close to the dura and are very sensitive to lesions. The surgical technique for implantation that we follow in our lab is similar for all species, but each species has its own peculiarities. For instance, the mouse head is very small and the skull is thin, requiring a delicate technique. Mouse surgeries require very small screws, drill bits, and custom-designed electrode arrays. Because of the small head size and bone thickness, inserting more than two small arrays of electrodes and more than two fixation screws per animal is not recommended. Compared to mice, the rat’s head is bigger and the bone is stronger, increasing the possibility of using bigger electrode arrays and reaching more cortical and subcortical areas. Because rats are stronger than mice, they can tolerate larger amounts of the acrylate used to secure the electrodes in place, which increases the number of arrays that can be inserted in the brain (Faggin, Nguyen et al., 1997; Nicolelis and Ribeiro 2002). Cortical and subcortical localization in rodents is based on commercially available stereotactic atlases. Theoretically, implanting electrodes into rodents and primates should be very similar; however, in practice they are vastly different. For investigators accustomed


24


to the hardiness of rodents and trained in rodent implantation techniques, performing similar procedures on primates can seem overwhelming. For instance, the commitment of time, personnel, and lab resources is much greater for primate surgery. Furthermore, from the point of anesthesia, primates require much more attention during surgery. Whereas rodents require only monitoring of a few physiological parameters and infrequent injections, maintaining a primate under anesthesia is much more labor intensive. Thus, for our primate surgeries, one member of the surgery team is dedicated to monitoring the animal throughout. This higher level of anesthesia technique is akin to pediatric anesthesia and requires specific equipment, planning, and personnel. Obvious differences in anatomical details include a thicker skull, thicker brain coverings, and a better-developed subdural space, all of which also influence the definition of the optimal surgical strategy for chronic multi-electrode implantation in nonhuman primates. The thicker skull requires forethought in terms of the appropriate electrode length. The brain coverings including the dura and pia are better developed, more variable in their thickness and at least the dura requires wide opening with microsurgical instruments for electrode penetration. We have found that the pia and arachnoid layers of primates are more variable, often tougher, and more prone to dimpling than rodents, sometimes necessitating microsurgical opening, as well as paying careful attention to be sure penetration has occurred. The primate brain is more prone to problems with swelling and retraction due to fluctuations in CO2 levels in the blood that are primarily affected by ventilation of the animals’ lungs. Plans must be in place to deal with these issues intraoperatively. Overall, we have also observed that intraoperative neural recordings are more difficult in primates than they are in rodents. Much more attention must be paid to noise detection and reduction, especially because many more electrical devices necessary for anesthesia are involved. Primates are more prone to infection than rodents and require more attention to sterile technique throughout, with sterilization of all the instruments, the use of sterile gloves and gowns, and draping of the surgical field. For those unaccustomed to working in a sterile environment for long periods of time, this can present unforeseen challenges and cost significant amounts of time. In summary, primate implantations require a team approach. It often takes several days to weeks to prepare and coordinate one’s lab prior to performing primate surgery, underscoring the critical importance of preoperative planning.

SURGICAL TECHNIQUES FOR RODENTS PREOPERATIVE SUPPLIES AND ROOM PREPARATION Typically, our rodent surgeries are performed on a surgery table in a regular room in the laboratory or in a designated surgical room. The table is kept clean and uncluttered and has the stereotaxic apparatus already installed (for rats we use the cat and small primates stereotaxic apparatus with a rat adaptor, and for mice, we use the stereotaxic apparatus for small rodents, both from David Kopf Instruments, Tujunga,



25

California). Between the two stereotaxic bars, a rectangular platform made of plexiglass is glued to a height adjustable stage where a warm pad is placed. This apparatus is used for the anesthetized animal for the duration of the surgery and immediate postoperative recovery. Other essential pieces of equipment include a binocular surgical microscope, a dental drill already installed in a source of compressed air, an amplifier connected to an audio monitor, an oscilloscope, and a micropositioner. The table, the stereotaxic apparatus and the plexiglass platform are cleaned before the surgery with 70% alcohol or Asseptiwipes (wipes moistened in a solution of N-alkyl(68%C12, 32%C14) dimethyl ethyl benzyl ammonium chloride 0.125%, N-alkyl(60%C14, 30%C16, 5%C12, 5%C14) dimethyl ethyl benzyl ammonium chloride 0.125%, isopropyl alcohol 14.850%, and other ingredients. All our surgeries follow the National Institute of Health (NIH) and Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) guidelines for rodents that are expected to recover from anesthesia. These guidelines include appropriate preoperative and postoperative care of the animals, asepsis (sterilization of the surgical tools, use of sterile gloves, mask, head covers, clean lab coat, and aseptic procedures), gentle tissue handling, minimal dissection of tissues, effective hemostasis, and correct use of suture materials and patterns when stitches are necessary (National Research Council 1996; Baumans, Remie et al., 2001). The surgical instruments (basically, four small hemostats, scalpel handle, Freer periosteal elevator, scissors, micro scissors, regular and micro tweezers, dental drill handpiece, special screw drivers, stainless steel wire cutters, and micro cup curette) and supplies (gauze, cotton-tip applicators, kimwipes, drill bits, small stainless steel screws, silicone cups for dental acrylic, and beakers) are sterilized by autoclave. All supplies that cannot be exposed to heat and humidity are sterilized by ethylene oxide gas (extrafine point markers, plastic rulers, electrode holders, electrodes, etc.). The microwire arrays (see chapter 1) are made of very delicate materials and cannot be sterilized by autoclave. Instead, these must be sterilized by ethylene oxide gas. The electrodes should be carefully packed in such a way that they are not damaged during handling, transportation to and from the sterilization facility, or by personnel helping in the surgery. Because of the length of these surgeries, all the equipment used during surgery should be routinely tested the day before the animal is anesthetized. To avoid unnecessary delays during surgery, it is imperative that all materials, equipment, and drugs for the surgery are readily available and in good working condition. A copy of the brain atlas appropriate to the animal undergoing implantation surgery is necessary to help in the location of the areas to be implanted. It is advisable to complete these steps the day before the surgery; however, they can be completed the day of the surgery prior to anesthetizing and preparing the animal.

PREOPERATIVE ANIMAL PREPARATION Typically, all animals selected to undergo chronic implantation of multielectrode arrays have several days to acclimate in the new laboratory facility before surgery. Often, these animals are subjected to weeks of behavioral training before they are implanted with arrays of electrodes. Only animals in good health are subjected to the surgery.


26


In general, we work with Long–Evans adult rats, males (in general 300–350 g) or females (in general 250–300 g). However, depending on the studies we can use younger or older animals. Mice are chosen according to their genotype since our laboratory currently utilizes a variety of mutant animals and their wild-type littermates as control (Costa, Lin et al., 2006; Dzirasa, Ribeiro et al., 2006). Because rodents have a high metabolism and are not at risk for vomiting and aspirating, food is not withheld until up to 2 h before the surgery. Plans for anesthesia, surgery, and determining the coordinates of the areas to be implanted must be carefully studied ahead of time. This will help decrease the time of the surgery for electrode-array implantation. A record sheet should be started for the procedure prior to beginning anesthesia. This not only meets the requirements of AAALAC, NIH, and Institutional Animal Care and Use Committee (IACUCs), but also provides a record of the procedure, the implanted areas and their coordinates, the depth of the electrode for each array, and the position of the ground wires related to the position of the connector of the array.

ANESTHESIA TECHNIQUES AND INTRAOPERATIVE MONITORING Following guidelines from Duke University’s IACUC and the advice of Duke’s veterinary staff, we use two anesthesia regimens for rat surgeries and one for mice surgeries. For rats: Pentobarbital IP Ketamine associated with Xylazine IM For mice: Ketamine IM associated with Xylazine IM Pentobarbital has a good hypnotic effect, but analgesia is obtained only when using high doses of the drug, which can cause cardiovascular and respiratory depression. The advantage of using Pentobarbital is that the effect of the anesthesia lasts longer than the Ketamine/Xylazine combination. Ketamine is a dissociative anesthetic that provides light surgical anesthesia with a short duration. When combined with sedatives or tranquilizers, the quality of the anesthesia is highly improved and the duration of the effect is increased. In general, in our surgeries supplemental doses of Pentobarbital are required every 2 h and Ketamine every 1½ h. The supplemental doses for both drugs are from ⅓ to ½ of the original dose, but Xylazine is not supplemented unless the surgery lasts longer than 7–8 hours, which is very uncommon in rodent surgeries. For both anesthesia regimens (rats and mice), anesthetic induction is carried out in a chamber with a mixture of 5% Isoflurane and O2. In our experience, a good induction with Isoflurane helps the selected anesthesia regimen last longer, decreasing the need for frequent injections of supplemental doses of the anesthetic. Once the animal is deeply anesthetized with Isoflurane, an injection of the selected drug is administered (Duke University DLAR 1995; National Research Council 1996; Hellebrekers and Booij 2001). Following the injection of anesthesia, the animals



27

may receive a dose of Atropine SQ to prevent or treat excessive secretion of the airways and to improve cardiac function. This is important especially with the use of Pentobarbital. For rats, if the surgery is expected to last for a longer period, small doses of sterile saline can be injected SQ, and IP injections of 10% dextrose solution will help to maintain hydration and blood glucose level during the surgery. In general, mice surgeries are much shorter than rat surgeries due to the small number of screws and arrays that are implanted, and extra volume or dextrose injections are not necessary. Once the animal is deeply anesthetized, it is placed in an area for preparation for surgery, apart from the operating theater. The animal’s head is shaved with small clippers, from the area just above the eyes to the back of the head and from ear to ear. If electromyography (EMG) electrode implants are planned, the skin on top of the target muscles is also shaved at this point. After the shaved hair is cleaned, the animal is transported to the surgery table and placed on the platform with a thermal pad to maintain the animal’s body temperature throughout the surgery. If an electrical pad is used, the pad should be covered to avoid direct contact with the animal in order to prevent burns to the animal’s skin in case the pad overheats, and a rectal temperature probe should be inserted. A good option is the Deltaphase® thermal pads (Braintree Scientific, Braintree, Massachusetts) that maintain temperature around 37ºC for about 5–6 h. This type of pad must be monitored and changed when the temperature decreases. After the animal is on a warm pad, it is mounted on the stereotaxic apparatus, using proper ear bars for rodents, placing the mouth and nose piece, and leaving the position of the head fixed until the top of the head is parallel to a flat surface. Once the rat head is secure, the sterile pack can be opened and the sterile materials should be kept in a sterile field or container. The skin of the head is cleaned three times using Iodophor (iodine soaps) or Chlorhexidine followed by alcohol 70%. The animal’s eyes are covered with an ophthalmic ointment to prevent corneal lesions since the animal loses the blinking reflex with the anesthesia.

ELECTRODE SPECIFIC FOR RODENTS The arrays of electrodes to be implanted are made in-house and vary in number and distribution of electrodes. For the past nineteen years, our experiments have utilized arrays or bundles made of thin microwires (see chapter 1 for details). Some arrays can be made to reach more than one area of the brain. In general, for rat surgeries, the arrays are comprised of 32 electrodes (4 × 8), but they can also be made up of either 16 (2 × 8) or 64 (8 × 8) electrodes. The tips are cut either blunt or sharp. When cut sharp, the electrodes may be able to penetrate the brain without dissecting the dura. In this case, the length of each electrode may be slightly different (see Figure 2.1). If necessary, a little cut on the dura with a bent fine needle may help relieve the tension of the meninges and facilitate the implantation of the electrodes. Obviously, the length of the microwires varies for each surgery and depends on the brain area to be implanted.


28


FIGURE 2.1 Sharp electrodes penetrating the dura in a rat surgery. Photo by Edgard Morya.

IMPLANTATION TECHNIQUES Brain Electrode Arrays The length of the surgery for the implantation of microwire arrays varies from animal to animal. It also depends on the number of arrays to be implanted, the location of such brain regions, and the experience of the surgeon. In general, the minimum time is about 3–4 h if only one or two arrays are to be implanted. Each additional array can add about 45–60 min to the total surgical time. It is essential not to rush the implantation procedure. In our accumulated experience, gentle and slow penetration of the brain tissue yields the best long-term results. During surgery, the status of anesthesia is checked periodically by pinching the inter-digit membrane of the hind paw or pinching the tails of the animal. Supplemental doses of anesthesia are given as needed. The areas of the brain to be implanted also vary for each study and may vary even in the same study because it is not possible to reach all areas of interest for each animal, especially in mice. The surgical procedures described below have been very successful in the past. Following this approach has allowed our experimental animals to tolerate these implants very well, survive without any postsurgical complication, and provide high-quality recordings for months after the surgery. After the animal is anesthetized, mounted to the stereotax, and cleaned, the surgical team suits up for surgery in a lab coat or gown, mask, sterile gloves, and head cap. Before starting surgery, the status of the anesthesia is checked as described above. To alleviate pain during the incision, an injection of lidocaine 1% or Bipuvocaine can be given under the skin at the site of incision. An incision is made on the midline of the scalp, from just above the eyes to the back of the head, and the skin is propped open. The borders of the skin are held open with small hemostats, and bleeding, which in general is very small, is cleaned with gauze or cotton-tip



29

applicators. The periosteum is scraped off the skull bone using a blunt tool (Freer) or the scalpel blade. It is very important to clean the bone surface very well and remove all soft tissue attached to the bone. Bleeding spots on the bone may be cauterized or pressed until the bleeding stops. The bone is then cleaned with hydrogen peroxide, applied very carefully to avoid touching soft tissues around the cut. This process can be repeated several times until the bone is clean, whitish, and all the blood is removed from the bone surface. The hydrogen peroxide is washed with sterile 0.9% saline, which is repeated as many times as necessary to remove the excess chemical. The bone must be dried with gauze and must look very whitish. This step is most important for the fixation of the electrodes with dental acrylic. Any soft tissue or bleeding may lead to infection of the site, and the head cap may become dislodged some time after the surgery. When the skull is clean and dry, an extrafine point marker, can be used to mark the skull. For rodents, the zero point is the center of the bregma. Using the stereotaxic apparatus, marks are made at the center of the area of the planned implants. The margins of the craniotomy for the electrodes are drawn around the central marks and other marks for the fixation screws are made. For mice, two screws are used; for rats, five to six screws are used. These metal screws (stainless steel, blunt tip) will hold the electrodes and the head cap in place and also provide a common electrical ground for the microwire arrays. Using a dental drill under a surgical microscope view, small holes for the screws are then drilled. At least one of the screw holes must be completely open, exposing the dura. The screw placed in this hole will be in contact with the dura and will be used as a ground for the electrodes. The screws are placed in the holes and tightened with a screw driver, turning the screw enough to be firmly attached to the bone without penetrating the brain. If the dura is accidentally opened and cerebrospinal fluid (CSF) leaks from any hole after the screws are in place, it is recommended to either seal the screw with cyanoacrylate glue, or remove the screw and seal the hole with glue and place the screw in another location in the skull. Leaking of CSF under the head cap can lead to infection and softening of the bone over time, and will cause the head cap to become loose in the future, compromising the electrode positioning and increasing the risk of the head cap becoming detached from the skull. After all screws are in place, the craniotomies for the electrodes are drilled. Care is taken to avoid damage to the dura during the drilling process. The size of the craniotomy must be just large enough to accommodate the electrode arrays. With a small drill bit, the edges of the craniotomy are drilled all the way to the bottom of the bone. To prevent damage to the dura, the bone should be tested frequently for softness. When the bottom feels soft, the rest of the bone can be taken out with a small cup curette, the tip of a needle, or with the tip of a dental explorer. Once all the bone from the edges of the craniotomy is drilled, the loose bone in the center can be removed either with a small sharp tool or micro tweezers. Once the craniotomy is opened, it is washed with saline until all the bone dust is removed. The skull needs to be dried again, and the craniotomy is covered with a piece of Gelfoam® moistened with saline or filled only with saline to prevent the dura from drying. The other craniotomies are opened the same way and covered with Gelfoam or saline. Drilling of the screw holes and all craniotomies should be done before the placement of the


30


electrodes to avoid vibration of electrodes after they are implanted in the brain. If no damage is done to the dura, the craniotomy will look very clean and the surface of the brain and blood vessels can be seen. Once all the craniotomies are opened, electrodes are set up for insertion, one array at a time. The Gelfoam is taken out of the craniotomy and is replaced by saline solution. The array, which is connected to headstages (cables and connectors) placed in special holders attached to the stereotax micromanipulators, is slowly lowered where the craniotomy is open to minimize brain damage and bleeding. Depending on how the tip of the electrodes are cut (sharp or blunt), the electrodes may be inserted in the brain without opening the dura. This maneuver is done slowly and with patience until the electrodes pierce the dura. If the wires are blunt, or it is impossible for them to break through, the dura must be opened. This can be done with a fine needle with a bent tip, taking care to avoid breaking the blood vessels. In general, a cut on the dura is enough to relieve tension, but it may be necessary to remove all the dura in the craniotomy. After the electrodes are touching the brain surface, the special grounding wire is wrapped around the screws, including the screw reserved for grounding. The electrodes can be grounded in other ways. For example, the tip of the ground wire can be placed under the skull in a small hole, and this hole can be fixed with cyanoacrylate gel glue to secure the ground wire in place. Then the rest of the ground wire can be wrapped around the screws. However, this technique requires more time and patience. Cell activity should be monitored during the implantation of the electrodes to help ensure placement of the array in the desired layer or structure of the brain. If during the placement of the electrodes, the superficial blood vessels break and bleeding occurs, it is better to remove the electrodes from the craniotomy, press the hole until the bleeding stops, and wash the craniotomy and electrode tip with saline. In our rodent surgeries, bleeding in the dura is not cauterized. Blood on the tip of the electrodes may interfere with monitoring of the brain signals during surgery. A cotton-tip applicator over the craniotomy, a little roll of sterilized kimwipe or a piece of Gelfoam will help stop bleeding. Once bleeding stops, the craniotomy is extensively washed with saline, and the electrodes can be placed in the craniotomy again and slowly lowered into the brain. The electrodes can be lowered by hand with very small turns of the micromanipulators or with a micropositioner, about 100 μm at a time waiting a couple of minutes before proceeding with the lowering of the electrodes. Dimpling of the brain may occur during electrode implantation, and it can result in traumatic brain injury. In our experience, this can be prevented by making the craniotomy as small as possible and performing the implantation slowly and gently, especially in cortical implants. If the first insertion of the arrays is not successful and no signals are recorded, removing the electrodes and reinserting them in the same craniotomy does not work well for cortical areas, but may succeed for subcortical implants. In this case, the electrodes can be reinserted in the same craniotomy in a slightly different position. After the array of electrodes penetrate the brain and reach the desired area (cortical or subcortical), a layer of Gelfoam or warm agar is placed on the craniotomy around the microwires, being careful to avoid disturbing the position of the electrodes already in place. The excess ground wire is cut and the skull surface is cleaned and extensively



31

dried again because cyanoacrylate glues or dental acrylic does not attach very well to wet bone. A drop of cyanoacrylate glue (gel) or dental acrylic is used to fix the electrodes to the nearest screws and to the bone, with care to avoid covering the next craniotomies. The craniotomy with the electrodes must be carefully sealed with acrylate to avoid leaking of CSF, especially when the dura has been surgically opened. After the glue or the dental acrylic is dried, the cables for recording (headstages) are removed from the electrode-array connector, and the electrode holder is gently released. If more electrode arrays are to be placed in the brain, the same procedures are repeated until all the arrays are in place. To decrease time during surgery, an accelerator can be used to speed curing of the glue when using cyanoacrylate glues. After all the arrays are in place, a layer of dental acrylic is applied around and between all the arrays and screws to create a strong head cap fixed to the skull. The edges of the head cap should be smooth and should cover all of the bone to avoid postoperative infection. At this stage, a small piece of hard wire may be placed in the head cap (a staple will work very well for this purpose), which will be used to help anchor the head stage to the head cap with tiny rubber bands during future recording sessions. If the skin is loose around the head cap, stitches may be necessary, and are done with suitable sutures. The skin around the cut area is washed with saline, dried, and a layer of antibiotic ointment is applied. Once the surgery is over, the animal is returned to a clean cage and partially positioned under a heating lamp, such that when the animal awakes, it can move away from the heat if necessary. Care is taken to position the lamp to avoid burning the animal. The small particles of the bedding on the cage bottom can be covered with paper towels to prevent possible choking on or ingestion of the bedding when the animal is waking from anesthesia. The animal is checked at least every 15 min until it is fully awake. Moist food pellets and water are offered when the animal is fully awake and moving around. It is very important to observe the animal when it is waking, especially when Pentobarbital anesthesia is used. Secretion in the respiratory airways may cause the animal to choke during this period. If necessary, another dose of atropine may be given to prevent secretion formation and possible choking. All the animals subjected to surgery are housed in individual cages and followed closely for 7–10 days after surgery. EMG Electrode Surgery If the study requires recordings of muscle activity, EMG microwires (typically 50 μm tungsten, isonel coated) can be placed inside muscle pads of rodents. Muscle activity can be recorded for many pairs of muscles. In general, we place 1–6 pairs of EMG electrodes in the following muscles: lavator labii superior, external epicantus, trapezius, biceps, triceps, and gastrocnemius. The EMG wires are attached to one connector (that will be attached to the head cap for the brain arrays) and are implanted during the same surgical procedure as the electrodes implantation in the brain. The EMG microwires are very thin and flexible, will not interfere with movements of the target muscle, and will not disturb the health and mobility of the animal. This step will add up to 30 to 90 min to the surgery depending on the number of pairs of muscles to be implanted.


32


After the brain electrode arrays are in place and fixed with dental acrylic, the shaved area on top of the target muscle is cleaned with iodine soap or Chlorhexidine and alcohol 70%. A small incision on the skin is made with a scalpel, and a long needle (20 Ga x 3½ in. disposable spinal needle) is inserted into the subcutaneous space on top of the muscle. The needle is pushed through the subcutaneous space up to the surgical opening on the head close to the electrode arrays. Once the tip of the needle is outside the cut area on the head, the EMG wire is passed through the needle until the tip is outside the body close to the target muscle. The needle is then removed, leaving the EMG wire in the subcutaneous space; the excess of the microwire is cut, and the tip inserted into the target muscle. The surgical opening is closed using suitable sutures. The area is cleaned and a layer of antibiotic ointment is applied. These steps are repeated for all the muscles to be implanted with EMG wires. After all EMG wires are in place, the head cap is finished, and the cuts are cleaned and covered with antibiotic ointment.

POSTOPERATIVE CARE In general, animals are fully recovered between 7–14 days following surgery, and they usually do not need any special procedure during postoperative days. As a rule, our rodents are treated for pain for the first 24 h after surgery, whether or not they are showing signs of pain. Rats will receive a dose of Tylenol every 6 h or an injection of Buprenorphine every 8 h. Determining which to use depends on the status of the animal. If no signs or minimal signs of pain are present, the animals will be treated with Tylenol. If there is strong evidence of pain (animal is quiet in a corner, walking in circles, not eating or drinking, and has altered aspects of the fur and posture), the animal will be treated with Buprenorphine. Mice are treated for pain in the first 24 h following surgery with Buprenorphine. The animals are observed during the next postoperative days, and pain medication is given as necessary. Typically, our animals are back to their normal behavior and routine 24–48 h after surgery. The surgical wound is checked everyday for signs of bleeding and draining, and to determine whether bedding materials may be contaminating the cut. If wound care is required, the animals will be anesthetized in an Isoflurane chamber (5% Isoflurane and O2 mixture), the wound will be cleaned and a new layer of antibiotic ointment will be applied. If the surgical area appears abnormally red or swollen, is draining fluid or if the animal does not appear to be recovering well, the animal will be treated with systemic antibiotics and pain medication will be given as necessary. After 10–14 days, stitches will be removed. In successful surgeries, the animals tolerate the presence of the head cap extremely well and neuronal activity can be recorded for months.

SURGICAL TECHNIQUES FOR PRIMATES PREOPERATIVE SUPPLIES AND ROOM PREPARATION For nonhuman primates, the surgical area is divided into three rooms. One room is used for general preparation. This room contains head covers, masks, face shield protectors, and shoe covers for the lab staff involved in setting up the surgery and



33

preparing the animal. This room also contains the table where the animal is prepared for surgery. The second room is used as the scrub room and contains a sink for the surgical team to scrub in before surgery, along with sterile garment supplies such as sterile gloves and gowns. The third room is the surgery suite. Surgery at our institution is performed in a dedicated operating room (OR) equipped with OR lights, an operating table, a microscope, suction equipment, ventilation equipment, an oxygen and air supply, a ventilator, and a physiological monitor. The presence of suction equipment is important for venting of inhalational anesthetics. A variety of animal ventilators are commercially available. Not all ventilators work well with very small animals, such as owl and squirrel monkeys, due to the small volumes and pressures that these animals require. A physiological monitor that can display waveforms for heart rate, breathing, oxygen saturation, temperature, and EKG is the cornerstone to intraoperative monitoring of the animal. Surgical instruments consist of basic clamps, scissors, and forceps. More important are the microsurgical instruments. A micro cup curette works well for trimming bone edges on the inside of the craniotomy after drilling. Micro scissors and micro forceps are needed for removal of the dura. Very fine suction catheters are useful to clear CSF from the operative field. The surgical microscope should have binocular vision and must be covered with a sterile drape designed for microscopes. Magnification and ease of use are generally limited by cost. The microscope is used extensively for drilling craniotomies, dural resection, and visually monitoring the entry of the electrodes into the cortex. A variety of medications should be prepared and dosed, based on the animal’s weight, and loaded into syringes prior to the day of surgery, minimizing the need to calculate and measure at that time. These include: 1. Steroids—given perioperatively to reduce inflammation and brain swelling. 2. Antibiotics—given prophylactically before surgery and throughout the operation via IV. 3. Fentanyl—an opioid analgesic that is also used as an adjunct for maintenance of anesthesia with isofluroane. It works well to suppress respiratory drive in a hyperventilating animal. 4. Epinephrine and Atropine—for cardiac emergencies. For details regarding veterinary anesthesia medications and supplies, please refer to Hellebrekers and Booij (Hellebrekers and Booij 2001). For electrode sterilization, electrodes are carefully packed to avoid contact of the tips with the packaging and damage to the microwires. They are sterilized in an oxide ethylene gas chamber. This step must be completed before the day of surgery to allow adequate time for the sterilization cycle and process. The success of a surgery depends upon a number of personnel working together in a coordinated fashion. It is useful to develop a “flight plan,” so that everyone knows what to expect. A typical schedule of our primate surgery day is as follows: Animal arrival, last minute prep, attaching electrodes for EKG, O2, and temperature monitoring Induction of anesthesia and intubation, placement of IV


34


Fix the animal in the stereotaxic, create sterile field, drape microscope, drill, etc. Exposure and marking craniotomies Drilling of craniotomy and insertion of bone screws Lowering electrodes, allowing approximately 1 h per array Closure and waking of the monkey

PREOPERATIVE ANIMAL PREPARATION Only animals in good overall health should be considered for surgical implantations. Older animals with less cardiopulmonary reserve are more likely to develop postoperative problems such as pulmonary edema. We typically shave the animal’s hair on the arms, chest, tail, and head to allow easy application of electrocardiograph (EKG) leads and oxygen-saturation monitors. This can be done the day before surgery to save time. Food is withheld the day before surgery to lessen the chance of vomiting and aspirating into the lungs during administration of anesthesia and placement of the endotracheal tube. Cortical localization should be planned well in advance. Atlas information and published coordinates of the areas of interest should be studied carefully. It is important to consider the depth of the cell layer of interest as well as the surface conformation—areas deep within a sulcus or very near the midline where the saggital sinus runs may be technically difficult to access. It is important preoperatively to have a three-dimensional idea of where the electrodes and connectors, head posts, and ground wires will fit on the surface of the skull. Electrodes with wires that are “offset” or angled may be necessary to access two areas, such as M1 and PMd that are close together. This issue is magnified when working with smaller species such as owl and squirrel monkeys. Unlike rodents, there is more variability in the size, symmetry, and proportions of primate brains, making it important to have other backup methods in place to verify the locations of critical landmarks and structures. Preoperative magnetic resonance imaging (MRI) or computed tomography (CT) localization is one way to address this issue. In this paradigm, adapted from human functional-neurosurgical techniques, fixed markers such as a halo attached to the skull or implanted markers (e.g., under the scalp) can be referenced on scans relative to known sulci and gyri, thereby allowing the investigator to generate preoperative coordinates to find areas of interest (Scherberger et al., 2003).

ANESTHESIA TECHNIQUES AND INTRAOPERATIVE MONITORING General anesthesia consisting of inhaled isoflurane through an endotracheal tube and supplemented by intravenous narcotics, generally fentanyl, is the mainstay of anesthesia. Details can be found in veterinary textbooks. With early surgeries, we had difficulty with unpredictable swelling or retraction of the brain. Smaller species such as owl and squirrel monkeys seem more prone to this than rhesus monkeys. These smaller species have a tendency to hyperventilate when intubated and under anesthesia, resulting in blowing CO2 and causing significant brain retraction. In order to address these issues, we have instituted use of a ventilator. End tidal CO2 measurements are monitored continuously as they correspond © 2008 by Taylor & Francis Group, LLC


35

well with swelling and retraction of the brain; CO2 being the main driver of vascular dilation and constriction. We adjust the respiratory rate and tidal volume to keep CO2 in the physiologic range. If brain swelling or retraction occurs, adjustments to the ventilation can be made to counteract this effect by manipulating the end tidal CO2. Intravenous narcotics such as fentanyl are a useful adjunct to depress respiratory drive and counteract hyperventilation. The various monitors and electrical devices used during surgery create a significant amount of electrical noise. Some devices can be run on batteries, thereby decreasing electrical noise. A strategy should be developed allowing basic monitoring that minimizes noise during the critical electrode-lowering phase of surgery. Both core body temperature and depth of anesthesia are known to affect spontaneous neuronal firing. Core body temperature should be as close to normal as possible during the critical electrode-lowering steps when neuronal firing is being recorded. The depth of anesthesia is more difficult to tightly control. We have found that during the first steps of the operation, we can get a sense of a target heart rate that is indicative of an adequate level of anesthesia and a heart rate above which the animal is likely to begin moving. We titrate the anesthetic carefully during electrode lowering and neuronal recording to target this number. Other possible solutions to this problem include the use of paralytics to ensure the monkey does not move while temporarily lightening the anesthetics and simultaneously using ketamine anesthesia.

ELECTRODE SPECIFIC FOR PRIMATES Although an in-depth discussion of various electrode designs can be found elsewhere in this book, it is worth mentioning some of the important general features of electrodes intended for primates. In our paradigm, the electrodes are fastened to the skull rigidly after implantation, relying on the surrounding bone to maintain their place in the brain. The appropriate length of electrodes for smaller species such as owl and squirrel monkeys is 5–7 mm. For rhesus macaques, 6–8 mm is appropriate. One should expect variability in depth of the skull across different areas of the brain.

IMPLANTATION TECHNIQUES Exposure The surgical technique begins when the animal is stable under general anesthesia, fixed in the stereotaxic, and the skull vertex area is shaved, prepped, and draped in sterile fashion (see Figure 2.2). The skin incision can either be midline with extensions laterally to allow adequate exposure, or the skin can simply be cut in an oval pattern and removed—the scalp is very forgiving in its healing properties. Subperiostial dissection of the soft tissues is essential to assuring a good bond between the skull cap and skull. All soft tissues down to the cortical bone of the skull must be removed. Dissection laterally and posteriorly should extend well beyond proposed craniotomies. Laterally, the temporalis muscle and its attachments must be removed to expose hand somatosensory and motor areas. The soft tissues are then


36


FIGURE 2.2 Rhesus monkey view of the vertex after sterile prep and drape is complete with Ioban ™ iodoform antibacterial drape on shaved scalp.

protected with gauze and tacked back using stitches to maintain the exposure. The convexity of the skull is then prepped. Once dry, the border of the soft tissues and skull can be demarcated using a protective tape-type barrier to avoid leakage onto the soft tissues of the caustic materials used in the next step (see Figure 2.3). Next, hydrogen peroxide is used to remove all remaining shreds of soft tissue from the skull. Bleeding from the bone is stopped with cautery. Acid solutions and mechanical instruments such as a scalpel can be used to etch the outer cortical bone and to confirm that no soft tissue remains. The reason for doing this is to minimize or prevent a soft-tissue layer from developing between the skull cap and skull, which can be a nidus for infection and source of loosening of the head cap. The importance of careful skull cleaning cannot be overemphasized (see Figure 2.4). Cortical Localization The midline and interaural line is marked on the skull. Using predetermined coordinates based on these landmarks, the locations of the areas to be implanted are marked on the skull. Based on the size and orientation of the electrodes, craniotomies are outlined, taking into consideration how the connectors will be oriented. Bregma is highly variable in primates, and is not considered a reliable landmark for cortical localization (see Figure 2.5.) Drilling of Craniotomies Using a high-speed air micro drill adapted from dental applications, and the microscope, the craniotomies are drilled carefully respecting the dura. To verify the



37

FIGURE 2.3 After removal of the scalp in an area wide enough to allow electrode placement, the skull is thoroughly cleaned with hydrogen peroxide and sharp instruments to remove all soft tissue from the scalp down to the outer cortex of the bone.

FIGURE 2.4 Skull surface after satisfactory cleaning. A bony lesion is evident in the center. All blood has been cleared. All periosteum has been removed.


38


FIGURE 2.5 Predetermined coordinates are mapped onto the surface of the skull, in this case, based on interaural line and midline for the planned craniotomies, in this case bilateral M1 and left S1.

skull-based coordinates, appreciating that individual animals vary, it is a good idea to drill the first craniotomy near a prominent sulcus that is easy to distinguish. If the sulcus is not clearly identifiable through the dura, the dura is opened avoiding injury to any of the surface vasculature. Once the key landmark is found, the other craniotomies can be adjusted appropriately, if needed. This is a good backup mechanism to be sure that the areas of interest are definitively identified given the known variability in surface landmarks in primates (see Figure 2.6). An appropriately sized craniotomy should allow enough working space to be able to visualize cortical structures both for opening identifying landmarks and to facilitate the technical challenge of opening the dura as well as allowing visual confirmation, if possible, of electrode penetration. Thicker bone will warrant larger craniotomies to avoid working in a “hole.” A 1 mm margin around the footprint of the electrode is generally reasonable. Very large openings can cause significant leakage of CSF and distortion of the cortical surface, both of which are detrimental to the technique. Ideally, the opening allows space for the electrodes, while leaving the brain underneath in its native position. CSF shifts, or herniation of large craniotomies, may change the local mechanical forces enough to impede optimal electrode placement and stabilization. Too big a craniotomy also makes it difficult to secure electrodes, because the technique relies on rigid attachment to the surrounding cranium. When all the craniotomies are drilled, securing skull screws are placed. A variety of options exist ranging from small stainless steel screws available at the hardware store to specialized titanium skull screws for human applications. Although both work, the ease of use and time savings of the self-drilling, self-tapping titanium



39

FIGURE 2.6 Operating room setup with sterile draping. Dura is opened microsurgically as depicted on projected plasma screen.

screws make them our first choice. “T-bolts” and other devices to secure the head cap to the skull can be employed to enhance security. More important than discussing specific fixation techniques are the general principles that are important for obtaining secure fixation of the head cap. It is well known that if too much stress is applied to a bone or screw interface, the bone will gradually resorb, connective tissue will form, and the screw will loosen (Betelak, Margiotti et al., 2001). For bony integration to be achieved, for “oseteointegration” to occur, at least 3 months without significant mechanical stress is required (Worthington, 1994). Generally, 10–16 3.5 mm screws are adequate for holding a head cap in place but if a head post is required, more screws are necessary. Because of spatial constraints, head posts used to fix animals during experimental sessions are often included in the same dental-acrylic mass as the electrode arrays. As described by Betelak et al., in such a paradigm, the screws must be specially positioned and time for osteointegration must be allowed (Betelak, Margiotti et al., 2001). If a head post is attached to the same head cap that holds the electrodes, a waiting period of 3 months should be taken into consideration in planning postoperative experiments with the head fixed. Skull screws of any kind should be carefully tailored to the depth of the bone. Placing them too deep can have devastating consequences such as piercing the brain and causing bleeding, which can lead to seizures. Skull screws are used as grounds for the electrodes (see Figure 2.7). It is not advisable to drill craniotomies after opening the dura or after implantation of electrodes due to the bone dust that is distributed and the vibration that is caused. Spatial limitations make it nearly impossible.


40


FIGURE 2.7 After drilling of the craniotomies and removal of the dura, M1 and S1 are evidently separated by the central sulcus on the left. On the right side, cellulose sponges fill the craniotomy defect. Skull screws have been placed for grounding and fixation purposes. Note that the final position of the craniotomies has been modified as the surface anatomy became clear during the course of the drilling. The craniotomies have been sized to accommodate the electrodes.

Opening of the Brain Coverings Unlike in rodents, we have found that it is always necessary to open the dura in primates to achieve successful penetration of the cortex. Under high-power magnification, the dura is opened in all of the craniotomies. A fine needle can be bent at the tip and used as a hook to slide under the dura and incise it. The dura has several layers and consists of white fibrous material arranged in overlapping stands. Care must be taken to be sure all layers have been opened. Freely flowing CSF and crystal-clear surface vasculature indicate complete dural removal. A cupped curette is useful to cut the dura off at its attachment to the edge of the craniotomy to allow full exposure. Rarely there is bleeding from the dura. This can be dealt with either with cautery or Gelfoam soaked in thrombin to enhance hemostasis. If there is bleeding from disruption of the cortical surface vessels, placement of a Gelfoam sponge soaked in thrombin works well. Every effort should be made to avoid injury to the cortical surface. Opening the dura is challenging, and practice should be performed on rats to gain familiarity with the technique and to fine tune what works best in individual hands. At this point, the brain should be assessed and adjustments to the ventilation made, if needed, to correct any obvious retraction or marked swelling. A neutral brain that fills the opening of the craniotomy with occasional flow of CSF from under the craniotomy edge indicates a neutral position. If there is consistently even



41

1½ mm between the brain and the inner table of the bone, the brain is too retracted for optimal electrode insertion, and physiological adjustments should be made (see Figure 2.6). We have opened the pia through a variety of techniques including enzymatic digestion using collagenase, and mechanically with fine forceps. Although both are feasible and result in good recordings, despite visual disruption to the cortical surface, we have found that appropriate electrode spacing obviates the need to pursue these technically challenging and time-consuming steps. By adjusting the spacing of individual electrodes alone, we have been able to achieve consistent pial penetration. Other factors that can be manipulated to help achieve penetration are electrode diameter and the shape of the electrode tip. Electrode Lowering Once the dura is opened in all the craniotomies, the pia is protected from drying out with a small sponge (Gelfoam soaked in saline). The modified commercially available Kopf ™ stereotaxic arm is attached to the stereotax frame, and an electrode array is loaded, connected, and grounded to skull screws. The electrodes should be moving only in line with the direction of the lowering arm to avoid shearing forces upon entry. Every attempt is made to have the electrodes enter as perpendicular as possible to the surface of the brain realizing that the brain is not a flat surface. The order of electrode lowering is strategically planned. It is best to work outward from the middle craniotomy, thereby avoiding working in between delicate electrodes. Touchdown of the majority of the electrodes is marked as zero, and electrophysiological monitoring is begun. Noise issues are dealt with at this point. Anesthesia is titrated to target heart rate. The electrodes are then lowered approximately 100 µm at a time using microscopic visual guidance and electrophysiological monitoring of the electrode channels to determine entry into the brain (See Figure 2.8). A certain degree of dimpling occurs initially until the electrodes penetrate. We do not find any advantage to rapid or slow insertion. Occasionally, there is bleeding as cortical vessels are punctured if they cannot be avoided. This does not seem to affect intraoperative or postoperative recordings. Once characteristic firing of superficial cortical neurons is established in the majority of channels, the electrodes are lowered based on known depth of the target layer and most importantly electrophysiologic monitoring. We have found the quality of the intraoperative recordings, i.e., intraoperative single-unit firing, to be the most reliable predictor of good postoperative recordings. Once the ideal location has been achieved, small pieces of cellulose sponges are placed around the electrode-brain interface. Cyanoacrylate glue in gel form is used to rigidly fix the electrode array to the skull (see Figure 2.9). The connectors are then removed, and the electrode is delicately released from the stereotax. This sequence is repeated until all electrode arrays are implanted. Skull and Wound Closure Once the arrays are all in place, being held by cyanoacrylate glue, there may be some leakage of CSF from the craniotomy openings. It is important at this point to seal the craniotomies in a water-tight fashion. The area to be covered by dental acrylic must be free of blood, soft tissue, or other debris that collected over the course of


42


FIGURE 2.8 Electrode lowering proceeds under direct microscopic visualization with simultaneous electrode recordings to use both visual and neuronal firing to assess entry into the cortex and determine the ideal electrode placement.

FIGURE 2.9 Placement of the electrodes is complete. They are temporarily secured to the skull with cyanoacrylate glue visible in the center of the three electrodes. Ground wires have been connected to the skull screws.



43

FIGURE 2.10 The electrodes are now secured to the skull and a head cap is created out of multiple layers of dental acrylic. The electrodes are completely embedded to protect them. A secondary grounding wire is visible. The interface between the skull and dental acrylic is smooth without intervening soft tissue at any point. No CSF leak is evident between the bone and head cap. Tape protects the electrodes during the cementing process. A thread is built into the head cap to allow a second protective cap to be fitted to protect the electrode connectors when not in use.

surgery. It must also be dry to allow the best possible bond to the bone. The better the interface between the dental acrylic and the skull, the less potential for CSF leak, development of soft-tissue migration, and infection. The acrylic is applied in layers until the electrodes are embedded up to the top of the connectors to protect them from breakage during the animal’s normal movements (see Figure 2.10). Finally, a layer of antibiotic ointment is placed at the skull, scalp, and acrylic interface. Excess or devascularized skin is trimmed. Stitches or clips can be used to reapproximate skin edges, usually anteriorly and posteriorly. The scalp is very forgiving in its ability to heal.

POSTOPERATIVE CARE Postoperative care first requires ascertaining cardiopulmonary stability, and personnel and equipment available to deal with such issues. If necessary, a warm water blanket is kept inside the animal’s cage to maintain body temperature. Close observation for seizure activity and a plan for intervention are important. Seizures are infrequent and likely indicate an unanticipated bleed over the convexity, most likely from a misplaced skull screw. Neurological deficits are surprisingly rare and generally resolve over time. The animals are usually returned to their cages by 24 h postop and advanced to a regular diet as quickly as possible. Post-op narcotics are used


44


for the first postoperative day and in the subsequent postoperative days if necessary, however, narcotic drugs should not inhibit or delay the return to normal feeding, or depress respiration or other normal activities, which could potentially lead to further problems. After the second postoperative day, analgesia is carefully discussed with a veterinarian staff and if necessary is generally provided through NS antiinflammatory drugs or Tylenol. However, narcotics will be used if necessary. It has been found with human surgery that, in general, early mobilization and return to function results in the least complications postoperatively, and the same principle applies to animal surgery. The wound is cleaned when necessary, and a new layer of antibiotic ointment is applied. Stitches are removed from 10–14 days after surgery. Primates will be treated with antibiotics for 5 days after surgery.

CONCLUSIONS AND FUTURE DIRECTIONS Analyzing and making use of neuronal data requires the critical first step of obtaining such information. The brain-electrode interface remains one of the rate limiting steps in our ability to advance neuronal-ensemble physiology. Techniques to acquire brain signals from rodents and primates have advanced rapidly, and this is only the beginning. Over the last 18 years, we have worked on perfecting the surgical technique for implant of multielectrode arrays for chronic recording of brain cell activity. Our animals recover very well from the surgery, and they are usually back to their normal activities in 24–48 h without any signs of pain after the first postoperative day. Generally, there are no problems with postsurgical infection, and over time the animals are not disturbed by the presence of the head cap. However, we make every effort to continuously improve the surgical techniques for rodents and primates. Our goal is to be able to reach the relevant areas of the brain for each study with better and faster techniques, new materials, and identify new ways to improve healing of the tissues and skull bone, without compromising the efficiency of the recordings in the future. Especially for primate surgeries, preparation for surgeries involves extensive multipersonnel discussion of previous surgeries and ideas for future improvements. The design of the new electrode arrays, distribution of the microwires in the array, and materials used for the electrodes is carefully studied before each primate surgery and for each study involving rodents. The use of mutant mice has provided a good data base for the study of several neurological diseases that affect many people in many different ways. Data collected and analyzed from chronic recordings in these animals have provided new insights in the mechanisms and physiopathology of conditions like Parkinson’s diseases, depression, and psychoses (Costa, Lin et al., 2006; Dzirasa, Ribeiro et al., 2006). The possibility of studies on these animals is unlimited and may contribute to therapies or cures for progressive brain diseases that effect the lives of millions of people. It is our goal to search for more biological-compatible materials for the manufacture of the electrodes and head caps, to develop wireless systems capable of recording brain activity, allowing us to study neuronal activity of animals in their regular environment, and to develop new methods for placement of electrodes in the brain that will result in shorter surgeries, less brain traumatic injury or lesions on the soft



45

tissues and bone, and faster recovery from surgeries. Our ultimate goal is to find safer and better implantation techniques that can one day be applied to investigate and treat a variety of catastrophic neurological diseases that affect millions of people throughout the world.

ACKNOWLEDGMENTS We would like to thank all the members of the Nicolelis lab, who over the years have helped us to improve all the techniques described above. Also, we would like to thank our former members Erin Phelps and Kirsten Shanklin—Dewey, and our actual primate lab manager, Dr. Weiying Drake, for their constant support during our primate surgeries, supervision of our primate work, and for their extremely gentle care of our animals. Finally, a special thank to Susan Halkiotis, Gayle Wood, and Terry Jones, who work diligently for the success of our lab.

REFERENCES Baumans, V. and Remie, R. et al. (2001). Experimental procedures. Principles of laboratory animal science. Van Zutphen, L.F.M., Baumans, V., and Beynen, A.C. Amsterdam, Elsevier Science. Revised ed.: 313–333. Betelak, K.F. and Margiotti, E.A. et al. (2001). The use of titanium implants and prosthodontic techniques in the preparation of nonhuman primates for long-term neuronal recording studies. J Neurosci Methods 112(1): 9–20. Chapin, J.K. and Lin, C.S. (1984). Mapping the body representation in the SI cortex of anesthetized and awake rats. J Comp Neurol 229(2): 199–213. Costa, R.M. and Cohen, D. et al. (2004). Differential corticostriatal plasticity during fast and slow motor skill learning in mice. Curr Biol 14(13): 1124–34. Costa, R.M. and Lin, S.C. et al. (2006). Rapid alterations in corticostriatal ensemble coordination during acute dopamine-dependent motor dysfunction. Neuron 52(2): 359–69. Duke University Division of Laboratory Animal Resources (DLAR) (1995) Duke University manual for animal research. Durham, N.C. Dzirasa, K. and Ribeiro, S. et al. (2006). Dopaminergic control of sleep-wake states. J Neurosci 26(41): 10577–89. Faggin, B.M. and Nguyen, K.T. et al. (1997). Immediate and simultaneous sensory reorganization at cortical and subcortical levels of the somatosensory system. Proc Natl Acad Sci U. S. 94(17): 9428–33. Ghazanfar, A.A. and Nicolelis, M.A. (1997). Nonlinear processing of tactile information in the thalamocortical loop. J Neurophysiol 78(1): 506–10. Ghazanfar, A.A. and Stambaugh, C.R. et al. (2000). Encoding of tactile stimulus location by somatosensory thalamocortical ensembles. J Neurosci 20(10): 3761–75. Hellebrekers, L.J. and Booij, L.H.D. (2001). Anaesthesia, analgesia and euthanasia. Principles of Laboratory Animal Science. Van Zutphen, L.F.M., Baumans, V., and Beynen, A.C. Amsterdam, Elsevier Science. Revised ed.: 277–311. Kralik, J.D. and Dimitrov, D.F. et al. (2001). Techniques for long-term multisite neuronal ensemble recordings in behaving animals. Methods 25(2): 121–50. National Research Council (1996) Veterinary medical care. Guide for the care and use of laboratory animals. Washington, D.C., National Academy Press: 56–70. Nicolelis, M.A. and Baccala, L.A. et al. (1995). Sensorimotor encoding by synchronous neural ensemble activity at multiple levels of the somatosensory system. Science 268(5215): 1353–8.


46


Nicolelis, M.A. and Chapin, J.K. (1994). Spatiotemporal structure of somatosensory responses of many-neuron ensembles in the rat ventral posterior medial nucleus of the thalamus. J Neurosci 14(6): 3511–32. Nicolelis, M.A. and Dimitrov, D. et al. (2003). Chronic, multisite, multielectrode recordings in macaque monkeys. Proc Natl Acad Sci U. S. 100(19): 11041–6. Nicolelis, M.A. and Fanselow, E.E. et al. (1997). Hebb’s dream: the resurgence of cell assemblies. Neuron 19(2): 219–21. Nicolelis, M.A. and Ghazanfar, A.A. et al. (1997). Reconstructing the engram: simultaneous, multisite, many single neuron recordings. Neuron 18(4): 529–37. Nicolelis, M.A. and Ghazanfar, A.A. et al. (1998). Simultaneous encoding of tactile information by three primate cortical areas. Nat Neurosci 1(7): 621–30. Nicolelis, M.A. and Ribeiro, S. (2002). Multielectrode recordings: the next steps. Curr Opin Neurobiol 12(5): 602–6. Nicolelis, M.A.L. and Stambaugh, C.R. et al. (1999). Methods for simultaneous multisite neural ensemble recording in behaving primates. Methods for simultaneous neuronal ensemble recordings. Nicolelis, M.A.L. and Boca Raton, CRC Press: 121–156. Niedermeyer, E. (1993). Historical Aspects. Electroencephalography, basic principles, clinical applications and related fields. Niedermeyer, E. and Lopes da Silva, F. Baltimore, Williams and Wilkins: 1–14. Piccolino, M. (1998). Animal electricity and the birth of electrophysiology: the legacy of Luigi Galvani. Brain Res Bull 46(5): 381–407. Scherberger, H., Fineman, I., Musallam, S., Dubowitz, D.J., Bernheim, K.A., Pesaran, B., Corneil, B.D., Gilliken, B., and Andersen, R.A. (2003) Magnetic resonance imageguided implantation of chronic recording electrodes in the macaque intraparietal sulcus. J. Neurosci Methods 130: 1–8. Shuler, M.G. and Krupa, D.J. et al. (2002). Integration of bilateral whisker stimuli in rats: role of the whisker barrel cortices. Cereb Cortex 12(1): 86–97. Worthington P., Lang, B.R., and La Velle, W.E., Eds. Osteointigration in dentistry: An introduction. Chicago: Quintessence Publishing, 1994: 12.


3

Technology for Multielectrode MicroStimulation of Brain Tissue Timothy Hanson, Nathan Fitzsimmons, and Joseph E. O’Doherty

CONTENTS Historical Highlights of Brain Stimulation.............................................................. 47 Principles of Cortical Microstimulation .................................................................. 49 Design Considerations ............................................................................................. 51 Hardware Implementation ....................................................................................... 52 References................................................................................................................ 55

HISTORICAL HIGHLIGHTS OF BRAIN STIMULATION From early arguments between Galvani and Volta in the late 1700s, to later experiments by Helmholtz in the mid-1800s, the scientific examination of the electrical excitability of biological tissues has a rich history (Galvani, 1791; Volta, 1800; Helmholtz, 1842; Hess, 1994), particularly regarding the use of electrical stimulation to induce focused cortical activation, which enabled the discovery of motor maps in different locations of the cortex. This tradition started in the late 19th century when German neuroscientists Fritsch and Hitzig used electrical stimulation in animals to generate crude motor cortex maps, demonstrating contralateral activation of muscles for the first time (Fritsch and Hitzig, 1870). Sir Charles Sherrington, one of the founders of modern neuroscience, later used more focused stimulation to form early maps of ape motor cortex (Brown and Sherrington, 1911). In the 1930s, Wilder Penfield, one of Sherrington’s students extended this work to human motor cortex, as he showed the existence of a spatial map of muscles in the body (Penfield and Boldrey, 1937). His rigorous examination demonstrated that the size of a body part’s cortical representation is proportional to the fine nature of that body part’s movements. Penfield’s work helped popularize the technique of electrical stimulation as a useful tool for extending understanding of the brain (Penfield and Rasmussen, 1950).


48


During the 1960s, Robert Doty performed a series of experiments that explored cortical stimulation of awake, behaving animals in a behavioral context. By stimulating a variety of cortical areas, Doty was able to operantly condition animals, using the microstimulation pulses as the conditioned stimulus, to perform limb flexion movements. This extraordinary body of work provided one of the earliest demonstrations that microstimulation might be used to inject information into the brain (Doty, 1965; Doty, 1969; Doty et al., 1956). In the early 1990s, Newsome and colleagues at Stanford University demonstrated some of the earliest examples of perceptual biasing using cortical microstimulation. Working with directionally selective neurons in the visual area MT, the researchers used microstimulation to modify neuronal firing rates, thus biasing psychophysical performance on a motion direction discrimination task. This not only exhibited a direct link between physiological properties and perception, but proved the usefulness of microstimulation current pulses as a means to affect perceptual judgments (Salzman et al., 1990; Salzman et al., 1992; Murasugi et al., 1993). More recently, Ranulfo Romo and his colleagues successfully demonstrated that cortical microstimulation in monkeys could mimic the sensory perception of flutter in these animals. Using macaque monkeys, they examined neurons in area 3b of primary somatosensory cortex that were presumed to be associated with Meissner’s corpuscles—neurons with responses that vary with tactile flutter frequency. After inserting microelectrodes near such neurons, a train of current pulses was delivered to the cortical tissue, and effectively substituting for the peripheral tactile stimulation the animal was trained to recognize. The monkeys then performed a frequency comparison task between peripheral and cortical stimulation. Performance on frequency comparisons between artificial (microstimulation) and natural (peripheral vibration) stimuli were indistinguishable from comparisons between two natural stimuli. Not only did this study help identify the specific nature of some of the cortical circuits in area 3b of the primary somatosensory cortex, but it served as a proof of concept for the idea that peripherally provided tactile information could be mimicked through direct cortical stimulation (de Lafuente and Romo, 2005; Romo et al., 2000; Romo et al., 1998). Recently, Graziano and colleagues used cortical microstimulation to augment our understanding of the organization of motor and premotor cortices by repeating classical mapping studies with a minor, yet ultimately important alteration. Rather than using the short durations (~50 ms) from previous stimulation studies, the researchers used longer durations (~500 ms), more in line with the actual duration of a typical monkey reach motion. These researchers observed that longer stimulations produced complex motions and postures, as opposed to the simple twitching elicited by short stimulations. The movements observed varied from stimulated area to area, and often the response coordinated disparate areas of the body, i.e., arms and mouth. In each of these stimulated movements, the initial position of the related body parts had no effect on the position moved to when stimulated, that is to say that stimulation was associated with a specific position rather than direction. Their results show that these motor areas do not have a sort of homunculus-type mapping, but rather have elements of a body-centered spatial map (Graziano et al., 2002).


Technology for Multielectrode MicroStimulation of Brain Tissue

49

Also in 2002, John Chapin and his fellow researchers at the State University of New York (SUNY–Downstate) successfully trained rats to navigate through a complex, three-dimensional terrain, following control cues delivered by cortical microstimulation. Microelectrodes were implanted in somatosensory cortical whisker representation areas to deliver left or right directional cues, and electrodes implanted in the medial forebrain bundle were used to deliver rewarding stimuli. Furthermore, the rewarding stimuli had the effect of urging the animal on, acting as a functional “forward” cue. In this way, the researchers were able to guide the rats’ movements to overcome a myriad of obstacles that would not typically be explored by rats. By using cortical microstimulation for a sequence of cues and rewards, Chapin and his colleagues have hinted at the potential bandwidth of microstimulation-based information delivery (Talwar et al., 2002). In fact, findings such as these helped lay the groundwork for feedback in the field of neural prosthetics. Future neuroprosthetics will not only need to decode and process the neural code, but also provide the somatosensory and proprioceptive inputs back into the system that our biological limbs naturally produce. The potential here for neural prosthetics is strong, offering an interesting hope for a streamlined decoding solution. Although much progress has been made towards an understanding of how microstimulation interacts with the brain and how it can be used for functional purposes, there is still much more to be learned. Importantly, we must probe the limits of this technique, determining just what kinds of signals it can be used to deliver. For example, to deliver the breadth of information that we typically receive from one of our own limbs, stimulation bandwidth will have to be expanded significantly. Experiments designed to elucidate some of these issues have been undertaken in the Nicolelis Lab at Duke University (Fitzsimmons et al., 2007). Owl monkeys in a twochoice task have shown long-term performance stability in a variety of stimulation discriminations, including amplitude, temporal, and spatiotemporal characteristics (Figure 3.1).

PRINCIPLES OF CORTICAL MICROSTIMULATION The means by which neurons can be stimulated is heavily reliant on the same component of the cell that allows action potentials to propagate along the length of an axon: the cell membrane. In typical physiological conditions, voltage-gated sodium channels in the membrane initiate the action potential and sustain it as it propagates by allowing an influx of positive ions when triggered by an incoming pulse in voltage. During extracellular electrical stimulation, current pulses are delivered via electrodes placed near the tissue of interest. The injected current can depolarize (or hyperpolarize) the membrane, opening (or closing) voltage-gated ion channels, thus rendering the cell more (or less) excitable. When electrodes are used for stimulation in the cortex, typically the number of neurons excited is large, as affecting a single neuron is often not enough for a behavioral response. Ideally, we would like to know the precise nature of what cells are activated by a given stimulus, however, there are a number of spatial effects, which prevent a complete understanding of stimulus effects. First and foremost, the brain


50

A

Methods for Neural Ensemble Recordings, Second Edition Microstimulation Waveform

D

Temporal Discrimination 150 msec 100 msec Cues right

50-200 µA

100 µsec

Discriminated from 300 msec

100 µsec

50-200 µA

200 msec

100 µsec

Cues left

10 msec ×3

B

E

150 msec Spatiotemporal Discrimination 100 msec EP1

Basic Task

150 msec 100 msec

Cues right

Discriminated from

C

Reversal Task

Cues right

Discriminated from EP1

No Stimulation Discriminated from

EP3 EP4

Cues left

No Stimulation

EP2

Cues right

EP2 EP3

150 msec 100 msec Cues left

Cues left

EP4

FIGURE 3.1 Temporal patterns of microstimulation. (A) In all sessions, stimulation pulses were delivered biphasically, with a cathodic phase preceding an anodic phase of equal amplitude. Pulse width, pulse delay, and frequency were kept constant at 0.1 ms, 0.1 ms, and 100 Hz, respectively. For all tasks except the one in which psychometric curves were measured, current amplitudes were held constant at 0.1 and 0.15 mA, for monkeys 1 and 2, respectively. To construct the psychometric curves, we varied the stimulation currents between 0.05 and 0.2 mA. Pulse bouts of 150 ms with 100 ms delays were used in the basic (B), reversal (C), and spatiotemporal (E) tasks, whereas a second stimulation waveform consisting of 300 ms pulse bouts with 200 ms delays was also used in the temporal-discrimination task (D). By keeping the low-level stimulation parameters constant between the two cues in the temporal and spatiotemporal tasks, the absolute charge injection was kept constant. In panel E, electrode pairs (EP1 through EP4) designate the sequential stimulus and ground electrode pairs along the linear electrode array. (Fitzsimmons, N.A., Drake, W., Hanson, T.L., Lebedev M.A., and Nicolelis, M.A.L. (2007). Primate reaching cued by multichannel spatiotemporal cortical microstimulation. J Neurosci 27(21) 5593–5602.)

is neither an isotropic nor a homogenous medium. The resistance of the brain to the spread of current depends on several factors including temperature and the material properties of the tissue in that region (gray matter, white matter, CSF, etc.). Furthermore, the local characteristics of the tissue being stimulated also have significant effects. For instance, axon diameter, whether or not the axon is myelinated, axon orientation with respect to the electrodes, and whether the soma or axon is closest to the site of stimulation all affect the excitability of a given neuron. However, with increasing stimulation strength, many of these effects are averaged out, and the area of stimulation approximates a sphere in most cortical regions.



51

Typically, stimulation is accomplished using microelectrodes of various materials including stainless steel, platinum, platinum-iridium, and tungsten. Electrodes can be coated with a variety of compounds, which serves not only to insulate, but also to increase biological compatibility. It is important, especially in the case of chronic electrode implantation, that the brain’s injury and immune responses to the electrodes are minimized. A strong injury response can lead to encapsulation, and furthermore to a loss in the electrodes effectiveness for both measurement and stimulation. Ideally, the interface between the uninsulated metal tips of these microelectrodes and the tissue of the brain they are placed in approximates a Helmholtz doublelayer capacitor, drawing excess charge in the electrolytic medium to the surface of a layer of water along the electrode. Typically, stimulation electrodes are used in pairs, producing changes in extracellular voltage as current on the order of 10–100 μA is passed between them through the brain. There are two primary types of stimulators in-use for cortical microstimulation: constant-current stimulators and constantvoltage stimulators. Given that stimulation occurs at each neuron as a reaction to changes in extracellular voltage, it might seem that constant-voltage stimulators should be better suited, however, the aforementioned capacitive relationship of the electrode–tissue interaction means that extracellular voltage is actually a function of the electrode current. If a constant-voltage stimulator is used, the fluctuations in resistance due to local variations in brain composition can have large effects on the actual current, which passes through the electrodes, making the true nature of the stimulation hard to predict. The stimulus waveform used in stimulation is also an important consideration. For extracellular stimulation, a quick (10 s to 100 s of μs) rectangular cathodic (negative) current pulse is considered most effective at eliciting neuronal action potentials because it quickly depolarizes cells to threshold before any significant channel inactivation can occur, thus minimizing the necessary charge injection. However, monophasic-stimulation pulses, when repeated, can result in charge buildup at the electrode or tissue interface, leading to electrode corrosion, tissue damage, and eventually lesions. As a result, most stimulators operate in a charge-balanced biphasic stimulation mode, where a square cathodic pulse is followed by an equal-sized anodic pulse. Although the second pulse can actually impair the ability of the stimulation to cause excitation, the prevention of damage to the electrode and surrounding tissue is a necessary step for most sorts of stimulation. While charged-balanced biphasic current pulses offer an effective, damage-minimizing means for electrical stimulation, there is certainly room for improvement. For instance, there is some evidence that slightly imbalanced pulses actually minimize damage over the long term. Given that chronic, long-term usage of electrodes for microstimulation will be needed to provide feedback for such devices as neural prosthetics, new ways to minimize damage or increase stimulation efficiency are important avenues of future stimulation research.

DESIGN CONSIDERATIONS As mentioned above, stimulators are primarily regulated in terms of either current or voltage, and limited in terms of the complementary voltage or current. Much like


52


for recording, the range of stimulator regulation and the parameters of its limits are dictated by the size and impedance of the electrodes to be used, the size and nature of the area to be stimulated, the desired complexity of the stimulation, and power or channel requirements. Each of these is determined by the intended application of electrical stimulation. There are many, including: functional electrical stimulation (FES) for control of muscles in paralyzed patients; stimulation of the spinal cord for similar muscle control, deep-brain stimulation (DBS) for treatment of the symptoms of Parkinson’s disease, such as akinesia and tremor; abatement of chronic pain; and sensory prosthetics such as cochlear implants. Applications that use large electrodes, like DBS and FES, employ higher current levels, (1–40 mA), whereas cortical electrodes, typically smaller, therefore have higher impedances and need lower currents.

HARDWARE IMPLEMENTATION In this section we focus on intracortical microstimulation (ICMS) applications, where electrode impedances range from 0.5–2 MΩ and effective stimulation currents are between 40–180 μA. The high impedance of the electrodes necessitates relatively high voltages, typically 50–100 V in many systems, though care must be taken not to exceed the often-low breakdown voltage of the thin insulation on microwires. In ICMS, current is usually delivered in 100 μs cathodic-first charge balanced biphasic pulses at a frequency of one to several hundred hertz. The required current, voltage, and speed are not difficult to achieve with modern electronic systems, and many commercial off-the-shelf product can satisfy the conditions. However, the other requirements of a laboratory or clinical system—such as size (as in subcutaneous FES), power (as in DBS, where the stimulator is powered by an implanted battery), complexity of stimulation train, and number of electrodes—enforce more rigorous requirements on stimulation technology. For experimental ICMS, the latter two requirements are pivotal, a design that satisfies these and allows dense-multichannel stimulation is outlined below. We have investigated two methods of sequencing control in microstimulation: one that is based on a program running on a desktop computer, and a second that uses an embedded microprocessor. In the former, a general-purpose program controls the isolation hardware through one or more National Instruments data acquisition (DAQ) cards (see Figure 3.2). The program updates the DAQ buffers every 100 μs; within this interval, the program can specify stimulation activity based on pulse, time, or channel frequency, which in turn are generated autonomously or as a consequence of recorded neural activity or behavioral events. Four or eight stimulus isolation units are assembled on a circuit board, providing independent and isolated monopolar or bipolar stimulation per electrode. An alternate form of control is provided by embedded microprocessors, which can autonomously produce the pulsed stimulation train, often used in experiments (see Figure 3.3). This stimulation pattern is described by five parameters: (1) the pulse width, (2) the primary period, which determines the frequency of the stimulation pulse, (3) the duration of the pulse, (4) the secondary period, which determines the period between pulses, and (5) the total length of the train. The parameters are



53

PI

C1 8F 45 2

100µs

FIGURE 3.2 Schematic of the microstimulator apparatus.

a

b

c

d

e

FIGURE 3.3 Pulsed stimulation train features. (1) pulse width, (2) primary period, (3) pulse duration, (4) secondary period, (5) total duration of stimulation.

held in the memory of the microprocessor and are set via serial control. Each microprocessor can control up to eight channels simultaneously with a resolution of 100 μs, and is mounted on the PCB with the isolation units. The stimulus train is initiated either with a byte through the serial port or a transistor-transistor logic (TTL) pulse applied to hardware interrupt pins on the microprocessor. There are four elements to each channel of the stimulus isolation units: the power supply, current regulator, voltage regulator, and optical switches (see Figure 3.4). The stimulus isolation unit uses a 1 W miniature DC-DC converter to provide isolated power per channel. Although it is not strictly necessary to have an isolated supply on every channel with monopolar stimulation (due to the common ground), this allows greater freedom in using bipolar stimulation. A high-voltage linear regulator controls the voltage across the optical switches and is adjusted via an serial-addressable digital potentiometer. In turn, a metal oxide semiconductor field effect transistor (MOSFET) regulates the current. Current is controlled through a MOSFET biased in a feedback loop with an operational amplifier (op-amp). The op-amp adjusts the gate voltage to keep the MOSFET in the saturation regime and the voltage across a current sensing resistor, Rfb, equal to a reference set by a digital potentiometer. This simple feedback system has been found to work very well in practice, though the feedback resistor must be tailored to the range of intended-current output; a smaller resistor would allow the isolation unit to be used with larger, lower impedance FES


54

Methods for Neural Ensemble Recordings, Second Edition Digital Potentiometer

Clock

LinReg TI TL783

SPI Data

Wiper Register

Address (3)

Vcc

+5v

+96v Isolated 1W DC-DC Converter

gnd

LinReg TI TL783

+5v Bipolar Electrodes

+control

–control Digital Potentiometer SPI Wiper Register

+ – Rfb = 3.3K 1%

Address (3)

FIGURE 3.4 Engineering schematic of the microstimulator circuitry.

or DBS electrodes. Current is switched to the electrodes via four optical complementary metal-oxide semiconductor relays arranged in an H-bridge configuration. Control from the switches is via two lines from either the microprocessor or DAQ card. The other two control lines (serial clock and data) for the digital potentiometers are shared between the four isolators per board (the three address bits per potentiometer allow eight addresses). It has been found that the digital potentiometers are fast enough to change current or voltage regulation within a biphasic pulse when the isolator is controlled through the DAQ card, permitting charge balancing via a brief, high-current negative pulse followed by a longer, slower positive recovery. The built-in microcontroller cannot do this, nor can it vary the time between positivenegative phases. Instead, the time between negative and positive pulses is determined by the turn on or turn off asymmetry in the opto-CMOS relays. To prevent charge imbalance due to this effect, a small ceramic capacitor should be put in series with the electrode.



55

Presently, we have only tested the ICMS system using the embedded microprocessor, though a prototype of the DAQ-controlled digital isolation units has recently been completed. The system has proven to be compact, reliable, and expandable to as many channels as required.

REFERENCES Brown T.G. and Sherrington C.S. (1911). Observations on the localisation in the motor cortex of the baboon (“Papio anubis”). J Physiol 43(2), 209–218. De Lafuente, V. and Romo, R. (2005). Neuronal correlated of subjective sensory experience. Nat Neurosci 8(12), 1698–1703. Doty, R.W. (1965). Conditioned reflexes elicited by electrical stimulation of the brain in macaques. J Neurophysiol 28, 623–640. Doty, R.W. (1969). Electrical stimulation of the brain in behavioral context. Annu Rev Psychol 20, 289–320. Doty, R.W., Larsen, R.M., and Ruthledge, L.T., Jr. (1956). Conditioned reflexes established to electrical stimulation of cat cerebral cortex. J Neurophysiol 19, 401–415. Fitzsimmons, N.A., Drake, W., Hanson, T.L., Lebedev M.A., and Nicolelis, M.A.L. (2007). Primate reaching cued by multichannel spatiotemporal cortical microstimulation. J Neurosci 27(21) 5593–5602. Fritsch, G. and Hitzig, E. (1870). Über die elektrische Erregbarkeit des Grosshirns. Arch Anat Physiol Med Wiss, 300–332. Galvani, L. (1791). De viribus electricitatis in motu musculari. Commentarius. De Bononiesi Scientarium et Ertium Instituto atque. Acad Commentarii 7: 363–418. Graziano, M.S., Taylor, C.S., and Moore, T. (2002). Complex movements evoked by microstimulation of precentral cortex. Neuron 34, 841–851. Helmholtz, H.L.F. (1842). De fabrica systematis nervosi evertebratorum. Berlin, D.Phil. thesis. Hess C.W. (1994). [Developments in neurophysiology in the 19th century] German. Schweiz Rundsch Med Prax. 83(16), 483–490. Murasugi, C.M., Salzman, C.D., and Newsome, W.T. (1993). Microstimulation in visual area MT: Effects of varying pulse amplitude and frequency. J Neurosci 13, 1719–1729. Penfield, W. and Boldrey, E. (1937). Somatic motor and sensory representation in the cerebral cortex of man as studied by electrical stimulation. Brain 60, 389–443. Penfield, W. and Rasmussen, T. (1950). The cerebral cortex of man: A clinical study of localisation of function. New York, Macmillan. Romo, R., Hernández, A., Zainos, A., and Salinas, E. (1998). Somatosensory discrimination based on cortical microstimulation. Nature 392(6674), 387–390. Romo, R., Hernández, A., Zainos, A., Brody, C.D., and Lemus, L. (2000). Sensing without touching: psychophysical performance based on cortical microstimulation. Neuron 26(1), 273–278. Salzman, C.D., Britten, K.H., and Newsome, W.T. (1990). Cortical microstimulation influences perceptual judgements of motion direction. Nature 346, 174–177. Salzman, C.D., Murasugi, C.M., Britten, K.H., and Newsome, W.T. (1992). Microstimulation in visual area MT: Effects on direction discrimination performance. J Neurosci 12, 2331–2355. Talwar, S.K., Xu, S., Hawley, E.S., Weiss, S.A., Moxon, K.A., and Chapin, J.K. (2002). Rat navigation guided by remote control. Nature 417, 37–38. Volta, A. (1800). On the electricity excited by the mere contact of conducting substances of different kinds. Philos Trans R. Soc (Lond.) 90: 403–431 (in French).


4

Strategies for Neural Ensemble Data Analysis for Brain– Machine Interface (BMI) Applications Miriam Zacksenhouse and Simona Nemets

CONTENTS Introduction.............................................................................................................. 57 Neuronal Encoding and Tuning Curves .................................................................. 58 Velocity Tuning Curves ................................................................................ 59 Spatiotemporal Tuning Curves .....................................................................60 Tuning to Spatiotemporal Patterns of Velocity............................................. 63 Neuronal Modulations .............................................................................................64 Movement Prediction ............................................................................................... 69 Ensemble Analysis................................................................................................... 71 Principal Neurons ......................................................................................... 71 Ensemble Permanence .................................................................................. 73 Linear Regression .................................................................................................... 74 Principal Component Analysis (PCA).......................................................... 74 Least-Square Solution................................................................................... 75 Regularization Methods ............................................................................... 75 Conclusions .............................................................................................................. 76 Appendix—Derivation of the Variance Relationship.............................................. 77 Variance Relationship for Doubly Stochastic Poisson Processes ................. 77 Cross-Variance Relationship for Doubly Stochastic Poisson Processes....... 79 References................................................................................................................80

INTRODUCTION The advance of BMIs was largely motivated by investigations of velocity encoding in single neurons during stereotypical reaching experiments. However, BMIs are designed to decode neural activity from an ensemble of neurons and direct general 57 © 2008 by Taylor & Francis Group, LLC

58


reaching movements. Hence, neural data analysis strategies for BMIs are required to: (1) analyze the neural activity from ensemble of neurons, (2) account for the dynamical nature of the neural activity associated with general reaching movements, and (3) explore and exploit other relevant modulating signals. Decoding neural activity can be performed either in a single stage or two stages. Two-stage decoding relies on a preliminary encoding stage to determine how the neurons are tuned to the relevant biological signals. Based on the estimated tuning curves, the neural activity across an ensemble of neurons can be decoded using either a population-vector, maximum likelihood estimation, or Bayesian inference (Pouget et al., 2003). The population-vector approach results in a linear relationship between the spike counts and the estimated biological signal, which can be estimated directly in a single stage using linear regression (Brown et al., 2004). This chapter focuses on single-stage decoding with linear regression, and in particular on two special challenges facing the application of linear regression to neural ensemble decoding during reaching movements (see “Movement Prediction”). First, given the dynamic nature of the decoded signal, it is necessary to include the history of the neural activity, rather than just its current spike count. Second, due to the correlation between the activities of different neurons (see “Ensemble Analysis”) and the activities in different time lags, the resulting regression problem is ill-posed and requires regularization techniques (see “Linear Regression”). Although neural decoding can be performed in a single stage, neural encoding is still important for investigating which signals are encoded in the neural activity. For this purpose, the notion of tuning curves is generalized to characterize how the neural activity represents the spatio-temporal profile of the movement. This analysis quantifies both the spatio-temporal tuning curves and the percent variance of the neural activity that is accounted by the movement profile (see “Neuronal Encoding and Tuning Curves”). For comparison, the percent variance in the neural activity that might be related to general neural modulations is assessed independently under the Poisson assumption (see “Neuronal Modulations”). These two-faced variance analyses provide a viable tool for quantifying the extent to which the neural code is effectively decoded, and the potential contribution of yet undecoded modulating signals. The strategies and algorithms described in this chapter are demonstrated using the neural activity recorded from an ensemble of cortical neurons in different brain areas during a typical target-hitting experiment with pole control as described in Carmena et al., 2003.

NEURONAL ENCODING AND TUNING CURVES The firing rates of cortical motor neurons represent a diversity of motor, sensory, and cognitive signals, and most notably the direction and speed of movement (Georgopoulos, 2000; Johnson et al., 2001; Georgopoulos et al., 1989; Paz et al., 2004). Neuronal encoding of specific movement-related signals, including movement direction, and speed, has been characterized in terms of the tuning properties of the neurons (Georgopoulos et al., 1986, Ashe and Georgopoulos 1994). Most prominently, center-out reaching experiments indicated that the firing rates of single cortical motor neurons


Strategies for Neural Ensemble Data Analysis for BMI Applications

59

are broadly “tuned” to the direction of movement. Tuning curves represent the firing rate as a function of the direction of movement, and are well described by a cosine function of the angle between the movement direction and the preferred direction of the neuron. Detailed investigations suggest that the activity of directionally tuned cortical motor neurons is also modulated by the speed of movement, both independently from and interactively with the direction of movement (Moran and Schwartz, 1999). Furthermore, other scalar signals, including the amplitude of movement, its accuracy, the location of the target, and the applied force may also contribute to the firing rate modulation of cortical motor neurons (Johnson, et al., 2001, Alexander & Crutcher, 1990, Georgopoulos et al., 1992, Scott 2003). BMI experiments can be used to further investigate (Nicolelis 2001; 2003): (1) how individual neurons represent the spatiotemporal profile of the movement during free arm movements; (2) the potential modulations by yet undecoded signals; and (3) the distribution of correlated activity across an ensemble of neurons. This section focuses on the first issue whereas the last two are addressed in the sections “Neuronal Modulation” and “Ensemble Analysis,” respectively.

VELOCITY TUNING CURVES It is customary to determine the tuning of motor neurons to the velocity of movement during planar center-out reaching movements, where the direction of velocity is approximately constant during each reaching movement (Georgopoulos, 1986; Moran and Schwartz, 1999). Tuning curves that account for both the cosine-directional sensitivity and the effect of the speed of movement are of the form N aVm cos(Q Q PD ) bVm c E

(4.1)

where N is the number of spike counts during the movement, Vm and Q are the magnitude (speed) and direction of the velocity, a and QPD are the magnitude and preferred-direction of the directional tuning, b is the magnitude of the tuning to the speed, c is the mean spike count across different directions, and E is the modeling error. This formulation implies that the neural activity depends linearly on the x- and y-components of the velocity, Vx Vm sin Q and Vy Vm cosQ, as: N axVx ayVy bVm c E

(4.2)

where ax a sin Q PD and ay a cosQ PD are the tuning to the components of the movement. Equation 4.2 is in the form of a linear regression, so the tuning coefficients ax and ay can be estimated directly using linear regression between the neural activity and the velocity. The resulting coefficient of determination R 2 ( N , V ) describes the fraction of the total variance in the spike counts that is attributed to the velocity and provides a measure of the goodness of fit. Because the neural activity is highly noisy (see the section “Neuronal Modulation”), the resulting coefficients of determination are small, even if the tuning to the velocity is significant. Free arm movements involve temporal patterns, which cannot be captured only by spatial features. The formulation of the directional tuning should be generalized


60


to describe the spatiotemporal features of the movement profile. The neural activity is binned (typically with 100 ms long bins), and the mean velocity in each bin is computed. The tuning of the neural activity to the velocity at a particular lag is assumed to follow the cosine tuning of Equation 4.2 (Lebedev et al., 2005) as: N ( k ) ax (l )Vx ( k l ) ay (l )Vy ( k l ) b(l )Vm (k l) c(l ) E(l, k ) ,

(4.3)

where N(k) is the spike counts in the kth time-bin, Vx ( k l ), Vy ( k l ) , and Vm ( k l ) are the mean components and speed of the velocity in the (k+l)th time-bin, l is the relative lag between the velocity and the spike counts (positive or negative l corresponds to rate-modulations preceding or succeeding the velocity measurement, respectively), Al [ ax (l ) ay (l )] is the vector of directional tuning parameters with respect to the lagged velocity, b(l) is the speed tuning parameter, c(l) is a bias parameter, and E(l, k) is the residual error. The coefficient of determination of the single-lag regressions R 2 ( N , V ([l ])) for l = –L1,…L2, and the strength of the tuning A(l ) can be used to evaluate the strength of tuning and determine the most significant lag, i.e., the lag between the neural activity and the velocity that it is tuned to the most.

SPATIOTEMPORAL TUNING CURVES To account for the dependence of the neural activity on the velocity profile across several lags, the model is extended according to the multi-lag regression given by (Figure 4.1): L2

L2

N (k )

£

ax (l )Vx ( k l )

l L1

£

l L1

L2

ay (l )Vy ( k l )

£ b(l )V (k l) c E(k ) m

(4.4)

l L1

where L1 and L2 are the number of preceding and succeeding lags between the spike counts and the velocity, respectively. This model describes the tuning of the neural activity to the complete velocity profile in the surrounding window of [–L1 L2]. The coefficient of determination of the multi-lag regression R 2 ( N , V ([ L1, L2 ])) can be interpreted as the fraction of variance in the binned spike counts that is attributed to modulation by the spatiotemporal velocity profile. Expressed as a percentage, it is referred to as the percent velocity modulation (PVM). It is noted that because the velocity is slowly varying (compared with the bin size), the velocities in different lags are highly correlated. Thus, the multi-lag regression analysis of Equation 4.4 and the lag-by-lag analysis of Equation 4.3 would yield different regression parameters, as demonstrated below. The multi-lag regression analysis accounts for the correlation between the velocities in different lags and describes the tuning to the complete velocity profile. Furthermore, R 2 ( N , V ([ L1, L2 ])) cannot be approximated by the sum of the individual coefficients of determination of the single-lag regressions R 2 ( N , V ([l ])) for l = –L1,…L2. Thus, it is necessary to perform the multi-lag regression in order to quantify the PVM. The multi-lag regression of Equation 4.4 can be formulated in a matrix notation as:



61

FIGURE 4.1 Tuning to the multi-lag velocity profile: the neural activity is related to the velocity trajectory in the surrounding window. Top panels: the x (right) and y (middle) components of the velocity, and the speed (left) in 100 ms bins. Bottom panel: the binned neural spike counts.

N [V x ( L1 )

V x ( L2 ) V y ( L1 )

V y ( L2 ) V m ( L1 ) V m ( L2 ) 1]C E

(4.5)

XV C E where N = [N(L1 + 1)…N(T – L2)]T, and V k (l ) [Vk ( L1 1 l ){Vk (T L2 l )]T (the index k = x, y, m indicates the x- and y- components of the velocity and its magnitude, respectively) are (T – L1 – L2) w 1 vectors of spike counts and velocity components, respectively, 1 is a (T – L1 – L2) w 1 vector of 1’s, and C = [ax(–L1) ax(–L1 + 1)…ax(L2) ay(–L1)…ay(L2) b(–L1)…b(L2)c]T is a vector of regression coefficients. The optimal least-square solution of Equation 4.5 is sensitive to measurement noise and become unstable when the condition number of the matrix is large (see the section “Least Square Solution”). During typical experiments the condition number of the matrix X V is on the order of 106 (Figure 4.4). Thus proper evaluation of Equation 4.5 requires regularization methods as detailed in the section “Regularization Methods.” The regression parameters of a typical spatiotemporal tuning curve, computed using regularization is shown in Figure 4.2, and compared with the lag-by-lag tuning curve. The parameters ax and ay are shown separately as a function of the lag


62


Regression Parameter

Tuning to Vx

Tuning to Vy

0.3

0.2

0.2

0.1

0.1

0

0

–0.1

–0.1 –1

–0.5

0

0.5

–0.2 –1

1

200

0.3

100

0.2 0.1 0 –1

0

0.5

1

Preferred Direction

0.4

Degrees

Velocity Tuning Index

Directional Tuning

–0.5

0 –100

–0.5

0 Lag (sec)

0.5

1

–200 –1

–0.5

0

0.5

1

FIGURE 4.2 Tuning to velocity signals (upper left–Vx, upper right–Vy) and corresponding velocity tuning index (bottom left) and preferred direction (bottom right) based on lag-by-lag (dashed), and multi-lag velocity profile (using t-SVD capturing 95% of the variance). Each lag is 100 ms long.

(top left and right, respectively), where negative lags are preceding and thus predictive, whereas positive lags are succeeding and thus reflective. The magnitude of the directional tuning, i.e., ax2 (l ) ay2 (l ) , and the estimated preferred direction (PD) at each lag, PD(l) = tan–1(ay /ax)are depicted in the bottom left and right panels, respectively. The results based on the lag-by-lag tuning analysis defined by Equation 4.3 are shown for comparison (the values of the resulting coefficients are scaled down by a factor of 5 to compensate for the smaller number of coefficients in each regression). It is evident that the lag-by-lag analysis provides only a coarse and highly smoothed estimate of the underlying multi-lag tuning curve. The estimates of the PDs are reliable only in lags where the magnitude of the directional tuning is large and, thus, the fluctuations in other lags are meaningless. For the lags in which the directional tuning is significant, the estimated PD is the same independent of the method and relatively constant across the lags. The spatiotemporal tuning curve expressed by Equation 4.4 describes how the neural activity is modulated by the velocity profile, and the associated regression



63

120 100

#Neurons

80 60

Mean (PVM) = 3.7% Median (PVM) = 1.8% Std (PVM) = 4.7%

40 20 0

0

5

10 15 20 % Velocity Modulation (PVM)

25

30

FIGURE 4.3 Distribution of percent velocity modulation (PVM) across an ensemble of neurons recorded during one experimental session. PVM was computed using a multi-lag tuning curve estimated by regularization.

analysis quantifies the percent variance attributed to the velocity modulation (PVM). The distributions of the PVM across an ensemble of (183) neurons recorded during a typical target-hitting experiment are depicted in Figure 4.3. Only few neurons exhibit velocity modulations in excess of 20% of their variance. On average, the velocity-profile accounts for only 3.7% of the total variance, and for half of the neurons it accounts for less than 1.8%.

TUNING TO SPATIOTEMPORAL PATTERNS OF VELOCITY The regularization methods used to compute the spatiotemporal tuning curve are based on decomposing the spatiotemporal velocity profile, given by the matrix X V , into its principal components (PCs) (see the section “Principal Component Analysis (PCA)”). The PCs of the velocity are uncorrelated linear combinations of the lagged velocity components. The first PC accounts for maximum fraction of the variance in the velocity profile, and the succeeding PCs account, in order, for the maximum fraction of the remaining variance. Figure 4.4 shows the variance accounted for by individual PCs, and the accumulated variance accounted by all the initial PCs. It is evident that the late PCs account for a negligible fraction of the total variance. In particular, the first 25 PCs in this example already account for 95% of the variance. The weights of the linear combinations that define each of the PCs, are the principal directions of the covariance matrix of the velocity profile, and represent the principal velocity-patterns (see the section “Principal Component Analysis (PCA)”). The initial principal velocity-patterns in a 1.9 s window during a typical target-hitting experiment are shown in Figure 4.5. Some of the principal velocity-patterns are easily interpreted in terms of directional velocity and acceleration: The first two principal patterns describe low-frequency directional acceleration and directional velocity, respectively. The third principal pattern describes a higher-frequency directional


64


% Variance

20 15 10 5 0

0

10

20

30 PC Index

40

50

60

0

10

20

30 PC Index

40

50

60

% Variance

100

50

0

FIGURE 4.4 Percent variance accounted for by individual PCs (top) and accumulated percent variance accounted for by the initial PCs (bottom) of the movement during one session of a target-hitting experiment with pole control. The indicated 95% accumulated variance (dashed line) is accounted for by the initial 25 PCs.

acceleration, with zero center velocity preceded and succeeded by negative and positive directional velocity, respectively. The fourth principal pattern describes a pause between movements in the same direction. The fifth principal pattern describes low-frequency speed, whereas the sixth describes a change in direction of movement. Other components describe higher frequencies of the velocity and acceleration profiles, including high-frequency speed (8th and 16th components) and higher-frequency acceleration (12th component). The correlated velocity components that compose the matrix X V (Equation 4.5) can be described alternatively by the uncorrelated velocity PCs. Each velocity PC describes the temporal evolution of the corresponding principal velocity-pattern defined. The correlations of the spike counts with these PCs represent tuning to spatiotemporal patterns of velocity. For example, the tuning of an M1 neuron to the principal velocity-patterns depicted in Figure 4.5, is shown in Figure 4.6. This neuron seems to be tuned to low-frequency directional acceleration (1st component) and velocity (2nd component), and high-frequency speed (12th component). However, the neuron is not selectively tuned to any single spatiotemporal velocity pattern.

NEURONAL MODULATIONS The velocity profile accounts for a fraction of the total variance in the spike counts of cortical neurons during arm movements, as suggested for example in Figure 4.3. The remainder of the variance may be attributed to either (1) the neuronal noise associated with the underlying firing activity or (2) other modulating signals, including



0.5

65

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

0 –0.5

Velocity Eigen Vector

0.5 0 –0.5

0.5 0 –0.5 0.5 0 –0.5 –1

0

1 –1

0

1 –1 Time (sec) Vx

Vy

0

1 –1

0

1

Speed

FIGURE 4.5 Principal patterns of velocity defined by the initial principal directions of the velocity and speed trajectories in 1.9 s windows during a typical target-hitting experiment.

FIGURE 4.6 Tuning to the spatiotemporal velocity patterns defined in Figure 4.4. Example based on the same neuron, which multi-lag tuning is depicted in Figure 4.2.


66


nonlinear velocity effects. These two possible affects can be differentiated by quantifying the percent variance that is attributed to the neuronal noise, while the remaining variance, is attributed to the other modulating signals. Statistically, neural spike trains are customary analyzed using the mathematics of point processes (Perkel and Bullock, 1968; Dayan and Abbott, 2001), where each point represents the timing of a spike. The simplest point process, the Poisson process is memory-less, i.e., the probability of spike occurrence is independent of the history of the spike train. The simplest Poisson process, the homogenous Poisson process, is characterized by a constant instantaneous spike rate, and thus is inadequate for describing firing rate modulations. Thus, the simplest point process that can describe rate modulations is the inhomogeneous Poisson process, which is characterized by time-varying instantaneous spike rate that is independent of the history of the spike train (Dayan and Abbott, 2001, Snyder 1975; Johnson 1996). Inhomogeneous Poisson processes in which the instantaneous spike rate is itself a stochastic process are referred to as doubly stochastic Poisson processes (Snyder 1975; Johnson 1996). In doubly stochastic Poisson processes, the probability of spike occurrence given the instantaneous rate is described by Poisson statistics. The spike trains recorded during arm movements can be considered as realizations of doubly stochastic Poisson processes because the instantaneous rate depends on a number of biologically relevant stochastic signals. These signals include, for example, the position and velocity of the arm and the muscle forces. These multiple signals affect the instantaneous rate according to the individual tuning of each neuron. Assuming that the spike trains are generated by doubly stochastic Poisson processes facilitates the analysis of their statistics, which is determined by two factors: stochastic changes in the instantaneous firing rate and Poisson probability of spike occurrence. The distribution of spike counts, Nb in bins of size b, is determined by the average instantaneous spike rate during the bin, ,b, and its statistics are related to the statistics of ,b according to (Snyder 1975; Appendix): E[ N b ] E[ , b ] Var[ N b ] Var[,b ] E[,b ]

(4.6)

Var[,b ] E[ N b ] The last relationship can be interpreted as a decomposition of the total variance in the spike counts into the variance of the underlying information bearing parameter, or rate-modulating signal, Var[,b], and the variance that would occur if Nb was generated by a homogenous Poisson process, E[Nb]. Thus, the variance of the ratemodulating signal is the excess variance of the spike counts beyond that of a homogeneous Poisson process. Furthermore, as the instantaneous rate of a homogeneous Poisson process does not vary in time, it cannot be modulated by any relevant biological signal and consequently its variance may be considered as noise. Thus, the signal-to-noise ratio (SNR) can be defined as: SNR


Var[ , b ] Var[ N b ] E[ N b ] F 1 Var[ noise] E[ N b ]

(4.7)


67

where F is the Fano factor, defined as the ratio between the variance and the mean of the spike count (Dayan and Abbott, 2001). In addition, the percent overall modulation (POM) is defined by expressing the variance of the rate-modulating signal as a percentage of the variance in the spike counts:

POM =

1 Var[ Λ b ] Var[ N b ] − E[ N b ]  ⋅ 100% =  1 −  ⋅ 100% ⋅ 100% =  F Var[ N b ] Var[ N b ]

(4.8)

Compared to the Fano factor, the POM emphasizes the variability due to the underlying stochastic modulations and distinguishes it from the high variability inherent in the Poisson process (Zacksenhouse et al., 2007a). The POM is zero for the homogeneous Poisson process and positive for the inhomogeneous Poisson process. However, if the underlying point process is not Poisson, the variance of the spike counts may be smaller than the mean, and the POM may be negative. It should be noted that negative values may also result from finite-length spike trains due to the variance of the estimate, as demonstrated in the following text. The distribution of the POM in the ensemble of 183 neurons analyzed in Figure 4.3 is depicted in Figure 4.7 (top). The POM of the recorded neurons was positive for 83% of the neurons in this ensemble. The distribution of the POM for simulated homogeneous Poisson processes having the same mean spike counts as the recorded neurons is shown in the bottom left panel, indicating that the standard deviation of the estimated POM is sPOM = 1.2%. The POM of the recorded neurons

Figure 4.7 Distribution of percent overall modulation (POM) across an ensemble of neurons recorded during one experimental session (top). Distribution of POM for simulated neurons, simulated as homogeneous Poisson processes having the same mean rate as the recorded neurons (left), or inhomogeneous Poisson processes having the same velocity tuning as the recorded neurons (right).


7046.indb 67

10/30/07 9:42:08 AM

68


% Velocity Modulation (PVM)

30 25 20

Coeff. of corr = 0.55 Slope = 0.22

15 10 5 0 –60

–40

–20 0 20 % Overall Modulation (POM)

40

60

Figure 4.8 Correlation between percent velocity modulation (PVM) and percent overall modulation (POM). Scatter plot of PVM and POM computed from an ensemble of neurons during the same experiment analyzed in Figure 4.3 and 4.7. Regression line (solid) and unit slope line (dashed) are superimposed.

(Figure 4.7, top) was above 2sPOM = 2.4% for 65% of the neurons, whereas only 6.5% of the neurons exhibited negative POM below –2sPOM = –2.4%. The POM provides a scale against which the PVM can be compared in order to determine the relative role of the velocity profile in modulating the firing rate. The scatter plot in Figure 4.8 depicts the correlation between the PVM and POM for the same ensemble of (183) neurons during the experiment analyzed in Figure 4.3 and Figure 4.7. The high correlation indicates that the activity of cortical neurons, which exhibited larger rate modulations, was, in general, better correlated with the velocity profile. As expected, the PVM is usually smaller than the POM, in agreement with the interpretation that the POM describes the percent variance attributed to overall modulations, including the velocity modulation. The slope of the linear relationship is only 0.22, suggesting that the PVM accounts for only a small fraction of the POM, and that additional signals, other than the velocity profile, modulate the neural activity (Zacksenhouse et al., 2007). The POM of the recorded neurons can be also compared to the POM of simulated neurons that are modulated only by the velocity profile (bottom right). The simulated neural activity is generated using inhomogeneous Poisson processes with a rate parameter derived from Equation 4.4 based on the estimated multi-lag tuning curves of the recorded neurons. When the activity is modulated only by the velocity profile, the statistics of the resulting POM distribution is similar to that of the PVM (Figure 4.3), but smaller than that for the recorded neurons. This comparison supports the above conclusion that the POM captures additional signals that modulate the neural activity.



69

MOVEMENT PREDICTION Neuronal analysis indicates that the spike counts generated by individual neurons encode the planned velocity, as detailed in the section “Neuronal Encoding and Tuning Curves.” BMI technology is based on decoding the neuronal activity and extracting the movement related signals, and in particular the planned velocity. Several decoding techniques are possible, including, linear (Weiner) filter, Kalman filter, and nonlinear filters. Typical BMI experiments indicate that the Kalman and nonlinear filters do not consistently outperform the linear filter (Carmena et al., 2003; Wessberg et al., 2000). Here we concentrate on describing the linear filter and its improvement using regularization methods, and demonstrated its superior performance. The movement signal of interest, like the velocity components, can be predicted by a linear filter of the spike counts recorded from an ensemble of neurons during the preceding time lags, according to: N

Mˆ ( k ) W 0

0

£ £ W (l )N (k l ) j

j

(4.9)

j 1 l L

where Mˆ ( k ) is the predicted movement signal (e.g., the components of the velocity Vx and Vy, or the grip force), W0 is the bias term, and Wj(l) is the weight given to the spike counts elicited by the j-th neuron during the preceding l-th lag. The filter described in Equation 4.9 is of the form of a moving average across multiple neurons, with a window determined by the number of lags L. The bias and weights are determined from the training section of the BMI experiment, in which both the neural activity and the movement signals are recorded, using the multi-variables regression given by: M (k ) W0

n

L

j1

l1

£ £ W (l )N (k l ) E(k ) j

j

(4.10)

where M(k) is the recorded movement signal, and E(k) is the residual error. The multi-lag multineuron regression of Equation 4.10 can be formulated in a matrix notation as: M [ N 1 (1)

N 1 ( L ) N 2 (1)

N 2 ( L ) N n (1)

N n ( L ) 1]C E

(4.11)

X N ,L W E where, M = [M(L + 1)…M(T)]T and Nj (l) = [Nj(l)…Nj (T – + l – L)]T are (T – L) w 1 vectors of the measured movement-signal and the properly lagged spike counts of the jth neuron, respectively, 1 is a (T – L) w 1 vector of 1’s, W = [W1(L) W1(L – 1)… W1(1) W2(L)…W2(1)Wn(L)…Wn(1) W0]T is a vector of regression coefficients, and n is the number of neurons. The coefficient of determination of the multi-variable regression of Equation 4.11, R 2 ( M , N ([ L,1])) , describes the fraction of variance in the measured movement-signal


70

Methods for Neural Ensemble Recordings, Second Edition 0.85 R2 Training R2 Testing

0.8

Fraction

0.75 0.7 0.65 0.6 0.55

0

100

200 300 Regularization Parameter

400

500

FIGURE 4.9 Effect of regularization on the fidelity of movement reconstruction and prediction on training (solid) and testing (dashed) records, respectively. The fidelity of the prediction is assessed by the coefficient of determination R2. The least square (LS) solution is obtained with a regularization parameter zero.

that is correlated with the linear filter and provides a measure for the fidelity of the reconstruction. However, in movement prediction we are interested in the ability to predict the movement-signal from new measurements of neural activity. Thus, the performance of the filter is assessed using testing records, which were not used to determine the filter coefficients. The quality of the prediction is evaluated on a testing record using the filter coefficients determined from a nonoverlapping training record. The fidelity of the prediction can be assessed using either (1) the coefficient of regression R( M , Mˆ ) between the predicted movement-signal Mˆ and the measured signal during the testing record M, or (2) The variance reduction VR( M , Mˆ ) given by VR( M , Mˆ ) 1

£ £

test test

( M i Mˆ i )2 ( M i M i )2

(4.12)

The optimal least-square solution of Equation 4.11 is sensitive to measurement noise and may become unstable when the condition number of the matrix is large (see the section “Least Square Solution”). During typical BMI experiments, the condition number of the matrix X N,L is on the order of 2000. The effect of different regularization parameters on the fidelity of velocity prediction, based on a 10 min training record and a 2 min testing record taken from a typical BMI experiment, is shown in Figure 4.9. It is evident that prediction can be improved by using proper regularization. The best coefficient of determination is obtained with L = 94, result2 ˆ ˆ ing in RL 94 ( M , M ) 0.676 (i.e., RL 94 ( M , M ) 0.82 ), implying that the predicted velocity (depicted in Figure 4.10, middle panel) accounts (explains) 67% of the variance of the measured velocity (Figure 4.10, top panel). In comparison, the least



71

Measured Velocity

20 10 0 –10

Regularized Prediction

–20

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

20 10 0 –10 –20

Kalman Prediciton

20 10 0 –10 –20 t(minutes)

FIGURE 4.10 Measured and predicted velocity using a linear filter with Tikhonov regularization (middle) or a Kalman filter (bottom). The coefficient of determination between the actual velocity and predicted velocity, is R2 = 0.676 and R2 = 0.593, respectively. 2 square (LS) regression, corresponding to L = 0, results in RLS ( M , Mˆ ) 0.593 (i.e., ˆ RLS ( M , M ) 0.77 ). Thus, with proper regularization, the prediction can capture an additional 8% of the variations in the velocity. The performance of a Kalman filter (Brown and Hwang, 1997; Wu et al., 2004) that was trained and tested on the same records of data is shown for comparison in the bottom panel of Figure 4.10. The resulting coefficient of correlation of R 2 ( M , Mˆ ) 0.6 (i.e., R( M , Mˆ ) 0.77 ) indicates that the Kalman filter performs as well as the leastsquares linear regression but underperforms the regularized linear regression.

ENSEMBLE ANALYSIS PRINCIPAL NEURONS The activity of the recorded neurons may be correlated either due to common modulating signals or due to correlation in the neural noise. Ensembles of neurons with correlated activity can be identified using principal components analysis (PCA), as


72

Methods for Neural Ensemble Recordings, Second Edition 4

% Variance

3 2 1 0

0

50

100 PC Index

150

200

FIGURE 4.11 Distribution of variance across different neural ensembles. Percent variance attributed to the principal components of the normalized neural activity recorded in one experimental session.

detailed in the section “Principal Component Analysis (PCA).” PCA transforms the sequences of normalized spike counts recorded from n neurons into an ordered set of n uncorrelated sequences, known as principal components (PCs). Each PC is the weighed linear combination of the n normalized (zero mean and unit variance) spike count sequences, and thus may be considered as the normalized activity of a principal neuron. The PCs are ordered according to their variance, with the first PC accounting for most of the variance. The associated unit-length weight vector is the first eigen vector of the covariance matrix of the neural activity, and describes the ensemble of neurons whose superposition carries most of the variance. Each subsequent PC accounts for the maximum of the remaining variance in the neural activity. The percent variance carried by the different PCs, or principal neurons, during a typical session of a target-hitting experiment is described in Figure 4.11. The percent variance drops significantly for the first few principal components before reaching an approximately constant, nonzero, level. This structure agrees well with the assumption that the correlated signals in the spike counts are embedded in a largely uncorrelated noise. If the neural activities from different neurons are assumed to be conditionally independent, i.e., any correlation in the spike counts are attributed only to correlation in the underlying firing rate, the covariance matrix of the normalized neural activity can be decomposed as (see [A.17]): T, Diag ¤¥¥ E[ N i ] ´µµ X NT X N , ¥¦ var( N i ) µµ¶

(4.13)

where, X N [ N 1 N 2 N n ] is the matrix of the normalized spike counts (Equation is the corresponding matrix of normalized rate-parameters A.16) of all the neurons, , (Equation A.18), and Diag(·) is a diagonal matrix with Fi 1 E[ N i ] / var( N i ) i 1, z, n on the diagonal. Thus, the gradually slopping level of the percent variance carried by the PCs can be attributed to the gradually varying Fano factors Fi . In contrast, the excess variance of the initial PCs above the background level reflects correlated activity, which can be attributed to common signals that contribute to rate modulations.



73

ENSEMBLE PERMANENCE

20 0.2 40

Window

Under the above assumptions, the initial principal neurons define the neural ensembles that carry the common modulating signals. Thus, their identity and in particular their permanence with time reflects the dynamics of neural computation and impacts the ability to extract the relevant modulating signals (Zacksenhouse et al., 2005). In order to evaluate the permanence of the principal neurons with time, their identity is determined using PCA on small windows of time, and compared across time. Each principal neuron is defined by a single vector vi in the n-dimensional neural space, which describes the relative weight given to each recorded neuron. Similarly, the initial m principal neurons define an m-dimensional subspace. The permanence of the m-dimensional subspaces defined by the initial m principal neurons can be assessed using either: (1) the cosine angle between the subspaces, or (2) changes in the variance carried by these subspaces. Specifically, let v1(k) be the first principal neuron in the kth window, as depicted in Figure 4.12 (left panel) for the experiment analyzed in Figure 4.11. Visual inspection suggests that the same ensemble of neurons contribute to the first principal neuron along the experiment. This conclusion is quantified by the percent variance carried by v1(k) in the lth window (top right) and the cosine angle between the first principal neurons cos(v1(k), v1(l)) = v1(k) · v1(l) (bottom right).

0.8 40

60

0.7 20 40 Window

0.15 Neuron

0.9

20

80 100 0.1 120

0.05 160

Window

140

0.95 20

0.9 0.85

40

180 20 40 Window

0.8 20 40 Window

FIGURE 4.12 (See color insert following page 140.) Ensemble permanence—first principal neuron during 30 s windows along one experimental session (left). The variance carried by the first principal neurons computed at one window (vertical axis) during another window (x axis) (top right), and the cosine-angle between the first principal neurons from the two window (bottom right).


74


LINEAR REGRESSION The regression problem of Equations 4.5 and 4.11 can be stated in vector form as: y = Xc

(4.14)

Where, the data matrix X is a K w J matrix with K > J. For proper analysis, the data matrix includes a column of ones, which account for the bias term, whereas each other column is normalized for zero mean (and optionally for unit variance).

PRINCIPAL COMPONENT ANALYSIS (PCA) The normalized data matrix X can be decomposed using singular value decomposition (SVD), as (Hansen 1997): r

X T U 3V T

£u S v

i i i

(4.15)

i1

where,U [u1,...u K ] K s K and V [ v1,...vJ ] J s J are orthonormal matrices, 3 = diag(S1, S J ) K s J with S1 ≥ S2 ≥…SJ ≥ 0, and the rank r ≥ J is the number of strictly positive singular values S i . Note that: X T vi U 3V T ui Siui

(4.16)

Xui V 3T U T ui Si vi Because U T U I K s K the correlation matrix of the data is given by: R XX T V 32V T JrJ where 32 3T 3 diag(S12 ,

S 2J ) JrJ

(4.17)

Equation 4.17 defines the principal component analysis (PCA) of the data. The vectors vi J s1 are the eigenvectors of the covariance matrix of the data, with a corresponding eigenvalue Si. The eigenvectors are referred to as principal velocity patterns when considering high velocity profile or principal neurons when considering neural data. The principal components (PCs) of the data, are the projection of the data on the eigenvector vi i = 1, …, J using Equation 4.15, i.e., pi X T vi U 3V T ui S iui . The PCs are referred to as the velocity PCs, in the case of velocity data, or the activity of the principal neurons in the case of neural data. The variance of each PC is determined by the corresponding eigenvalue as var( pi ) S i2 , and the percent variance carried by each PC is given by: %var( pi )

Si2

£S j1


(4.18)

N

2 j


75

LEAST-SQUARE SOLUTION The least-square solution can be expressed in terms of the PCA (or SVD) as (Hansen 1997; Fierro et al., 1997): r

cLS

T i

£ uS y v

(4.19)

i

i

i1

The resulting mean square error of the regression (MSE) is given by: K

2

MSE LS MSE (cLS ) y XcLS

£ (u y) T i

2

(4.20)

ir1

Where the vector y was expanded in the orthonormal basis defined by the vectors ui as: r

y

£ (u y)u . T i

i

i1

The LS optimal solution is highly sensitive to measurement errors or uncertainties because, from Equation 4.19, it depends on the inverse of the singular values. Small singular values can dominate the solution and magnify errors in the measurement vector y. Thus, it is common to use regularization to stabilize the solution. Two common regularization methods are (Hansen 1997; Fierro et al., 1997) (1) Truncated SVD (tSVD), and (2) Tikhonov regularization.

REGULARIZATION METHODS The truncated-LS regression is obtained by truncating the LS regression of Equation 4.19 at k ≤ r ≤ J: k

ck

T i

£ uS y v

(4.21)

i

i

i1

The resulting MSE is: r

K

2

MSEk MSE (ck ) y Xck

£ (u y) MSE £ (u y) T i

i k 1

2

T i

LS

2

(4.22)

ik 1

which increases as more terms in the LS regression are truncated. Tikhonov regularization stabilizes the optimal LS solution by minimizing the combination of the MSE and the size of the regression vector:


76

Methods for Neural Ensemble Recordings, Second Edition 2 § · min ¨ y X T c L 2 Lc ¸ . © ¹

When L = I, the regularized regression vector with a given L2 is (Elden 1982): r

cL

2 i

£ S S L i1

2 i

2

uiT y vi Si

(4.23)

The resulting MSE is: 2

r

MSEL MSE (cL ) y X T cL

L4

£ S L i1

2 i

2

2 T 2 (ui y )

K

£ (u y) T i

2

(4.24)

ir1

Because tSL 2L 2 tL 2

r

£ i1

´µ ¤ ¥¥ Si2 T 2µ µµ 0, ¥¥ u y ( ) i µµ ¥¥ Si2 L 2 3 ¶µ ¦

the MSE increases with L2 and the minimum is achieved for the optimal LS solution cM= 0 = cLS. By varying the Tikhonov parameter L, it is possible to trade-off between robustness to uncertainties (large L) and performance on training data (small L).

CONCLUSIONS The neural activity during general, free reaching movements represents the spatiotemporal profile of the movement. This representation can be captured by a multi-lag regression of the neural activity on the velocity profile. However, given the significant inter-lag correlations, this representation cannot be captured well by combining single-lag analysis. Interestingly, the neural activity is represented better in terms of the multi-lag tuning curves, which reveal how the preferred direction and depth of modulation changes with the lag, rather than in terms of tuning to the Principal components of the velocity profile. Likewise, velocity prediction involves the spatiotemporal activity across an ensemble of neurons at multiple lags. The resulting high-dimensional regression problem requires regularization techniques for stabilizing the solution in the presence of neural noise and uncertainties due to the potential contributions of other modulating signals. A critical issue for future BMI improvements is whether the neural activity encodes other relevant signals, aside from the spatiotemporal profile of the movement velocity. To assess this issue we developed a measure, termed the percent overall modulations, which quantifies the percent variance that can be attributed to



77

neural modulations under the Poisson assumption. Comparing the percent overall modulations with the percent variance that can be attributed to the velocity profile suggests that the neural activity is modulated by additional signals, whose exact nature is still under investigation.

APPENDIX—DERIVATION OF THE VARIANCE RELATIONSHIP The relationship between the statistics of the spike counts and those of the underlying stochastic information process are stated in (Snyder, 1975), with a proof based on the moment generating function. Here we provide a direct proof for the variance relationship, and for the cross-covariance between the spikes counts elicited between two doubly stochastic Poisson processes.

VARIANCE RELATIONSHIP FOR DOUBLY STOCHASTIC POISSON PROCESSES Let {Nt} be a doubly stochastic Poisson process with intensity {Lt (xt )}, where xt is the underlying information process. The mean and variance of the number of spikes in bins of width b, {Nb}, are related to the mean and variance of the underlying stochastic Poisson parameter: ,b

°

t 0 b

L S ( xS )dS t0

according to: E[ N b ] E[ , b ] (A.1) Var[ N b ] Var[,b ] E[,b ] Var[,b ] E[ N b ] Proof: By definition: c

E[ N b ]

£

c

n Pr( N b n ) and E[ N b2 ]

n0

£ n Pr(N 2

b

n)

(A.2)

n0

Using the method of conditioning (Snyder 1975, Equation 6.1 there) Pr( N b n ) E[Pr( N b n | , b ] E[(n!) 1 ,bn exp( ,b )] where the expectation is with respect to the stochastic parameter ,b. Substituting Equation A.3 in Equation A.2: c

E[ N b ]

£ nE[(n!)

1

n0


,bn exp( ,b )]

(A.3)

78


and c

E[ N b2 ]

£ n E[(n!)

1

2

,bn exp( ,b )]

(A.4)

n0

Because the expectation is with respect to ,b it can be moved outside the summation, so the first part of Equation A.4 implies that: c

E[ N b ] E[

£ n(n!)

1

,bn exp( ,b )]

n0

(A.5) c

E[exp( ,b ),b

£ (m!)

1

,bm ] E[,b ]

m 0

where m = n – 1, and the last step is based on the equality c

£ m 0

,b2 exp(,b ). m!

(A.6)

This proves the relationship between the mean of the spike counts and the underlying rate parameter. The second part of Equation A.4 implies that: c

E[ N b2 ] E[

£ n (n!) 2

1

,bn exp( ,b )]

n0

c

E[exp( ,b ),b2

£

c

(l !) 1 ,lb ] E[exp( ,b ),b

l 0

£ (m!)

1

,bm ] (A.7)

m 0

E[,b2 ] E[,b ] where l = n – 1. Finally, the last two equations can be manipulated to derive the following relationship between the respective variances: Var[ N b ] E[{( N b E[ N b ]}2 ] E[ N b2 ] E[ N b ]2 E[,b2 ] E[,b ] E[,b ]2 E[{,b E[,b ]}2 ] E[,b ] Var[,b ] E[,b ]

(A.8)

This completes the proof of the variance relationship stated in (A.1). The cross-correlation and co-variance relationship between two spike trains that are generated by two doubly stochastic Poisson processes are derived next.



79

CROSS-VARIANCE RELATIONSHIP FOR DOUBLY STOCHASTIC POISSON PROCESSES Let {N1} and {N2} be the number of spikes in bins of width b that were elicited by two doubly stochastic Poisson processes with underlying stochastic Poisson parameters ,1 and ,2 (the index b, indicating the bin-width is omitted for simplicity, but is the same for both processes). If the two processes are conditionally independent the cross-covariance between the spike counts is E[ N1N 2 ] E[,1, 2 ]

(A.9)

cov[ N1, N 2 ] cov[,1, , 2 ] Proof: By definition: c

E [ N1 N 2 ]

c

£ £ n n Pr(N n & N 1 2

1

1

2

n2 )

(A.10)

n10 n2 0

Using the method of conditioning Pr( N1 n1 & N 2 n2 ) E[Pr( N1 n1 & N 2 n2 | , 1 , , 2 ]

(A.11)

Where the expectation is with respect to stochastic parameters ,1 and ,2. Conditional independence implies that given the two rate parameter, the two spike counts are independent, i.e., any correlation between the two spike counts is generated only by correlation between the underlying rate parameters. Hence: Pr( N1 n1 & N1 n1 ) E[Pr( N1 n1 | , 1 ) Pr( N 2 n2 | , 2 )] (A.12)

E[(n1 !) 1 ,1n1 exp( ,1 )(n2 !) 1 , n22 exp( , 2 )]

Substituting Equation A.12 in Equation A.10, and making the change of indices m1 = n1 – 1 and m1 = n1 – 1, result in: c

E [ N1 N 2 ]

c

£ £ n n E[(n !) 1 2

1

1

,1n1 exp( ,1 )(n2 !) 1 , n22 exp( , 2 ]

n10 n2 0

c

£

E[,1, 2 exp( ,1 )exp( , 2 )

c

( m1 !) 1 ,1m1

m10

£ (m !) 2

1

, 2m2 ]

(A.13)

m2 0

E[,1, 2 ]. Hence (using also Equation A.1), cov[ N1, N 2 ] cov[,1, , 2 ]


(A.14)

80


In summary, Equations A.14 and A.8 imply that: cov[ N i , N j ] cov[,i , , j ] Dij E[,i ]

(A.15)

The spike counts can be normalized to zero-mean and unit variance according to: N E(Ni ) N i i var( N i )

(A.16)

Thus, the covariance of the normalized spike counts is: cov[ N i , N j ]

cov[,i , , j ] Dij E[,i ] var(,i ) E (,i ) var(, j ) E (, j )

(A.17)

i,, j ] Dij E[ N i ] cov[, var( N i ) Where the normalized rate-parameter is: i ,

,i E (,i ) var(,i ) E (,i )

(A.18)

REFERENCES Alexander, G.E. and Crutcher, M.D. (1990) Neural representations of the target (goal) of visually guided arm movements in three motor areas of the monkey. J. Neurophysiol. 64: 164–178. Ashe, J. and Georgopoulos, A.P. (1994) Movement parameters and neural activity in motor cortex and area 5. Cereb. Cortex 4: 590–600. Brown, E.N., Kass, R.E., and Mitra, P.P. (2004) Multiple neural spike train data analysis: state-of-the-art and future challenges. Nat. Neurosci. 7(5): 456–461. Brown, R.G. and Hwang, P.Y.C. (1997) Introduction to random signals and applied kalman filtering, 3rd ed. Wiley: New York. Carmena, J.M., Lebedev, M.A., Crist, R.E., O’Doherty, J.E., Santucci, D.M., Dimitrov, D., Patil, P.J., Henriquez, C.S., and Nicolelis, M.A.L. (2003) Learning to control a brain– machine interface for reaching and grasping by primates. PLoS Biol. 1: 193–208. Dayan, P. and Abbott, L.P. (2001) Theoretical neuroscience. MIT Press: Cambridge, MA. Elden, L. (1982) A weighted pseudoinverse, generalized singular values, and constrained least squares problems. BIT, 22: 487–502. Fierro, R.D., Golub, G.H., Hansen, P.C., and O’Leary, D. P. (1997) Regularization by truncated total least square, SIAM J. Sci. Comput. 18: 1223–1241. Georgopoulos, A.P. (2000) Neural aspects of cognitive motor control. Curr. Opin. Neurobiol. 10: 238–241. Georgopoulos, A.P., Ashe, J., Smyrnis, N., and Taira, M. (1992) The motor cortex and the coding of force. Science 256: 1692–1695. Georgopoulos, A.P., Lurito, J.T., Petrides, M., Schwartz, A.B., and Massey, J.T. (1989) Mental rotation of the neuronal population vector. Science 243: 234–236. Georgopoulos, A.P., Schwartz, A.B., and Kettner, R.E. (1986) Neuronal population coding of movement direction. Science 26: 1416–1419.



81

Hansen, P.C. (1997) Rank-deficient and discrete Ill-posed problems, SIAM: Philadelphia. Johnson, D.H. (1996) Point Process models of single neuron discharge. J. Comp. Neurosci. 3: 275–299. Johnson, M.T.V. Mason, C.R., and Ebner, T.J. (2001) Central processes for the multiparametric control of arm movements in primates. Curr. Opin. Neurobiol. 11: 684–688. Lebedev, M.A., Carmena, J.M., O’Doherty, J.E., Zacksenhouse, M., Henriquez, C.S., Principe, J., and Nicolelis, M.A.L. (2005) Cortical ensemble adaptation to represent velocity of an artificial actuator controlled by a brain machine interface. J. Neurosci. 25(19): 4681–4693. Moran, D.W. and Schwartz, A. (1999) Motor cortical representation of speed and direction during reaching. J. Neurophysiol. 82: 2676–2692. Nicolelis, M.A.L. (2001) Actions from thoughts. Nature 409(18): 403–407. Nicolelis, M.A.L. (2003) Brain-machine interfaces to restore motor functions and probe neural circuits. Nat. Rev. 4: 417–422. Paz, R., Wise, S.P., and Vaadia, E. (2004) Viewing and doing: similar cortical mechanisms for perceptual and motor learning. Trends Neurosci. 27, 496–503. Perkel, D.H. and Bullock, T.H. (1968) Neural coding. Neurosc. Res. Progr. Bull. 6, 221–248. Pouget, A., Dayan, P., and Zemel, R.S. (2003) Inference and computation with population codes. Annu. Rev. Neurosci. 26, 381–410. Scott, S.H. (2003) The role of primary motor cortex in goal directed movements: insights from neurphysiological studies on non-human primates. Curr. Opin. Neurobiol. 13: 671–677. Snyder, D.L. (1975) Point Processes. Wiley: New York. Wessberg, J., Stambaugh, C.R., Kralik, J.D., Beck, P.D., Laubach, M., Chapin, J.K., Kim, J., Biggs, S.J., Srinivasan, M.A., and Nicolelis, M.A.L. (2000) Real-time prediction of hand trajectory by ensembles of cortical neurons in primates. Nature 408: 361–365. Wu, W., Shaikhouniy, A., Donoghue, J.P., and Blackz, M.J. (2004) Closed-loop neural control of cursor motion using a Kalman filter, Proc. 26th Int. Conf. IEEE EMBS, San Francisco, CA, September. Zacksenhouse, M., Lebedev, M.A., Camena, J.M., O’Dohery, J.E., Henriquez, C.S., and Nicolelis, M.A.L. (2005) Correlated ensemble activity increased when operating a brain machine interface, extended abstract, Comp. Neuro-Science CNS–05. Zacksenhouse, M., Lebedev, M.A., Camena, J.M., O’Dohery, J.E., Henriquez, C.S., and Nicolelis, M.A.L. (2007a) Cortical modulations increase in early sessions brain machine interfaces. PLoS-ONE, 2(7): e619.


5

Chronic Recordings in Transgenic Mice Kafui Dzirasa

CONTENTS Introduction.............................................................................................................. 83 Construction of Implant ...........................................................................................84 Construction of Electrodes ...........................................................................84 Electromyographic Electrodes........................................................... 86 Implantation Surgery ............................................................................................... 86 Anesthesia..................................................................................................... 86 Microarray Implantation............................................................................... 87 EMG Wires................................................................................................... 87 Recording Protocol .................................................................................................. 87 Recording Paradigms ................................................................................... 87 Gross Locomotor Behavior................................................................ 87 Motor Coordination ........................................................................... 88 Learning and Memory ....................................................................... 88 Sleep and Circadian Rhythms ........................................................... 88 Pharmacological Manipulations ........................................................ 89 Data Collection ............................................................................................. 89 Data Analysis ...........................................................................................................90 Spike Sorting/Waveforms.............................................................................90 Behavioral State Analysis.............................................................................90 Results and Conclusions ............................................................................... 91 References................................................................................................................ 95

INTRODUCTION Previous chapters of this book have highlighted technical advances made in recording ensembles of neurons in primates and rats. These advances have created a novel window through which to study brain function, allowing us to interpret an ever-growing body of data describing how the normal brain operates. On the other hand, these techniques have only provided a limited picture of how alterations in brain function generate behavioral pathology, requiring new recording techniques to be manufactured. In this chapter, we describe a novel method for extending previously described electrophysiological recording techniques for use in transgenic mice. This fusion of genetic and neuroscience technologies provides a promising tool for elucidating 83 © 2008 by Taylor & Francis Group, LLC

84


how changes in gene expression result in the behavioral alterations characteristic of central nervous system (CNS) disease. During the course of the last century, great strides were made in treating diseases involving the CNS. Much of this progress resulted from an explosion of neuroscience research, which led to the discovery of action potentials, neuromodulators, and the principles of organization of neural circuits. Although these findings facilitated the development of drugs to treat neurological and psychiatric illnesses such as schizophrenia, Alzheimer’s disease, Parkinson’s disease, and depression, little headway was made in elucidating the pathophysiological mechanisms associated with many of these diseases. Thus, treatment options typically focused on managing symptoms, and not on curing the underlying disease. Currently, one of the primary challenges of behavioral brain research lies in the complexity of understanding how changes in patterns of gene expression alter the spatiotemporal firing of widely distributed populations of single neurons that define the large-scale neural interactions underlying the generation of behavior. This issue is further complicated by ethical concerns associated with genetic manipulation in humans and limitations inherent to the techniques currently used to study electrophysiological changes in the brain. The primary techniques used to study human brain activity are functional magnetic resonance imaging (fMRI) and electroencephalography (EEG). Unfortunately, fMRI, which measures changes in blood flow resulting from activation of neural networks, has significant limitations in temporal resolution, and EEG, which utilizes scalp electrodes to measure brain activity, has significant limitations in spatial resolution. Although intraparenchymal recording protocols can be used to enhance spatial and temporal resolution of brain activity, these techniques are invasive and are thus not typically suitable for studying neurological and psychiatric illness in human populations. To overcome these issues, we have created a protocol for chronically recording brain activity in animal models of CNS disease. Recently, transgenic mice have gained greater acceptance as models of CNS disease (Dennis 2005). These genetically modified mice display behavioral alterations similar to those observed in humans with neurological or psychiatric illness, and recapitulate several endophenotypes characteristic of these diseases. By chronically implanting these mice with intraparenchymal microelectrode arrays, one is able to assess how changes in gene expression result in changes in neural ensemble firing patterns. Moreover, one can conduct longitudinal observations in animals given drugs used to treat the CNS diseases they model, providing a detailed picture of how changes in neuronal firing patterns result in neurological and psychiatric pathology.

CONSTRUCTION OF IMPLANT The transgenic mouse recording protocol is very similar to that utilized in rats, with one major difference: mice are typically one-tenth the size of rats. This, of course, necessitates several modifications to the previously described implant.

CONSTRUCTION OF ELECTRODES The main design used to conduct electrophysiological recordings in mice consists of an array of 16 S-isonel-coated tungsten microwire electrodes, a printed circuit board


Chronic Recordings in Transgenic Mice

85

(PCB) connected to the microwire electrodes, and a high-density, miniature connector attached to the opposite side of the PCB (Nicolelis et al. 2003). We investigated the advantages and disadvantages of several array designs: (1) 16 single-ended wires spaced 250 µm apart (Figure 5.1A); (2) Four groups of four microwires separated by 100 µm in a square arrangement (each group separated by 250 µm; Figure 5.1B); (3) Eight pairs of two adjacent electrodes (maximum 50 µm between wire centers in each pair, pairs separated by 250 µm; Figure 5.1C). We also tested two wire diameters: 50 µm, with an impedance of about 1.2 MΩ; and 35 µm, with an impedance of about 1.5 MΩ. The arrays and headstage were miniaturized (weight ~0.26 g per array), allowing mice to move freely. In our experiments, 32 microwires were implanted per mouse (an array of 16 microwires in each hemisphere) in the dorsal striatum and the motor cortex. To test the different designs, we recorded in single-channel mode (for all designs), tetrode mode (for design 2) or stereotrode mode (for design 3), (using MAP, Plexon Inc.). We used differential recording in which one of the channels with no neural activity was used as a reference electrode. Using the single-channel recording mode, we were able to record well-isolated single units from single channels in all three designs (Figure 5.1D). With both designs 2 (tetrode mode) and 3 (stereotrode mode), we were unable to record simultaneous activity from the same unit in adjacent electrodes, even when very well-isolated single neurons could be recorded from each A

B

Stereotrode

C

Stereotrode Mode

Single Channel Mode

D

Channel 1 Channel 2

0

20 ms

FIGURE 5.1 Comparison of different arrays and recording modes to record multiple single-neuron activity in awake behaving mice. A, B, and C are different microarray designs. (A) 16 single-ended wires spaced 250 µm apart. (B) Four groups of four microwires separated by 100 µm in a square arrangement—each group separated by 250 µm. (C) Eight pairs of two adjacent electrodes. (D) Example of a striatal single unit recorded in stereotrode mode. Although the signal recorded in channel 2 of the stereotrode is very prominent, no signal from this unit was captured in channel 1 of the stereotrode (note that the threshold is very close to zero). Also, the isolation quality of the cluster of this single unit is similar when the unit is recorded in stereotrode or single-channel mode.


86


of the adjacent electrodes in single-channel mode. In this context, our results differ from previous studies using stereotrode and tetrode recording methods (Harris et al. 2000). Likely, this difference results because we used electrodes with an impedance 4 to 5 times higher than in other studies, larger wire diameter, different array design geometry, and we also recorded from sparsely distributed (in striatum 2–4 cells 35 µm3 or 6–12 cells 50 µm3 [Rosen and Williams 2001]) and very small (~12 µm cell body diameter for striatal medium spiny neurons [Rafols et al. 1989]; and ~15 µm for pyramidal cortical cells [Faherty et al. 2003]) neurons. Using the specific parameters described here, we observed no advantage in using triangulation to isolate single units. This observation could be different when recording in other brain areas or when using other parameters. We therefore opted to employ single-channel recording mode in all the experiments. In the striatum, we obtained better results with the 50 µm wires (both in terms of number of units isolated and longevity of the recordings), whereas in motor cortex both wire diameters produced similar outcomes. When constructing implants, individual electrodes should be geometrically spaced and cut according to mouse stereotaxic coordinates to ensure appropriate spacing for target brain structures. Electromyographic Electrodes Electromyographic (EMG) electrodes provide an external measure of behavior. These electrodes are particularly useful when studying locomotor behavior, or classifying electrophysiological changes across sleep–wake states. To avoid damage to muscle tissue, smaller single-wire EMG electrodes are utilized in transgenic mice implants. EMG electrodes are created using 50 μm tungsten wire, and are connected to the microwire arrays as described in chapter 1. They should be cut to a length of no less than 7 cm to facilitate muscle implantation. Importantly, multiple EMG wires can be attached to each microwire array, allowing for muscle activity to be simultaneously recorded from the eyes, whiskers, neck, and other peripheral muscles.

IMPLANTATION SURGERY Transgenic mice necessitate several adjustments to the rodent implantation surgery described in chapter 3. These modifications result from the decreased body mass of the mice in comparison to rats and primates, and the smaller surgical implantation field.

ANESTHESIA After selecting a mouse for implantation, rapidly induce the animal using halothane/ isoflourane (as described in chapter 3). Once the animal is induced, the mouse should be injected with a single dose of ketamine 100 mg/kg IM, and xylazine 10 mg/kg IM. This combination of anesthetic agents should be sufficient to provide 2–3 h of controlled anesthesia. If the mouse should require additional anesthesia during the course of the surgery, prepare a 20 mg/ml ketamine solution, and inject 0.1 mL IM as needed. At least 5 min should elapse between subsequent maintenance doses of ketamine to avoid causing unanticipated respiratory depression.



87

MICROARRAY IMPLANTATION The mouse cranium has a very thin skull, making the implantation surgery a technically challenging procedure. Once the cranium is exposed as described in the rat implantation surgery (chapter 3), target locations are marked off using a felt tip pen. First, small craniotomies for placement of anchoring screws are drilled using a ¼ mm drill bit. Screws (1 mm diameter) are used to anchor the implant. The size of such screw craniotomies should be approximated to tightly secure each screw. Two screw craniotomies are usually drilled for each surgery, but only one is necessary if numerous multiarrays are being implanted or cranial space is limited. After the screw craniotomies are opened, the anchor screws should be inserted using a screwdriver. Next, the electrode ground wire should be wrapped around the secured ground screw. The process should be repeated for each electrode. Two and a half revolutions should be sufficient. Finally, after lowering the electrodes to the desired depth, the ground wire should be covered and screwed with dental acrylic (as described in chapter 4). Care should be taken to avoid covering any of the other implant target sites with dental acrylic.

EMG WIRES EMG wires should be placed after microwire arrays are implanted and covered with dental acrylic. To implant EMG wires, pass a 27-gauge needle through the target muscle. Next, insert the EMG wire into the opening of the needle, and retract the needle through the target muscle. Repeat this procedure 2–3 times to ensure that the EMG wire is properly secured. Only one pass is necessary to secure eye and whisker EMG wires. Finally, remove the insulation coating from the EMG wire using a scalpel, and cut the EMG wire at the muscle.

RECORDING PROTOCOL The transgenic mouse recording protocol is similar to that utilized for conducting electrophysiological recordings in rats. After mice are implanted, connect them to the recording setup using a 16-channel headstage and connector. The neural ensemble recording setup is identical to that utilized to record rats.

RECORDING PARADIGMS Various recording paradigms can be used to uncover behavioral phenotypes in transgenic mice (Crawley 1999). By conducting the transgenic mouse recording protocol using these alternative recording paradigms, one can investigate the electrophysiological alterations underlying specific behavioral changes. Thus, this section will focus on describing the specific behavioral phenotypes isolated by each recording paradigm. Gross Locomotor Behavior Neural activity associated with changes in gross locomotor activity can be assessed by conducting recordings in an open field. Open-field recordings are conducted by


88


placing mice in an open environment in which they can move freely. This environment can be stimulated using a specialized setup for tracking movement, or even an empty mouse cage and a video recorder. The open-field recording paradigm is particularly useful when studying genetic alterations that are known to generate changes in gross locomotor activity, such as hyper- or hypolocomotion, or stereotypic behavior such as grooming and licking. These phenotypes are displayed in human diseases such as Parkinson’s disease, obsessive-compulsive disorder, schizophrenia, and attention-deficit hyperactivity disorder. Motor Coordination The accelerating rotarod (Crawley 2000) is particularly useful for studying changes in neural activity underlying deficits in motor coordination. During the rotarod recording setup, mice are placed on a rotating cylinder that is accelerated over time. Motor coordination is then assessed as the latency for the animal to fall off the rotating cylinder. To avoid injury to the animal, the rotarod cylinder is suspended only 6 in. from the bedding-covered floor. By conducting neural ensemble recordings in transgenic mice using the rotarod, one can study electrophysiological alterations associated with disorders of the motor system. This recording paradigm would be particularly useful for studying pathophysiological mechanisms underlying human neurological disorders such as amyotrophic lateral sclerosis and Huntington’s disease. Learning and Memory Many behavioral paradigms have been developed to assess cognitive deficits in genetically altered mice. These paradigms include the light/dark and foot shock avoidance task, water maze, radial arm maze, and T/Y mazes (Crawley 2000). By conducting electrophysiological recordings while the transgenic mice are performing each of these tasks (with the exception of the water maze, which may present a problem with the recording equipment), one can probe alterations in neural circuits underling deficits in learning and memory that result from genetic manipulations. This recording protocol may be particularly useful in studying mouse models of human diseases such as autism, schizophrenia, and Alzheimer’s disease. Sleep and Circadian Rhythms A long-term recording setup can be used when experiments necessitate that recording periods last several days. The long-term recording setup is particularly useful when one is studying how genetic alterations generate changes in neural activity associated with the circadian and sleep–wake cycle. Long-term recordings are typically carried out in a modified mouse cage, which is placed in a behavior box that allows the temperature and light dark cycle to be closely controlled. Importantly, the modified mouse cage is filled with bedding and equipped with a modified feeding grate that allows for food and water to be accessed by the mouse ad libitum. Additionally, care must be taken ensure that the modified feeding grate does not cause the headstage connector to become tangled.



89

Pharmacological Manipulations One can also conduct electrophysiological recordings in transgenic mice treated with pharmacological agents known to treat the human neurological diseases they model. The aforementioned recording paradigms can than be utilized to assess how these agents affect alterations in neural circuitry observed in untreated mice.

DATA COLLECTION Once a mouse is implanted, connected to the recording setup, and the appropriate recording paradigm is selected, single-cell, multiunit, LFP, and EMG activity is recorded using the Multi Acquisition Processor (MAP) system (Plexon Inc., Texas) (Figure 5.2). The spike activity should be initially sorted using an online sorting algorithm. Only units that have a clearly identified waveform with a signal-to-noise ratio of at least 3:1, where no waveforms are dropped by the acquisition system (sampling rate of 40 KHz), should be used. At the end of the recording, units should be resorted offline based on waveform, amplitude, and interspike interval histogram using an offline sorting algorithm (Nicolelis et al. 2003). Concomitantly, we have also measured local field potentials (LFPs) in the same areas by low-pass-filtering the data (0.1–400 Hz range). Histological analysis demonstrated that chronic implants did not produce significant tissue damage besides the original implantation track.

A

B

FIGURE 5.2 (See color insert following page 140.) (A) Upper-left panel. Craniotomy in the area above the dorsolateral striatum. An array of 16 microwires is ready for implantation. Lower-left panel. Awake mouse with 32 microwires inserted bilaterally in the dorsolateral striatum during a recording session. (B) Screen of our acquisition system during a recording session of the mouse presented in the left, one and a half months after surgery. The selected channel shows three clearly isolated cells recorded simultaneously from one electrode. Twenty-one other channels are also recording single-neuron activity. (Costa, R.M., Cohen, D., and Nicolelis, M.A. (2004). Differential corticostriatal plasticity during fast and slow motor skill learning in mice. Curr. Biol. 14(13): 1124–1134.)


90


DATA ANALYSIS SPIKE SORTING/WAVEFORMS In our studies, single-cell and multiunit activity were recorded using the MAP system (Plexon Inc., Texas). In our protocol, the recorded neural activity is initially sorted using an online sorting algorithm (Plexon Inc, Texas). At the end of the recording, potential single units should be resorted offline based on waveform, amplitude, and interspike interval histogram using an offline sorting algorithm (Nicolelis et al. 2003). Only units in which all the spikes have amplitudes well above the voltage threshold should be considered (to ensure that no spikes are lost). When the waveform, amplitude, and inter-spike-interval (ISI) do not allow unambiguous isolation (even with signal-to-noise ratios larger than 3:1), the signals should either be discarded or labeled multiunit activity. The criteria should be as follows: When the waveforms from a unit clearly result from two or more neurons that cannot be discriminated, the unit should be discarded. In the occasional cases in which not all the spikes of a single unit can be unequivocally isolated during offline sorting (owing to the proximity of two clusters or proximity to the noise cluster), the unit should be labeled multiunit. During three sessions, we found that the average number of multiunits per session was 11, 13, and 11%. Analyses of these units should be conducted separately. We observed that the waveform recordings across a session were generally very stable (Figure 5.3), allowing us to follow single units through time. The implants seemed to be well tolerated by the mouse brain, and single units could be recorded several months after surgery (Figure 5.4). These experiments revealed that several well-isolated single units can be obtained from a very high number of implanted electrodes, indicating that our methodology is appropriate for conducting long-term longitudinal studies in behaving mice. Additionally, we observed that changes in neural activity could be isolated from single units recorded across a behavioral learning task (Figure 5.5) (Costa 2004).

BEHAVIORAL STATE ANALYSIS Behavioral state scoring algorithms typically combine analysis of brain and EMG activity, with direct behavioral observations. We recently developed a novel technique in rats that can be utilized to predict behavioral states and their transitions based on LFP spectral ratios (Gervasoni et al. 2004). The first ratio (Ratio 1) produces cluster separation based on high-frequency gamma (33–55 Hz) spectral oscillations (Steriade et al. 1993), whereas the second (Ratio 2) produces cluster separation based primarily on theta (4–9 Hz) spectral oscillations (Vanderwolf 1969; Timo-Iaria et al. 1970; Cantero et al. 2003). We found that this method could also be used to predict behavioral states in mice using LFP oscillations recorded from hippocampus (Figure 5.6a and Figure 5.6b). Additionally, EMG data can be used to further discriminate points of behavioral state cluster overlap (Figure 5.6c). To determine EMG activity, the Fast Fourier transform was applied to the LFP signal using a 2 s window with a 1 s step. The Fourier transform parameters were chosen to allow for a frequency resolution of 0.5 Hz. EMG activity was then calculated by taking the root mean square of the spectral amplitude over selected frequency bands: 30–56 and 64–250 Hz. All EMG traces



0

91

20 ms

FIGURE 5.3 Stability of the recordings during a session. Example of a striatal unit recorded 4 months after surgery. Left Panel. Waveforms (top) and inter-spike-histogram (bottom) of a selected unit. Center Panel. Top—Projection of the three first principal components (PCs) of the waveform shows that the waveforms of the unit cluster distinctively. Bottom—Projection of the two PCs of the waveform across time shows that they remain relatively constant across time. Right Panel. Stability of the PC cluster based on the first three PCs (top) and waveform (bottom) across time. (Costa, R.M., Cohen, D., and Nicolelis, M.A. (2004). Differential corticostriatal plasticity during fast and slow motor skill learning in mice. Curr. Biol. 14(13): 1124–1134.)

were high-pass-filtered at 30 Hz. Importantly, behavioral states predicted by the twodimensional state space method correspond with those determined using standard sleep scoring methods (Figure 5.7). Thus, this powerful technique facilitates highthroughput behavioral state analysis in mice, allowing for the observation of neuronal subpopulations across the sleep–wake cycle in animal models of CNS disease.

RESULTS AND CONCLUSIONS We recently utilized the method outlined in this chapter to examine the influence of dopamine on sleep–wake states. Dopamine transporter knockout (DAT-KO) mice were implanted with multiarray electrodes in the hippocampus, and electrophysiological recordings were conducted across the sleep–wake cycle (Dzirasa et al. 2006). DAT-KO mice lack the gene encoding the dopamine transporter, a transmembrane protein that is responsible for regulating the reuptake of synaptic dopamine and replenishing dopamine stores in the presynaptic terminal (Gainetdinov and Caron 2003). As such, the dopamine transporter plays the key role in controlling


92


FIGURE 5.4 Electrophysiological recording protocol in mice. Screen of our acquisition system during a recording session 10 weeks after surgery. Along with five isolated cells (top), hippocampal LFP oscillations are also being recorded from four channels, and electromyographic activity is being recorded from the trapezius muscle. (Costa, R.M., Cohen, D., and Nicolelis, M.A. (2004). Differential corticostriatal plasticity during fast and slow motor skill learning in mice. Curr. Biol. 14(13): 1124–1134.)

dopamine homeostasis. Because of the loss of the dopamine transporter, DAT-KO mice exhibit profound depletion of intraneural dopamine stores as well as a five-fold increase in extracellular dopamine levels (Gainetdinov and Caron 2003). Our results demonstrated that novelty-exposed DAT-KO mice display a new awake state characterized by electrophysiological signatures characteristic of REM sleep (Dzirasa et al. 2006). Moreover, when these mice were dopamine-depleted via administration of a tyrosine hydroxylase inhibitor (Sotnikova et al. 2005), they displayed another new sleep–wake state characterized by the complete suppression of REM sleep. These results indicated that dopamine directly regulates the sleep–wake states. Our findings elucidate mechanisms central to the role that dopamine plays in mediating symptoms classically associated with neurological and psychiatric diseases. This work demonstrates that by combining in vivo electrophysiological recording techniques with techniques fundamental to genetics and pharmacology, one can gain novel insights into pathophysiological mechanisms underlying neurological and psychiatric disease. Furthermore, this approach is particularly promising given the ever-growing library of genetic alterations underlying brain diseases, and subsequent development of multiple novel transgenic mouse lines. By expanding our electrophysiological recording technique to collect and analyze ensembles of single-unit neuronal data recorded simultaneously from multiple brain sites in various transgenic mouse lines, we believe that our method will ultimately demonstrate how genes interact to generate normal and altered behavior.



93 Motor Cortex Frequency (imp/sec)

Frequency (imp/sec)

Dorsolateral Striatum

3

0 –200

–100

0

100

200

300

400

6 3 0 –200

–100

0

–100

0

100

200

300

400

Frequency (imp/sec)

Frequency (imp/sec)

10

0 –200

100

200

300

400

Frequency (imp/sec)

Frequency (imp/sec)

3 0

0 –200

–100

0

100 Time (sec)

200

300

400

Frequency (imp/sec)

Frequency (imp/sec)

3 0

100

200

300

400

200

300

400

200

300

400

4

0 –200

–100

0

100 Time (sec)

6

–100

400

3

Time (sec)

0 –200

300

Time (sec)

6

–100

200

6

Time (sec)

0 –200

100 Time (sec)

Time (sec)

10 5 0 –200

–100

0

100 Time (sec)

FIGURE 5.5 Modulation of neural activity during running on the accelerating rotarod. Left column: striatal neurons, right column: motor cortex neurons. Each panel displays the raster plots and corresponding peri-event time histogram (PETHs) for a particular neuron throughout one session. The beginning of each running period is aligned at time zero. In each graph, trials are presented top to bottom as they were presented during training. Dorsal striatum neurons are presented on the left column and motor cortex neurons on the right column. First row: Example of a striatal and a cortical neuron that abruptly increase their firing rate during the running period. Second row: Example of a striatal and a cortical neuron that abruptly decrease their firing rate during the running period. Third row, left: Example of a striatal neuron with a rather high firing rate at the onset of each trial; gradually decreases firing rate toward the end of the trial. Third row, right: Example of a cortical neuron that gradually increases its firing rate during the running period. Fourth row: Example of a striatal and a cortical neuron that transiently changed their firing rate at the beginning of the running period and then gradually decreased it throughout the trial.


94

25

3

0.72

2.4

Ratio 1

SWS SWS

10 5

EMG Power (µV2)

15 0.64 REM 0.56

WK

WK 0.48

0

10 0 Ratio 2

40 20 30 Frequency (Hz) Ratio 1 (a)

50

1.8 1.2 WK 0.6 REM

60 0.40 0.15

0.25

0.35 0.45 Ratio 2 (b)

0.55

0.65

0 0.15

0.25

SWS 0.35 0.45 Ratio 2

0.55

0.65

(c)

FIGURE 5.6 State-dependent local field potential oscillations and behavioral state map of adult mice. After a 1-week recovery period, normal mice were subjected to 12-h continuous LFP (hippocampus) and EMG (trapezius) recordings in their home cage. (A) The mice displayed state-dependent power spectral patterns characteristic of REM, SWS, and WK in rodents. REM was characterized by high-amplitude theta (4–9 Hz) and gamma (33–55 Hz) oscillations, SWS was characterized by high-amplitude delta (1–4 Hz) and low-amplitude gamma oscillations, and WK was characterized by high-amplitude gamma oscillations. (B) A two-dimensional behavioral state map was generated by plotting the following spectral ratios: x-axis, 0.5–4.5 Hz/0.5–9 Hz and y-axis, 0.5–20 Hz/0.5–55 Hz; (C) EMG data were used to disambiguate WK and REM clusters. All unassigned time points, typically corresponding to interstate transitions, were coded gray. (Dzirasa, K. and Ribeiro, S. et al. (2006). Dopaminergic control of sleep-wake states. J. Neurosci. 26(41): 10577–10589.)



Power (µV2)

20

0.80 REM

Chronic Recordings in Transgenic Mice LFP

95 EMG

WK

SWS

2.8 µV

0.4 µV

REM 1.0s

1.0s

FIGURE 5.7 LFP and EMG activity during state-map-predicted behavioral states. Mice were introduced into a novel cage, and subjected to 12-h continuous LFP (hippocampus) and EMG (trapezius) recordings. Real-time two-dimensional behavioral state maps were generated by plotting the following spectral ratios: x-axis, 0.5–4.5 Hz/0.5–9 Hz and y-axis, 0.5–20 Hz/0.5–55 Hz. Raw LFP and EMG activity was analyzed during periods of WK, SWS, and REM sleep predicted by the two-dimensional state map. As previously demonstrated, WK was characterized by high brain activity and high muscle activity, SWS was characterized by low brain activity and low muscle activity, and REM was characterized by high brain activity and negligible muscle activity (atonia). (Dzirasa, K. and Ribeiro, S. et al. (2006). Dopaminergic control of sleep-wake states. J. Neurosci. 26(41): 10577–10589.)

REFERENCES Cantero, J.L. and Atienza, M. et al. (2003). Sleep-dependent theta oscillations in the human hippocampus and neocortex. J. Neurosci. 23(34): 10897–10903. Costa, R.M., Cohen, D., and Nicolelis, M.A. (2004). Differential corticostriatal plasticity during fast and slow motor skill learning in mice. Curr. Biol. 14(13): 1124–1134. Crawley, J. (1999). Behavioral phenotyping of transgenic and knockout mice: experimental design and evaluation of general health, sensory functions, motor abilities, and specific behavioral tests. Brain Res. 835(1): 18–26. Crawley, J.N. (2000). What’s wrong with my mouse? New York: Wiley-Liss. Dennis, C. (2005). All in the mind of a mouse. Nature 438: 151–152. Dzirasa, K. and Ribeiro, S. et al. (2006). Dopaminergic control of sleep-wake states. J. Neurosci. 26(41): 10577–10589. Faherty, C.J. and Kerley, D. et al. (2003). A Golgi-Cox morphological analysis of neuronal changes induced by environmental enrichment. Brain Res. Dev. Brain Res. 141(1–2): 55–61. Gainetdinov, R.R. and Caron, M.G. (2003). Monoamine transporters: from genes to behavior. Annu. Rev. Pharmacol. Toxicol. 43: 261–284. Gervasoni, D. and Lin, S.C. et al. (2004). Global forebrain dynamics predict rat behavioral states and their transitions. J. Neurosci. 24(49): 11137–11147. Harris, K.D. and Henze, D.A. et al. (2000). Accuracy of tetrode spike separation as determined by simultaneous intracellular and extracellular measurements. J. Neurophysiol. 84(1): 401–414. Nicolelis, M.A. and Dimitrov, D. et al. (2003). Chronic, multisite, multielectrode recordings in macaque monkeys. Proc. Natl. Acad. Sci. US A 100(19): 11041–11046.


96


Rafols, J.A. and Cheng, H.W. et al. (1989). Golgi study of the mouse striatum: age-related dendritic changes in different neuronal populations. J. Comp. Neurol. 279(2): 212–227. Rosen, G.D. and Williams, R.W. (2001). Complex trait analysis of the mouse striatum: independent QTLs modulate volume and neuron number. BMC Neurosci. 2(5): Epub. Sotnikova, T.D. and Beaulieu, J.M. et al. (2005). Dopamine-independent locomotor actions of amphetamines in a novel acute mouse model of Parkinson disease. PLoS Biol. 3(8): e271. Steriade, M. and McCormick, D.A. et al. (1993). Thalamocortical oscillations in the sleeping and aroused brain. Science 262(5134): 679–85. Timo-Iaria, C. and Negrao, N. et al. (1970). Phases and states of sleep in the rat. Physiol. Behav. 5(9): 1057–62. Vanderwolf, C.H. (1969). Hippocampal electrical activity and voluntary movement in the rat. Electroenceph. Clin. Neurophysiol. 26: 407–418.


6

Multielectrode Recordings in the Somatosensory System Michael Wiest, Eric Thomson, and Jim Meloy

CONTENTS Introduction..............................................................................................................97 Electrodes and Surgical Procedures ........................................................................99 Fixed Electrode Array Specifications...........................................................99 Moveable Electrode Arrays for Layer Analysis.......................................... 100 Implantation and Advancing through the Layers ....................................... 100 Histology..................................................................................................... 101 Multichannel Recording ........................................................................................ 101 Hardware and General Recording Procedure............................................. 101 Somatosensory Recordings in Anesthetized and Behaving Animals ................... 101 Precise Whisker Stimulation in Anesthetized Rats.................................... 101 Somatosensory Stimulation of Head-Immobilized Awake Rats ................ 103 Tactile Stimulation and Recording in Freely Moving Rats ........................ 105 Nerve-Cuff Stimulation ................................................................... 105 Active Tactile Discrimination.......................................................... 106 Rat Behavioral Testing Chamber Details ................................................... 109 Combining Ensemble Recordings with Reversible Focal Inactivation.................. 112 Neural Ensemble Data Analysis ............................................................................ 115 Single-Neuron Responses ........................................................................... 115 LVQ-Based Ensemble Analyses ................................................................. 117 Conclusion and Outlook......................................................................................... 120 Resources ............................................................................................................... 121 References.............................................................................................................. 122

INTRODUCTION A fundamental goal in systems neuroscience is to explain animal behavior in terms of the dynamics of neural ensembles. Multielectrode techniques greatly facilitate the approach toward this goal. Aside from the fact that each experiment provides a higher yield of data as compared to single-site recordings, some questions simply cannot be addressed using only one electrode at a time. For example, only multisite 97 © 2008 by Taylor & Francis Group, LLC

98


recordings can determine whether different neurons respond independently to stimuli, or covary from trial to trial. The purpose of this chapter is to review methods used in multielectrode studies of the rat somatosensory system, with an emphasis on the whisker system. We present a basic toolbox of methods we have used to probe the functions of populations of somatosensory neurons in a behavioral context. The basic toolbox includes techniques for applying controlled whisker stimuli, behavioral training in tactile discrimination tasks, multielectrode recordings, reversibly inactivating specific brain areas, and analysis of the ensemble neural data. These methods have already revealed fundamental properties of the somatosensory system that would have been difficult or impossible to uncover using singleelectrode recordings. For example, cortical (Zhu and Connors, 1999; Ghazanfar and Nicolelis, 2001; Diamond et al., 1992; Ghazanfar et al., 2000; Schubert et al., 2001) and thalamic (Armstrong-James and Fox, 1987; Nicolelis and Chapin, 1994) neurons have large multiwhisker receptive fields that are dynamic over poststimulus time (Nicolelis and Chapin, 1994; Ghazanfar and Nicolelis, 1999; Ghazanfar et al., 2000). These data, together with observations of supralinear summation of multiwhisker inputs (Ghazanfar and Nicolelis, 1997; Shimegi, 2000), suggest that tactile receptive field dynamics function to integrate time-varying multiwhisker inputs (Ghazanfar and Nicolelis, 2001). For example, analysis of multineuron response data revealed additional stimulus-coding properties of somatosensory ensembles. S1 ensembles code stimulus location in single-neuron temporal patterns and the relative response latencies of their neurons, but not in single-trial covariations among the neurons (Nicolelis et al., 1998; Ghazanfar et al., 2000). In S2 of the primate, on the other hand, single-trial covariations among multiple neurons did contribute significantly to coding the location of a punctate stimulus (Nicolelis et al., 1998; Ghazanfar et al., 2000). Even in S1, the contribution of coordinated firing may increase with greater stimulus complexity, because multiple whisker stimuli lead to a higher prevalence of synchronous responses between neurons in the infragranular layers of S1 than in other layers (Zhang and Alloway, 2005). Combining methods for inactivating specific neural inputs with ensemble recordings led to the further conclusion that spatiotemporal RF properties of somatosensory neurons arise not only from intrinsic local properties of neurons and their neighbor connections, but rather from interactions among multiple levels of the somatosensory system. For example, recording thalamic tactile responses in the presence and absence of cortical feedback revealed that corticofugal projections contributed to both the short- and long-latency components of ventral posterior medial nucleus (VPM) responses (Krupa et al., 1999; Ghazanfar et al., 2001). These interlevel interactions were reflected in simultaneous recordings in trigeminal areas in brain stem, thalamus, and cortex, which revealed widespread oscillatory synchronization of neural firing (Nicolelis et al., 1995). The correlated activity remains even after transection of the facial nerve, which suggests that such synchronous activity is generated centrally. Although the high coherence among large populations of neurons associated with this oscillatory 7–12 Hz brain state suggested absence seizures to a number of authors (Marescaux et al., 1992; Shaw et al., 2006; Shaw, 2007), a direct test showed that rats respond robustly to mild tactile stimulation during bouts of 7–12 Hz oscillations in S1, contradicting the absence interpretation (Wiest and Nicolelis, 2003).


Multielectrode Recordings in the Somatosensory System

99

Thus, widespread synchronized neural firing need not preclude perception; in fact, it can enhance aspects of sensory representation (Fontanini and Katz, 2006) as well as long-term plasticity (Erchova and Diamond, 2004). These demonstrations of fast interactions among neurons distributed across the somatosensory maps at multiple processing stages were paralleled by demonstrations of a tight coupling between the two hemispheres of S1 (Shuler et al., 2001; Wiest et al., 2005). This cross-talk challenges the classical conception of the S1 barrel cortex as an encoder for exclusively contralateral whisker activity and suggests the potential importance of bilateral interactions in S1 for whisker-guided discriminations (Krupa, 2001b; Shuler et al., 2002). Multielectrode recordings in different layers of S1 while rats performed a bilateral whisker-guided discrimination revealed that a feed-forward model of tactile signal processing cannot explain S1 response properties (Krupa et al., 2004). For example, firing rate modulations that began before tactile stimulation clearly could not be explained in terms of bottom-up propagation of a stimulus signal. Rather, other inputs to S1 must contribute to shaping the task-related responses. Similarly, tactile responses were found to vary significantly in different spontaneously occurring behavioral states (Nelson, 1996; Fanselow and Nicolelis, 1999; Moore et al., 1999; Nicolelis and Fanselow, 2002; Castro-Alamancos, 2004; Moore, 2004). These data collectively suggest that widely distributed neurons coordinate their activities on millisecond time scales, and that the functional connectivity among them can be quickly adjusted in different behavioral contexts. The preceding examples are meant to indicate the range of results that have already been achieved using multielectrode arrays (MEAs). In the following sections we present specific methods developed in the past 15 years. The examples have been selected to represent methods from each major phase of a typical study, from electrode design and surgical implantation, through ensemble recording involving somatosensory stimulation, behavioral monitoring, and reversible inactivation of specific brain areas, to analysis of the recorded many-neuron data.

ELECTRODES AND SURGICAL PROCEDURES FIXED ELECTRODE ARRAY SPECIFICATIONS To the extent possible, the multielectrode arrays should be tailored to the scientific goals and the anatomical areas under study (see chapter 1). To study the properties of output neurons in layer V of S1 cortex in the rat (Ghazanfar et al., 2000; Ghazanfar et al., 2001; Shuler et al., 2001; Wiest et al., 2005), we would use 32-electrode arrays with 2 rows of 16 electrodes, with 250 µm spacing between rows and between electrodes within a row. This configuration affords broad sampling of the large cortical whisker representation in rats. The standard electrodes are 35-µm-diameter tungsten wire insulated by S-Isonel, attached to Omnetics 32-pin connectors. Two stiff prongs protrude from either side of the connector by which to hold the array during implantation; these are cut off after the array has been cemented in place during surgery. For a layer-V target (~1.3–1.5 mm below the cortical surface), 3.5 mm of electrode length is left exposed with blunt, i.e., flat-cut, tips (see chapter 1 for details).


100


Similar arrays, often with a 4 w 8 configuration, are used to record VPM neurons. Although electrode bundles were commonly used in the past to access thalamic neurons, even deeper structures such as the brain stem can be probed using electrode arrays with long wires.

MOVEABLE ELECTRODE ARRAYS FOR LAYER ANALYSIS For a study comparing the firing and coding properties of neurons in different cortical layers during tactile discrimination behavior, it was necessary to record ensembles of single-unit activity through different depths in S1 cortex. Krupa (Krupa et al., 2004) developed a movable electrode array that incorporated ultrafine wire electrodes (12 µm diameter) into movable arrays of 16 or 24 microelectrodes. The 16-channel arrays were arranged in a 5/6/5 pattern, with microelectrodes 250 µm apart. The 24-channel arrays were arranged in a 4 w 6 pattern, 250 µm apart. The very small exposed tips of these ultrafine microwires yielded high-impedance electrodes (Z = 2.5–3 MΩ @ 1 KHz) capable of recording high signal-to-noise singleunit activity throughout the different layers of S1.

IMPLANTATION AND ADVANCING THROUGH THE LAYERS The basic surgical procedures for electrode implantation are described in further detail in chapter 2. Briefly, for implantation into S1 cortex, rats are anesthetized with ketamine (100 mg/kg) and xylazine (8 mg/kg) and placed in a stereotaxic head holder. (In some cases Pentobarbital (50 mg/kg i.p.) is used as the anesthetic for female rats.) Using a dental drill, a small craniotomy is made over the barrel cortex (3 mm caudal from bregma, 5.5 mm mediolateral), exposing the brain surface covered by the dura. For rat implantations we generally transect the dura before inserting the array into the brain, or the electrodes may not penetrate and the brain would be compressed instead. Our standard method of removing the dura is to use a large hypodermic needle with a bent tip to hook, lift, and cut the dura, so that it can be retracted with forceps. The array is then lowered into S1 no faster than 100 µm/min. Finally, the electrode assembly is fixed in place with dental acrylic. For the study using moveable arrays (Krupa et al., 2004), prior to implantation of the array a single sharp tungsten microelectrode (Z = 1.5 MΩ @1 KHz) was lowered approximately 0.7 mm into several locations within the region of the barrel cortex while neural activity recorded by the electrode was monitored. The contralateral large facial whiskers were manually stimulated to locate the C3 barrel as well as the general orientation of the barrel region. This procedure allows the array to be centered over the C3 barrel. The moveable arrays were only lowered 100 µm into the cortex at the time of surgery. They were oriented normal to the S1 cortical surface so that as they traversed multiple depths they would sample activity from a single cortical column. In general, rats receive at least 7 days of postsurgical recovery, during which no behavioral training takes place. During this time rats implanted with movable electrodes would have their arrays advanced approximately 20–40 µm per day until single-unit activity was detected. Prior to data collection, rats involved in a behavioral study received at least two behavioral discrimination sessions with the multichannel recording headstages and cables attached, but without recording, to allow rats to



101

adapt to performing the task with the attached cables. All rats adapted quickly; task performance before surgery and after these adaptation sessions did not differ.

HISTOLOGY At the end of all recording sessions, small electrolytic marking lesions (20 µA/5 s) were made through several electrodes within each array. Rats were injected with a lethal dose of pentobarbital and perfused with saline and 10% formalin. Brains were removed, fixed in formalin/sucrose, sectioned, and stained with cresyl violet to identify the marking lesions. The depth of the marking lesions was used to confirm the depth of the electrode arrays during different recording sessions. We have thus far chosen to use coronal slices to verify the depth of our electrode tips, but for some scientific questions it may be desirable to section brains tangentially to determine electrodes’ positions with respect to the cytoarchitectonic barrel/septa distinction in layer IV, which has been proposed to have physiological correlates in other layers as well (Castro-Alamancos, 2004).

MULTICHANNEL RECORDING HARDWARE AND GENERAL RECORDING PROCEDURE Single-unit (SU), multiunit (MU), and local field potential (LFP) activity are recorded through a Many-Neuron Acquisition Processor (MNAP; Plexon, Inc., Dallas, Texas). The data are sampled at 40 KHz. Units are first sorted online during each recording session, according to the shapes of supra-threshold spikes. Waveforms whose amplitude is more than two standard deviations above the background noise are saved for offline sorting (Nicolelis et al., 2003). Unless more stringent criteria are applied to the online-sorted units, such as a test for spike conflicts—i.e., multiple spikes falling within a neuron’s 1–2 ms refractory period—then the online-sorted units, which may include spikes from multiple neurons, are termed MUs. Digitized spike waveforms are also stored on computer hard disk for offline analysis, so units are further sorted offline according to their clustering in the principle component space representing the spike waveforms. Those sorted units that show a distinct cluster in principle component space from the cluster of noise waveforms, and displayed fewer than 0.1% of interspike intervals within a refractory period of 1 ms, are termed SUs. Following each recording session with movable arrays, the arrays were typically lowered a total of 100 µm (in 20–40 µm steps) before the next recording session to ensure that the arrays were sampling different populations of neurons. This procedure was repeated until the electrode arrays had been driven through the entire thickness of cortex.

SOMATOSENSORY RECORDINGS IN ANESTHETIZED AND BEHAVING ANIMALS PRECISE WHISKER STIMULATION IN ANESTHETIZED RATS To study somatosensation in the rat we have focused on the system of large facial whiskers that rats rely on for many purposes, such as navigation and object recognition


102


A

B

FIGURE 6.1 (See color insert following page 140.) Multiple individual whisker stimulator for (A) anesthetized and (B) head-restrained awake rats.

in the dark. The combination of a discrete array of sensors—the vibrissae—and a clear corresponding topographical representation at each ascending level of the somatosensory system provides a convenient framework for addressing questions of sensory coding and integration. The absence of gross movements and fluctuations of behavioral state in anesthetized rats affords the opportunity to observe neural responses to arbitrary finely controlled multiwhisker stimuli. Of particular interest are questions of multiwhisker integration, neural plasticity related to repetitive stimulation of particular combinations of inputs, and neural responses to complex naturalistic patterns of stimulating many whiskers. To deliver precise multiwhisker stimulation, Krupa et al. (2001a) developed a multiwhisker stimulator incorporating 16 stainless steel wires (130 µm diameter) attached to computer-controlled miniature solenoids (Figure 6.1a). The other ends of the wires are attached to individual facial whiskers, such that activation deflects whiskers by 1–2 mm at approximately 1 mm/ms. This system affords greater stimulus control than single-whisker stimulators or air puffs, for instance. For example, in our study (Krupa et al., 2004) comparing somatosensory neural responses during active tactile discrimination (see below) to those under anesthesia and passive awake stimulation, we wanted to duplicate in an anesthetized rat the spatiotemporal dynamics of whisker deflections that occurred during the active discrimination (i.e., an obstacle sweeping rostrocaudally across the whiskers). This passive stimuli pattern consisted of ramp-and-hold deflection of 16 individual whiskers in



103

three rats that were lightly anesthetized with pentobarbital (40 mg/kg, subcutaneous). The multiwhisker stimulator was attached to each of 16 facial whiskers (rows B–E, arcs 0–3) on one side of the face. The patterned stimulation consisted of the following: four whiskers in arc 3 (B3-E3) were deflected simultaneously with a ramp-andhold deflection for 150 ms; 25 ms after the onset of arc 3 deflection, four whiskers in arc 2 (B2-E2) were deflected similarly; 25 ms after arc 2 stimulation onset (50 ms after arc 3 stimulus onset), arc 1 whiskers (B1-E1) were stimulated; followed 25 ms later (75 ms after arc 3 stimulus onset) by arc 0 whisker (four straddlers) stimulation. Thus, the facial whiskers were stimulated in a pattern that simulated a tactile stimulus moving caudally across the whisker pad over a period of ~75 ms, similar to the pattern of whisker deflection during the active task. Three hundred stimuli were delivered at 1 Hz while S1 activity was recorded. This stimulation was very well controlled and precise, but may not have mimicked the variability or other properties of a solid edge sweeping across the whiskers (as in the active task), so we also recorded responses in that study using a moving aperture (of the same dimensions as the aperture in the active discrimination) swept across the whiskers. The moving aperture was computer controlled and powered by pressurized air. The onset of aperture movement, velocity of movement, and trajectory of movement were pseudorandomly varied from trial to trial based on velocity and trajectory parameters that were obtained by video analysis of rats performing the active discrimination. This resulted in whisker deflection dynamics that approximated the trial-to-trial lateral jitter in whisker deflections that occurred during the active discrimination. This sliding aperture stimulus was also applied to awake rats by acclimating them to head immobilization as described in the following text. The multichannel stimulator has been used in the Nicolelis lab to study multiwhisker (Ghazanfar et al., 2000; Ghazanfar et al., 2001) and bilateral integration (Shuler et al., 2001), and has been adapted for use in awake rats (Wiest and Nicolelis, 2003; Wiest et al., 2005) (see following text).

SOMATOSENSORY STIMULATION OF HEAD-IMMOBILIZED AWAKE RATS In extending somatosensory electrophysiology to awake animals, we strive to limit the greater variability associated with the awake state. One strategy is to prevent bodily movements and the uncontrolled sensory activity they cause, by immobilizing rats’ heads (Bermejo et al., 1996; Bermejo et al., 1998; Harvey et al., 2001; Wiest and Nicolelis, 2003; Wiest et al., 2005). (An alternative approach to limiting both stimulus variability and confounding motor-related activity during somatosensory ensemble recordings is to train rats to repeat stereotyped active samplings of a tactile stimulus, as described in a later section.) For head-immobilized awake recordings, a brass head post (1/4” width, Small Parts, Inc.) was fixed to the dental-cement headcap of subjects during the surgery for electrode array implantation. These rats were first habituated to head-fixed restraint in a Plexiglas restraint tube in which the rat’s body and all four feet were comfortably confined in a light drawstring “jacket” that minimized the subject’s traction against the Plexiglas tube, while its head was held fixed by a mounting post embedded in the dental acrylic. The head post fit snugly into a pin vise (Small Parts, Inc.) attached to the restraint


104


tube to prevent all head movement. This design was an improvement over a previous practice of using round screws as head-bolts. Rats were often able to rotate their heads slightly when immobilized by a round head-bolt, and any slight freedom of movement seems to aggravate further struggling. To facilitate acclimation to restraint, rats were placed on a mild food deprivation schedule and calorie-dense liquid reward was delivered at random intervals during sessions in the restraint tube. Habituation was achieved by gradually increasing the amount of time that rats were confined in the restraint tube each day over a period of ~2 weeks. Prior to implantation, it was helpful to handle the rats daily, and acclimate them to restraint in drawstring jackets as well as to the reward solution (rats are often initially averse to novel tastes). However, attempts to acclimate the rats to full restraint in the Plexiglas tube before implantation of the head bolt appeared to be counter-productive, as the rats would tend to struggle and panic, leading to an unpleasant association with the restraint tube rather than a comfortable one. We adapted the computer-controlled multiple individual-whisker stimulator (Krupa, 2001a) described earlier for use on an awake animal. Specifically, each of two stimulator arms was fitted with a thread noose, which could be constricted around an individual whisker. In Figure 6.1b (right panel) the stimulator noose (at the tip of the silver stimulator arm) is attached to a single whisker. The noose could easily be fitted over groups of whiskers as well. For example, the whisker stimulator was used (Wiest et al., 2005) to apply controlled 5 ms rostral deflections of the bottom four whiskers in a single column on each side of the subject’s face (Figure 6.3b, left panel). To lasso four whiskers, the noose was first opened wide, then fine forceps were used to pull through the whiskers one at a time before tightening the noose about 1 cm from the rat’s face. After acclimation rats allow this manipulation, relax their whisker pads, and only attempt to whisk when there is a noise or large movement in the room. In a study of bilateral integration between the S1 hemispheres (Wiest et al., 2005), two basic stimulation protocols were run. In both experiments, stimuli were presented in random order at random intertrial intervals of between 2 and 4 s. In Experiment 1 the stimulus set consisted of a left arc deflection, a homologous right arc deflection, and simultaneous bilateral deflections. In Experiment 2 the stimulus set consisted of a left, then right, arc deflection at interstimulus intervals (ISIs) of 0, 30, 60, 90, and 120 ms. TEMPO software (Reflective Computing, St. Louis, MO) controlled the stimulus administration and sent time-stamped stimulus events with millisecond accuracy to the file of neural recordings. Usually, at least 200 stimulus presentations could be delivered during an hour-long recording session. For another set of experiments (Wiest and Nicolelis, 2003) the head-immobilization setup was expanded to accommodate a behavioral response from the rat. This study was aimed at comparing rats’ behavioral responsiveness during a highly coherent oscillatory cortical state and a desynchronized state. A solenoid-gated reward tube was positioned at the animal’s mouth to deliver drops of liquid reward. An optical beam sensor was positioned directly in front of the reward tube such that it reliably registered the animals’ licks for reward. Programming the reward solenoid to release a drop gated by the animals’ licks greatly facilitated the process of



105

acclimating the animals to head restraint. Drops of reward were initially made available only once per second to avoid satiation or spillage. After acclimation, drops of reward could be made contingent on a perceptual task. In particular, we trained rats to inhibit their licks for reward until the stimulator delivered a 5 ms, ~1 mm deflection of four whiskers from a single arc. Sessions were conducted in the presence of white noise to mask any auditory cues, and control sessions with stimulator unattached to whiskers verified that whisker cues were used exclusively to detect the stimulus. To initiate a trial, the rat had to refrain from licking for at least 1 s. A peaked response (lick latency) histogram following stimulation shows that the rats were responding to the stimulus rather than licking randomly. Occasionally, the rat is startled by the stimulus, and instead of a lick, the sensor registers a startle movement of the jaw. These events are obvious to an observer and can be clearly segregated in the response histogram as a sharp peak at a latency of about 30 ms. Thus, this headfixed paradigm combined with a lick detector and reward dispenser appears to have the potential for a variety of rat somatosensory psychophysics.

TACTILE STIMULATION AND RECORDING IN FREELY MOVING RATS Nerve-Cuff Stimulation For a study comparing tactile responses during different behavioral states (such as the quiet awake state and the exploratory whisking state), it was desirable to stimulate the somatosensory system without constraining the animal’s natural behavior in any way. To achieve this, Fanselow and Nicolelis (Fanselow and Nicolelis, 1999) custom made cuff electrodes to encircle the infraorbital nerve, the branch of the sensory division of the trigeminal nerve that carries tactile information from the facial vibrissae to the trigeminal ganglion. Two platinum bands (Goodfellow, Berwyn, PA) separated by 0.8 mm, were connected to flexible three-stranded Teflon-coated wires (AM Systems, Everett, WA) so that current could be passed between the bands. The whole construct was embedded in a thin film of Sylgard (Factor II, Lakeside, AZ) to hold the electrode together and provide electrical insulation. The inside platinum faces were left exposed for nerve stimulation, and a piece of surgical silk was attached to the outside of the band for maneuvering the electrode during implantation. The inner diameter of the finished cuff was about 1.7 mm. During surgery a dorsoventral incision was made on the face lateral to the infraorbital nerve, and tissue was dissected until the nerve was exposed and perineurium was cleared away. Surgical silk was inserted under the nerve so the cuff electrode could be drawn under and around the nerve, just rostral to the infraorbital fissure. The cuff was tied off using the attached silk. Wires from the cuff electrode were led subcutaneously to the connector on the skull. During the recording session, stimuli were generated using a Grass S8800 stimulator in conjunction with a Grass PSIU6 stimulus isolation unit. 100 µs pulses of 5–9 mA in amplitude were typically successful in evoking neural responses. The current amplitude was set to 1 mA higher than the threshold for evoking responses. The neural responses were found to deliver highly consistent stimuli that moreover did not depend on the position of the animal’s whiskers. Neural responses to this


106


artificial stimulation were very similar to those obtained by mechanical deflection of individual whiskers, except for a shorter latency. To minimize interference with the animal’s movements, recording cables were connected to a rotating commutator at the top of the recording chamber. This stimulation apparatus was subsequently applied as an alternative means for disrupting pentylenetetrazole-induced seizure activity in rats (Fanselow et al., 2000). In contrast to vagus nerve stimulation studies, no substantial cardiovascular side effects were observed by unilateral or bilateral stimulation of the trigeminal nerve. Moreover, by triggering trigeminal nerve stimulation on the occurrence of seizure episodes, this method resulted in more effective and safer seizure reduction per second of stimulation than with previous methods, and may thus lead to improved human treatments. Active Tactile Discrimination The nerve cuff stimulation is appropriate for a study comparing neural responses to identical tactile stimulation in different behavioral states. To study the neural mechanisms by which rats actively sample and discriminate distinct tactile stimuli, however, another approach is required. To this end, a whisker-dependent tactile discrimination task was developed (Krupa et al., 2001b) in which a rat repeatedly samples a variable-width aperture with its facial whiskers, by advancing its head through the aperture (Figure 6.2). At the end of each trial the rat is required to go to a spout located to the left or the right depending on whether the aperture discriminandum was narrow or wide, respectively. Computer-controlled behavior boxes (see section “Behavioral Testing Chamber Details” below) were built in-house so that the discrimination task with well-trained rats proceeded as follows. At the start of each session, rats were placed in the outer reward chamber (the main chamber in Figure 6.2a). The first trial of a session begins when the sliding door between the outer reward chamber and the inner discrimination chamber opens by computer control. Rats quickly proceed into the center discrimination chamber and toward the center nose poke. In doing so, rats disrupt an infrared photobeam in front of the variable-width aperture and, immediately after, their large facial whiskers contact the variable width aperture. Upon breaking the infrared photobeam and sampling the variable width aperture, rats then back out into the reward chamber. To receive water reward, rats have to correctly poke their nose into either the left or right nose poke: left nose poke if the aperture was narrow; right nose poke if the aperture was wide. Rats perform the task in complete darkness to eliminate the use of visual cues. To verify that they relied not only on tactile cues, but specifically on their large vibrissae, the vibrissae are cut at the end of an experiment and the rats are run through the task again to see that their performance is reduced to chance levels. To motivate them for training, the rats are placed on the following water restriction schedule. Between 15 and 60 min after the end of each daily training session, rats receive free access to water for 15–60 min; food is available ad libitum. Once a week rats are given free access to water for 24 h followed by water restriction for 24 h, at which point training is resumed. Thus, for each 7-day period, rats receive 6 daily training sessions (with 1 h access to water after each session), followed by



107

A Reward Solenoid Valves (1–3)

Camera

Whisker Discriminator 1

Nose Poke Center

Whisker Discriminator 7

Whisker Discrimination Module

Nose Poke Left

Center Door Main Chamber Camera

Chamber Door

Nose Poke Right Reward Solenoid Valves (4–6)

B

FIGURE 6.2 Whisker-dependent aperture width discrimination task. (A) Schematic of the behavioral task. A rat begins in the main chamber, and enters the rear whisker discrimination chamber when the sliding center door opens. The rat samples the variable width aperture with its whiskers and retreats to the left for reward if the aperture was narrow (60 mm) and to the right if the aperture was wide (68 mm). (B) Video frame capture showing a rat sampling a narrow aperture.

1 day of free access to water. This restriction schedule is repeated throughout an entire series of experiments. All rats have shown significant weight gains over the course of training, and no health problems have occurred with this protocol.


108


Typical narrow- and wide-aperture widths of 60 and 68 mm are moderately difficult for rats to discriminate, but they are capable of significantly finer discriminations (62 and 65 mm) (Krupa et al., 2001b). Rats receive a 50 µL water reward for a correct response and no reward for an incorrect response. Immediately after rats poke into either the correct or incorrect reward nose poke, the center door between chambers closes and the aperture is randomly reset to either wide or narrow. The next trial begins up to 30 s later when the center door was again opened. Wide and narrow trials are presented randomly throughout each 75 min training session. Well-trained rats typically performed ~80 or more trials per session with approximately 80–85% correct discriminations using an intertrial interval of 30 s. This long intertrial interval was used initially so that rats would be motivated to correctly discriminate rather than accepting the rate of reward associated with 50% correct performance. However, recently we have been able to train rats to routinely perform on the order of 200 trials per session using intertrial intervals as low as 5 s. The ability to record a relatively large number of trials per session is an important and attractive feature of this behavioral paradigm: reliable scientific conclusions require sufficiently powerful statistics. The number of trials recorded per session becomes increasingly important, for example, in variants of the task that present more than two different tactile stimuli to be discriminated. Training continues until the rats correctly perform the discrimination at criterion (typically 75% correct) accuracy for at least three consecutive sessions. Rats are then removed from water restriction before surgical array implantation (as described earlier). Following at least 7 days of postsurgical recovery, rats are returned to water restriction and the behavioral discrimination training. After reacclimation to the task, neural ensemble activity can be recorded during active discrimination behavior. A detail may serve to illustrate the kinds of considerations that go into choosing or designing a particular training protocol. In initial studies (Krupa et al., 2001b), rats were required to trigger the center nose poke (see Figure 6.2) during their sampling of the aperture discriminandum. Later (Krupa et al., 2004), rats were only required to break the aperture photobeam and not the center nose poke beam. This was done to minimize the possibility that rats’ whiskers might be stimulated as they fully poked their nose into the center nose poke. Moreover, because rats did not have to break the center nose poke photobeam, the total amount of time that rats were required to sample the aperture, as well as trial-to-trial variability in contact and sampling times, were minimized. This resulted in a very repeatable, stereotypic whisker stimulation. A highly consistent stimulus is as desirable as a large number of trials. On the other hand, in the mandatory center nose poke protocol there is a longer delay between whisker contact with the discriminandum and the beginning of the rat’s movement toward the left or right reward spout. This delay is actually an advantage in terms of dissociating sensory representation of the narrow or wide stimulus from motor-related somatosensory responding due to left or right movements. During that delay, left or right movement-related activity will not be confounded with wide or narrow sensory representations. There may be an unavoidable trade-off between the speed of trials in the photobeam-only protocol and this potential interpretational advantage of the mandatory central nose poke protocol.



109

This behavioral paradigm for active tactile discrimination is flexible, in that the training boxes and protocols can be adapted for a number of different tasks. For example, Shuler et al. (2002) adapted the wide versus narrow discrimination by adding a third, asymmetric stimulus configuration, such that rats were forced to compare information from the two sides of their face to successfully perform the task. Recently, Pereira et al. (2005) have adapted the apparatus to develop a texture discrimination task, by attaching removable magnetic strips (built by Jim Meloy) to the aperture. Sandpaper strips of different coarsenesses are glued to the magnetic strips to provide texture cues. Two different textures (60 grit “rough” and 120 grit “smooth”) are affixed to two different apertures in a single behavior box. By retracting one aperture and presenting the other, a single texture can be presented on each trial. As a control to verify that the animal uses the texture cue rather than the position of the two apertures, the position of the texture strips can be reversed during some sessions. Another variant of the task uses four equally spaced aperture widths to study the neural correlates of category perception (Thomson et al., 2005).

RAT BEHAVIORAL TESTING CHAMBER DETAILS The rat behavioral testing chamber (Figure 6.2) is a multifunctional rat behavioral testing apparatus designed to provide a platform for tactile discrimination experiments. System capabilities include 20 control inputs, 32 signal outputs, and a choice of 36 selectable timing outputs. The chamber (Figure 6.2) consists of the main chamber and the whisker discrimination module. The main chamber is the primary holding area and provides the mounting location of two nose poke modules for liquid reward delivery. The whisker discrimination module contains a center nose poke module and seven discrete whisker discrimination bars with infrared photo-beam detectors. The main chamber and the whisker discrimination module are separated by a pneumatically operated center door. There are three reward solenoids valves on each side of the chamber that deliver liquid rewards to the left and right nose pokes. There is a printed circuit board black and white 1/3 in. CCD (60 fields/s) camera located in both the main chamber and the whisker discrimination module with outputs to Panasonic BW-BM 990 video monitors. The chamber base consists of a frame and cap. The frame is 2.0 in. w 0.75 in. PVC and the cap is high-density polyethylene. The Main Chamber and the Whisker Discrimination Module are constructed of 0.125 in. aluminum, and the chamber door and chamber cap are constructed of 0.25 in. clear acrylic sheet. The nose poke modules are constructed of 1.0 in. black Acetal (Delrin). Overall unit dimensions are 22.5 in. (L) w 16.0 in. (W) w 16.5 in. (Height). A system overview of the behavioral chamber follows: Main circuit board: The control board is the primary system electronic interface to the behavioral chamber; it provides power distribution, photo-beam processing, input/output isolation for timing signals, and the Med Associates interface for input/output control.


110


Pneumatic controller: There are three double-acting pneumatic cylinders used in the chamber that operate the left nose poke door, right nose poke door, and center door. The pneumatic controller receives input commands from the main circuit board and activates four-way pneumatic valves. The pneumatic valves are fitted with manually adjustable speed controllers to regulate the opening and closing speed of the doors. An external 20 ft3/min compressed air source is regulated to 35 lb/in2 prior to supply routing to the pneumatic valve. Manual control: A printed circuit board mounted on the chamber cabinet to provide manual activation of outputs such as reward solenoids, pneumatic door solenoids, and house lighting. Plexon interface: This printed circuit board inputs 36 timing events from the main circuit board and allows the user to select and hardwire 14 of the timing events to send to the Plexon DSP module. Plexon DSP module: The module is configured for 14 individual TTL events, asserted Low True. Parallax BSII: The BASIC Stamp 2 (BS II) is a 24-pin DIP (dual inline package) microcontroller module. It is the main controller for the 14 servos that drive the whisker discriminator bars (Figure 6.2; see following text). The BS II receives a series of command pulses from a Med Associates DIG-726 output card and then issues serial commands to the Mini SSII Servo Controller, which in turn position the servos to the appropriate settings. The BS II is programmed in the PBASIC language on the host PC, and the programs are loaded via the host computer’s serial port to a DB-9 connector on the BS II carrier board. Mini SSC II servo controller: The Scott Edwards Electronics, Inc., Mini SSC II is an electronic module that controls seven pulse-proportional servos according to instructions received serially from the BSII. There is a Mini SSC II for the left set of whisker discrimination bars and another for the right set. Servo: The drive servo is a Futaba S148 pulse-proportional servo with a linear servo conversion kit. The servo receives position information from the Mini SSC II Servo Controller and positions the whisker discrimination bar through a series of push rods and gears. Med Associates input/output: System command and control is accomplished through products provided by Med Associates, Inc. MedState Notation protocols are translated and compiled with Trans IV, and MED-PC IV is the run-time operating system. Outputs are controlled with DIG-726 SuperPort Output Modules and inputs with DIG-712 SuperPort Input Modules. Whisker discrimination module: The whisker discrimination module (Figure 6.3) is used for tactile discrimination experiments. The primary operating feature is the set of whisker discrimination bars (WDBs) (Figure 6.3). The WDB is constructed from aluminum stock with dimensions 4.0 in. Height (H) w 3.5 in. Length (L) w 0.25 Thickness (T). On the face of the WDB, there are three infrared photo-beam ports (Figure 6.2b). The bars on the right side of the module contain infrared emitters, and the bars on the left contain infrared phototransistors. Phototransistor status information is sent to the Main Circuit Board and a



111

A

Whisker Beam Break Indicator

B Whisker IR Emitter

Triple IR Beams

Whisker IR Detector

FIGURE 6.3 Whisker discrimination module. (A) Photo of the module showing center nosepoke and seven adjustable whisker discriminandum bars in the fully retracted position. (Krupa, D.J, Brisben, A.J., and Nicolelis, M.A.L. (2001a) A multi-channel whisker stimulator for producing spatiotemporally complex tactile stimuli. J Neurosci Methods 104: 199–208.) (B) Schematic of single-pair whisker discriminandum bars showing placement of infrared beams triggered by the rat’s head passing between the bars (horizontal beams) and by whisker contact with a bar (vertical beams). (Wiest, M.C., Bentley, N., and Nicolelis, M.A. (2005) Heterogeneous integration of bilateral whisker signals by neurons in primary somatosensory cortex of awake rats. J Neurophysiol 93: 2966–2973.)

circuit containing an OR Gate integrated circuit triggers a beam break output if any of the three phototransistors signal a beam break. Each WDB is driven by a pulse-proportional servo through a gearing mechanism. The travel distance achieved is 40.0 mm with a resolution of ± 500 µm. Experimental apertures are typically in the range of 85 to 50 mm. Whisker contact detector: An optional feature of the WDB is a whisker detection device (Figure 6.3b). The whisker detector consists of a photo


112


beam produced by an 880 nm infrared emitter and detected by an 880 nm phototransistor. The infrared emitter and phototransistor are installed in sealed housings with an orifice of 0.150 mm and attached to the leading edge of the WDB. A special printed circuit board is used to drive the emitter, adjust the sensitivity of the phototransistor, and send a beam break signal to the main circuit board. Additionally, there is a beam break indicator infrared LED installed on the top edge of the WDB to provide the operator a visual beam break indication through the video system.

COMBINING ENSEMBLE RECORDINGS WITH REVERSIBLE FOCAL INACTIVATION Techniques for manipulating selected populations of neurons during a recording experiment are a powerful means of testing theoretical models of neural dynamics. One such tool that has been used repeatedly in the Nicolelis lab is reversible pharmacological inactivation of local neural populations with muscimol. Muscimol is a GABAA agonist. This agent has no effect on fibers of passage that may cross the area being infused and allows cortical activity to return to control levels once the effect of the injection wears off. The combination of muscimol infusions with multielectrode recordings of cortical neuronal activity through arrays of microwires spaced at known distances from the S1 infusion site provides a precise measure of the effective spatial spread as well as the time course of muscimol inactivation (Figure 6.4). Previous experiments have shown that by varying the dose of muscimol, the effective spread of the drug can vary from a few hundred micrometers to several millimeters. Because the neural activity on each of the microwires is continuously monitored, the effects of the drug can be tracked throughout the entire experiment. This quantitative spatiotemporal monitoring of the drug’s effects is a significant improvement over purely behavioral or single-electrode measures of a drug’s effect. For inactivating the deeper layers of S1 cortex, for example, a 27-gauge (0.016 in. OD.) thin-walled stainless steel cannula is implanted in the infragranular layers of the area of interest, just adjacent to a 32-microwire array. The cannula is positioned either in the middle or at the end of the microwire array, approximately 300–500 µm away from the rostral-most microwire. The depth of the cannula is approximately 1.0 mm below the pial surface—slightly superficial to the depth of the electrode tips. FIGURE 6.4 (see facing page) Muscimol time course and spatial spread. (A) Control experiments demonstrate the stability of our neural ensemble recordings during cortical inactivation with muscimol (150 ng dose). Each row shows poststimulus time histograms (PSTHs) from a neuron on a different electrode, at different postinfusion times. In this case inactivation lasted about 9 h before preinfusion responses reappeared. (B) Ensemble recordings are also used to map the spread of muscimol inactivation in the S1 cortex. Each column shows PSTHs from a different electrode, spaced 250 microns from its neighbors. Inactivation spread less than a millimeter for the lower dose (50 ng), but more than a millimeter for the higher dose (150 ng; adapted from Krupa, D.J., Ghazanfar, A.A., and Nicolelis, M.A. (1999) Immediate thalamic sensory plasticity depends on corticothalamic feedback. Proc Natl Acad Sci U.S. 96: 8200–8205. With permission). (Krupa, D.J., Ghazanfar, A.A., and Nicolelis, M.A. (1999) Immediate thalamic sensory plasticity depends on corticothalamic feedback. Proc Natl Acad Sci U.S. A. 96: 8200–8205.)


A

100 50 0 100 50 0

dsp 100 21a 50 0 dsp 100 23a 50 0 dsp 24a

1 Hour

40 20 0 –50

10

9 Hours

dsp 18a 60

20 10 0

100 50 0

2 1 0

1 0.5 0

150 75 0

150 75 0

dsp 19a 60

12 Hours

1 0.5 0

1 7.5 1 40 0.5 20 3 0.5 0 0 0 0 –50 0 50 100 –50 50 100 –50 50 100 –50 0 50 100 Time (msec) Time (msec) Time (msec) Time (msec)

dsp 17a

B

6 Hours

1 0.5 0

Effective Spread of Muscimol dsp 20a dsp 21b 300

400

50 100 Time (msec)

dsp 22a 300

dsp 23a 150

60 30 0 –50

50 100 Time (msec)

dsp 24a 100

Pre-infusion 0

0

0

0

0

0

0

0

7.5

40

60

200

300

200

200

60

Saline

Muscimol 50 ng

Pre-infusion

Muscimol 150 ng

0

0

0

0

0

0

0

0

7.5

20

40

200

100

1

1

1

0

0

0

0

0

0

0

0

40

15

80

80

200

60

20

20

0

0

0

0

0

0

0

0

1

1

1

1

1

1

1

1


Duration of Muscimol Inactivation Saline

Preinfusion

0 0 0 0 0 0 0 0 –10 0 20 40 –10 0 20 40 –10 0 20 40 –10 0 20 40 –10 0 20 40 –10 0 20 40 –10 0 20 40 –10 0 20 40 Time (ms) Time (ms) Time (ms) Time (ms) Time (ms) Time (ms) Time (ms) Time (ms)


113

FIGURE 6.4

114


A partially bent wire stylet is left inserted in the cannula when not in use to prevent clogging. The bend in the stylet holds the stylet in place. During recording sessions, the stylet is taken out and an inner injector cannula (33 ga, 0.008 in. O.D.) is lowered through the guide cannula so that the tip of the injector extends up to about 0.5 mm beyond the base of the cannula. The depth of the tip of the injector cannula will be approximately 1.2 mm below the pial surface. Muscimol (or saline vehicle) is then infused at a rate of 0.3 µL/min. The dose of muscimol will depend upon the amount of cortex to be inactivated, and typically range between 50 and 500 ng in 50–500 nL of saline vehicle. Infusions are carried out with a precision syringe pump (Sage 361) driving a 1 µL Hamilton syringe. By monitoring the movement (with a calibrated gauge) of a small bubble in the polyethylene tubing connecting the Hamilton syringe to the injector cannula along with the precise displacement of the syringe, infusion volumes as low as 15 nL can routinely be achieved. Approximately 3–5 min after the infusion, the injector cannula is removed and the internal stylet replaced. Previous experiments in our laboratory have demonstrated that this technique can reliably inactivate localized regions of S1 cortex for periods up to 6–8 h in a completely reversible manner (Figure 6.4a and [Yuan et al., 1986]), and that concurrent multielectrode recordings provide a real-time measure of the spread of the drug (Figure 6.4b). At the end of all recording sessions, infusions of fluorescent dextrans (10,000 MW) into S1 cortex can be used to visualize the central area of infusion. One application of muscimol inactivation was to evaluate the contribution of corticothalamic feedback projections to the definition of tactile responses in populations of VPM neurons. Before cortical inactivation, Ghazanfar et al. (2001) recorded control data by stimulating a large number of single whiskers. During the period of cortical inactivation, the same single- and multiple-whisker stimuli were delivered again. Upon reversal of the cortical inactivation, a subset of these stimuli were delivered to verify whether thalamic and cortical responses had returned to control values. The reversible nature of the cortical inactivation was important for their experimental paradigm because multiple experiments could be carried out in the same animals. Waveform analysis, autocorrelation, and ISI histograms were used to confirm that the same single neurons were being recorded throughout the experiments. The results showed that supralinear summation of VPM neural responses to multiwhisker stimuli depends on feedback from S1. Similarly, Krupa et al. (1999) inactivated S1 cortex to show that peripheral deafferentation-induced immediate plasticity of tactile responses in VPM is substantially affected by input from S1 neurons. Muscimol inactivation has also been applied to determine the path of sensory whisker signals as they propagate centrally into the brain. By inactivating individual hemispheres of S1, Shuler et al. (2001) showed that S1 single-neuron responses to ipsilateral whisker deflection are due to projections across the corpus callosum from the contralateral S1 hemisphere, where neurons receive the primary afferent sensory signal from the whisker pad. These examples demonstrate the utility of muscimol inactivation for dissecting the electrophysiological functioning of a complex brain circuit. Another powerful class of applications dissects the role of specific neural populations in producing different levels of behavioral performance during tactile discrimination. For example,



115

Shuler et al. (2002) inactivated individual S1 hemispheres in a task in which certain stimuli required integration of bilateral whisker signal for successful discrimination. Performance deficits specific to those particular trials during unilateral S1 inactivation suggested that perceptual bilateral integration depended on callosal cross-talk between the S1 hemispheres.

NEURAL ENSEMBLE DATA ANALYSIS SINGLE-NEURON RESPONSES Having collected neural ensemble data using a particular behavioral and stimulation paradigm such as those described earlier, it remains to analyze and interpret the data. Ultimately, we would like to understand the behavior of neural ensembles in terms of the properties of individual neurons. Therefore, computing single-neuron firing properties is a reasonable starting point for most analyses. A basic question is, which neurons respond to a stimulus with a statistically significant modulation of their firing rate? Figure 6.5 reproduced from the study (Krupa et al., 2004) of S1 neural response properties during tactile discrimination, illustrates how a variety of properties can be derived from single-neuron response data. In this figure Krupa et al. have tabulated Depth 0

Early-on

LVQ

Duration Magnitude MPI*

Excitatory Multi

II/III 0.7 IV 1.0 V/VI 1.8 30 70

0 (%)

(%)

85 150 350 3 10 10 60 20 5020 50 (ms) (spiko) (%) (%) (%) *MPI: Multi-phasic Inhibitory

FIGURE 6.5 Summary diagram showing salient response properties of single neurons in different layers of S1 (Krupa, D.J., Wiest, M.C., Shuler, M.G., Laubach, M., and Nicolelis, M.A. (2004) Layer-specific somatosensory cortical activation during active tactile discrimination. Science 304: 1989–1992. With permission). The vertical axis measures recording depth in millimeters. Early-on: Percentage of units showing early-onset responses that begin before whisker stimulation; LVQ: Mean (+SEM) wide versus narrow categorization performance of an artificial neural network using population neural activities recorded at the different depths. Duration: Mean (+SEM) excitatory response duration. Magnitude: Mean (+SEM) magnitude of excitatory responses. MPI: Distribution of units with multiphasic responses that began with an inhibitory phase. Excitatory: Percentage of units with excitatory responses. Multi: Percentage of multiphasic responses. The difference between responses during the active discrimination and responses to passive whisker stimulation (not shown) could not be attributed to differences in the stimuli, suggesting that top-down inputs to S1 sculpt the active discrimination responses.


116


the distribution across cortical layers of several single-neuron properties. These include the duration and magnitude of firing modulations in the different layers, but also the incidence of “early-onset” modulations that occur before whisker contact with the stimulus, and the incidence of different temporal patterns of excitation and inhibition. (Only the column labeled “LVQ” in Figure 6.5 reflects a multineuron measure; see following text). The results were calculated using the commercial analysis package NeuroExplorer. Excitatory response durations were defined as the time interval in which firing rate significantly exceeded (p < 0.01) background firing rate, based simply on the standard deviation of counts in individual bins. The results from many neurons were tabulated and combined by hand. Another method based on the statistical distribution of cumulative-summed spike counts (e.g., Ghazanfar, 2001; Ushiba et al., 2002) has also been used in the Nicolelis lab to find the poststimulus latency at which the actual response deviates more than 99% of the null distribution of cumulative spike counts. The null distribution is based on a prestimulus window of activity. This approach takes into account the statistics of all the spiking activity up to a given moment, as opposed to considering each bin independently. It provides precise response latencies without smoothing or otherwise distorting the peri-event spike histograms, which would introduce complexities into the determination of confidence intervals. Moreover, the cumulative sum method requires no assumption of a particular parametric form of the spiking statistics. We have recently modified and automated (together with Ranier Guttierez) this method by writing a Matlab routine to calculate temporally precise response offsets, and response magnitudes as well as onsets. To assess the significance of deviations from the expected cumulative sum, an empirical distribution of the cumulative-summed spike count at each time bin before the stimulus is constructed from 1000 bootstrapped samples (with replacement) of the prestimulus spike histogram. This empirical distribution of prestimulus firing is then used to find the poststimulus bin (if any) at which the cumulative-summed poststimulus spike count exceeded or was less than 99% of the cumulative spike counts from the baseline distribution (Martinez and Martinez, 2002). This bin was recorded as the onset of an excitatory or inhibitory response. The response offset is identified as the first zero crossing of the derivative of the cumulative deviation from the baseline spike count, meaning that at that time point the cumulative sum is no longer deviating from its expected growth with time. This procedure gives the time at which the poststimulus time histogram has returned to a baseline firing rate. Given onset and offset, of course, response duration is also implicitly defined. The response magnitude was calculated as the number of excess spikes between response onset and offset, compared to the baseline expected number, divided by the number of trials. Thus, the response magnitude measures the average number of spikes per trial fired in response to a particular stimulus. Wiest et al. (2005) used this method to automate the calculation of response latencies and magnitudes in an electrophysiological study of bilateral integration of whisker signals in head-immobilized waking rats. This facilitated the calculation of bilateral integration measures for many units recorded over multiple sessions, resulting in a characterization of the distribution of different levels of sub- and supralinear summation of bilateral inputs by different infragranular S1 neurons.



117

LVQ-BASED ENSEMBLE ANALYSES We use the learning vector quantization (LVQ) (Kohonen, 1997) pattern recognition algorithm to predict the identity of a stimulus based on neural activity recorded during a single trial. We use the percentage of correct classifications, over all trials, as a measure of performance on the stimulus estimation task (Krupa et al., 2004). The percentage correct reflects the amount of overlap, in the space of neuronal responses, of the responses to different stimuli (Thomson and Kristan, 2005). For instance, when there are two stimuli such as narrow and wide apertures, if LVQ gives 50% correct, then that means there is no difference in the neural response to wide or narrow apertures. If the algorithm yields 100% correct, then there is no overlap between the response distributions to narrow and wide apertures. We have found that for neuronal ensemble data, LVQ outperforms methods such as linear discriminant analysis, and performs as well as support vector machines (Mark Laubach, unpublished observations) or neural networks trained via backpropagation (Nicolelis et al., 1998). We implement LVQ as an artificial neural network consisting of two layers of artificial neurons. We implement it in Matlab (Mathworks) with the Matlab Neural Network Toolbox (Demuth and Beale, 1998). The input to the network is the neural response to be classified. The first layer in the network is called the competitive layer, which is a winner-take-all network. The activation of a unit in the competitive layer is the distance between the input to the network and the weights onto that unit. The output is 1 for the unit whose weights are closest to the input, and zero for all the other units. Hence, the unit whose input weights are closest to the actual input return a 1 as output. LVQ networks are trained via a supervised learning algorithm, LVQ1 (Kohonen, 1997). Before training, each competitive neuron is assigned to a class in the training data (e.g., trials with narrow- or wide-aperture stimuli). Weights are modified only for the unit that wins the competition in the competitive layer (i.e., the unit whose output is 1). If the winning unit corresponds to the correct class for the training trial, then its input weights are nudged closer to the input on that iteration. If, on the other hand, the winning unit corresponds to the incorrect class for that trial, the input weights are pushed a small distance away from the inputs for that trial, so that similar inputs will be less likely to activate that node in the future. The weights remain unchanged for the competitive units that output 0. Mark Laubach modified the built-in Matlab Toolbox code so as to implement the optimized LVQ (OLVQ1) learning rule (Kohonen, 1997). This learning rule reduces the learning rate during each iteration of the algorithm so that weight changes are smaller for each iteration. The second, output, layer of the network typically consists of the same number of units as there are classes of inputs to be estimated. For instance, with the narrow-wide discrimination there are two output units. The output layer is simply a “reporter” of which class won the competition at the competitive layer. Quantitatively, all the competitive units preassigned to the same class have a weight of 1 onto the corresponding class output unit, and 0 to the other output units. Hence, if any “narrow-class” competitive unit wins the competition, the corresponding output unit will return a 1, and the other output units will return a 0.


118


When training is complete the test trial data are input to the network, and the network’s output classifies the test trial on the basis of regularities in the training data set. We use all but one trial as the training set, and testing the network on a single “hold-out” trial. By taking each trial in turn as the “hold-out” and training the network on the remaining set, one tests the network using every trial (instead of only on some arbitrarily chosen training subset). This procedure is known as leave-oneout cross validation. The results of an analysis are quantified in terms of percentage of single trials classified correctly. For example, we applied LVQ to study information about actively sampled stimuli coded in ensemble activity recorded from different cortical layers of S1 as rats discriminated a wide from a narrow aperture using their facial vibrissae. Single-trial peri-event histograms (50 ms bins) were constructed for the epoch starting 200 ms before until 500 ms after the time when the rats sampled the tactile stimuli. Leaveone-out cross-validation was used as outlined earlier. The training data were used at each iteration to initialize the LVQ networks by setting the coefficients for each competitive neuron dedicated to a given type of trial equal to the mean neuronal ensemble response for that type of trial (plus a small random noise term). In this case we used four competitive units: two for each class to be represented, i.e., two for “wide” and two for “narrow.” Performance of the network for each analysis was compared to a chance level of 50% correct. This analysis allowed us to compare the ability of ensembles at different cortical layers to discriminate between the two aperture widths. We found a significant trend toward better LVQ performance through the deeper cortical layers (r(20) = 0.60, p < 0.01). By manipulating the single-trial data to selectively omit excitatory or inhibitory modulations of neural firing rate, we were also able to compare stimulus discrimination for these different response types. Because inhibitory modulations are a rare form of response to passive stimulation (Sachdev et al., 2000; Krupa et al., 2004), it was interesting to find that widespread inhibitory modulations during discrimination behavior carried significant information about the actively sampled stimulus. In contrast, previous studies that examined potential coding mechanisms in S1 following passive whisker stimulation suggest that tactile information is encoded mainly by excitatory S1 activity (Ghazanfar et al., 2000; Petersen and Diamond, 2000; Arabzadeh et al., 2003). We also used a “moving window” analysis to assess the time course of information about the wide and narrow apertures. LVQ was applied sequentially to a 500 ms epoch of neuronal ensemble activity that “moved” in 100 ms steps through the 1 s epochs before and after the time of whisker contact. An example of this analysis is shown in Figure 6.6. This analysis produced a continuous quantitative readout of the recorded population’s ability to distinguish between the wide and narrow apertures. The time course of LVQ performance around the time of whisker contact can be used to verify that the animal’s brain does not have access to task-related cues before sampling the discriminandum, and to show how task-related information rises and falls after the stimulus. In interpreting the LVQ performance time course, it is important to note that stimulus information reflected in the LVQ performance need not be directly related to the stimulus. For example, in the aperture discrimination task in which the animal is required to pass its face through the discriminandum to trigger a central nose poke beyond the aperture, there is a period after contacting the aperture



119

90

Correct Predictions (%)

80

70

60

50

–1.0

–0.8 –0.6 –0.4 –0.2 0.0 0.2 0.4 0.6 Time (before & after breaking photobeam; seconds)

0.8

FIGURE 6.6 Example of LVQ output (from one recording session), showing the ability to discriminate between the wide and narrow aperture based on the recorded population activity before and after whiskers contacted the aperture (Krupa, D.J., Wiest, M.C., Shuler, M.G., Laubach, M., and Nicolelis, M.A. (2004) Layer-specific somatosensory cortical activation during active tactile discrimination. Science 304: 1989–1992. With permission). Prior to the whiskers contacting the aperture (time 0.0 s), the ability of the LVQ to discriminate between the wide and narrow apertures was near the level of chance. As the whiskers contact the aperture (0.1 s), the performance of the LVQ classifier improved above the level of chance. By 0.2 s after whisker contact, the LVQ classifier performed similar to (even better than) the rat at discriminating between different aperture widths.

until triggering the center nose poke during which the animal’s motor behavior is statistically identical in wide and narrow trials. During that period, stimulus-related information in S1 is most likely a direct sensory image of the stimulus. After that period (~300–500 ms), the animal begins to move to the left or right—and as soon as it does, the stimulus-related information quantified in S1 using the LVQ may be “contaminated” with movement-related activity. A study by Ghazanfar et al. (2000) illustrates other data manipulations that can be used to investigate the role of specific coding mechanisms in a given neural ensemble data set. The experimenters stimulated individual whiskers in anesthetized rats while recording ensemble activity in layer V S1 and VPM. They applied LVQ to quantify ensemble performance at discriminating between four different whiskers (B1, B4, E1, and E4; chance performance = 25%). To ask whether stimulus location information was local or distributed, they applied a “neuron-dropping” analysis. The procedure is to sequentially remove the “best predictor” neuron from the data set and recalculate LVQ performance to calculate a performance curve as a function of the number of remaining neurons in the ensemble. The best predictor neuron is determined by running the analysis with each neuron left out in turn. The neuron whose


120


omission had the most detrimental effect on performance is identified as the best predictor. A smooth degradation in performance as neurons were dropped suggested the coding of the location of single-whisker stimuli in layer V is distributed rather than local to a single barrel. To determine whether the temporal modulation of neural firing rates contributed to the ensemble information about stimulus location, Ghazanfar et al. varied the bin size of the single-trial poststimulus time histograms that define the input data for the LVQ analysis. The bin size was varied from 1 to 40 ms. Because increasing the bin size degrades the temporal resolution of the recorded response, a decrease in LVQ discrimination performance upon increasing the bin size shows that temporal patterning of the ensemble response contributed to the ensemble information about stimulus location. Finally, one can manipulate the data to perturb relations among different neurons, to test whether coordination across neurons contributes to the ensemble stimulus information. (Ghazanfar et al. actually used linear discriminant analysis (LDA) rather than LVQ for neural activity pattern classification in this part of their analysis.) At least two methods can be distinguished for this disruption of across-neuron coordination. Randomizing the trial number separately for each neuron (known as “trialshuffling”) preserves single-trial latency information, but disrupts any across-neuron coordination that varies from trial to trial. To the extent that neural responses are consistently stimulus-locked, on the other hand, their covariance can be preserved after trial shuffling. Alternatively, one can introduce random “jitter” into the entire spiketrain of each neuron individually, disrupting both single-neuron latency information and across-neuron correlations, without affecting single-neuron temporal patterns. Latency jitter reduced the ensemble stimulus information, but trial-shuffling did not, supporting the importance of temporal coding in terms of relative latencies, but not across-neuron coordination, in the context of single-whisker stimuli. Another study, on location coding in primates (Nicolelis et al., 1998), used the same LVQ approach to examine different coding strategies used by ensembles in different areas to carry information about the location of a tactile stimulus. Again, it was found that S1 ensembles code single locations primarily by a distributed stimulus-locked latency code. In contrast, however, trial-shuffling showed that S2 ensembles carry location information in the temporal pattern of activity distributed across many neurons that varies from trial to trial.

CONCLUSION AND OUTLOOK Multielectrode, behavioral, pharmacological, and quantitative methods have evolved to the point where we can fruitfully address questions about somatosensory coding by neuronal ensembles in the awake behaving rat. Applying these methods has already shown that principles of somatosensory coding depend strongly on the behavioral state and context the rat finds itself in at each moment. Technical improvements in three directions will facilitate the discovery of the neuronal underpinnings of somatosensory behavior. First, more fine-grained circuit analysis techniques will reveal specific cell-types, their precise locations, geometry, and channel distributions; leading to refinements of biologically realistic models



121

of the specific transformations being carried out by the somatosensory system. Markram (2006) discusses a research program along these lines. Second, techniques to evoke neuronal activity in precise spatiotemporal patterns will shed light on the effects of such perturbations on downstream networks and behavior. Boyden et al. (2005) discusses recent advances in such techniques. Third, as the focus of somatosensory electrophysiology shifts to questions of somatosensory function in awake behaving animals, behavioral methods will assume increasing importance. For example, advances in whisker sensory physiology will likely follow improvements in our ability to monitor whisker movements using optoelectronic whisker monitoring (Bermejo et al., 1998), video analysis (Knutsen et al., 2005), and physical modeling of whisker dynamics (Hartmann et al., 2003). With these advances, the whisker system will continue to provide fertile ground for understanding how rats use tactile sensation to pursue their constantly varying goals.

RESOURCES Scott Edwards Electronics, Inc. Mini SSC II Serial Servo Controller www.seetron.com (520) 459-4802 Parallax, Inc. Basic Stamp II www.parallax.com (888) 512-1024 Med Associates, Inc. DIG-726 Output Card DIG-712 Input Card Med PC IV Software www.med-associates.com (802) 527-9724 Electronic Model Systems Futaba S148 Servo and EMS Linear Servo Conversion Kit www.emsjomar.com (800) 845-8978 Parker Hannifin General Valve www.parker.com Model 003-0111-900 Solenoid Valve, 24V (973) 575-4844 MSC Industrial Supply www.mscdirect.com Pneumatic valves and cylinders (800) 645-7270 Digikey www.digikey.com Electronic components (800) 344-4539


122


REFERENCES Arabzadeh, E, Petersen, R.S., and Diamond, M.E. (2003) Encoding of whisker vibration by rat barrel cortex neurons: implications for texture discrimination. J Neurosci 23: 9146–9154. Armstrong-James, M. and Fox, K. (1987) Spatiotemporal convergence and divergence in the rat S1 “barrel” cortex. J Comp Neurol 263: 265–281. Bermejo, R., Houben, D., and Zeigler, H.P. (1998) Optoelectronic monitoring of individual whisker movements in rats. J Neurosci Methods 83: 89–96. Bermejo, R., Harvey, M., Gao, P., and Zeigler, H.P. (1996) Conditioned whisking in the rat. Somatosens Mot Res 13: 225–233. Boyden, E.S., Zhang, F., Bamberg, E., Nagel, G., and Deisseroth, K. (2005) Millisecondtimescale, genetically targeted optical control of neural activity. Nat Neurosci 8(9): 1263–1268. Castro-Alamancos, M. (2004) Absence of rapid sensory adaptation in neocortex during information processing states. Neuron 41: 455–464. Demuth, H. and Beale, M. (1998) Neural network toolbox for use with MATLAB. Natick, MA: The Mathworks. Diamond, M.E., Armstrong-James, M., and Ebner, F.F. (1992) Somatic sensory responses in the rostral sector of the posterior group (POm) and in the ventral posterior medial nucleus (VPM) of the rat thalamus. J Comp Neurol 318: 462–476. Erchova, I.A. and Diamond, M.E. (2004) Rapid fluctuations in rat barrel cortex plasticity. J Neurosci 24: 5931–5941. Fanselow, E.E. and Nicolelis, M.A. (1999) Behavioral modulation of tactile responses in the rat somatosensory system. J Neurosci 19: 7603–7616. Fanselow, E.E., Reid, A.P., and Nicolelis, M.A. (2000) Reduction of pentylenetetrazole-induced seizure activity in awake rats by seizure-triggered trigeminal nerve stimulation. J Neurosci 20: 8160–8168. Fontanini, A. and Katz, D.B. (2006) State-dependent modulation of time-varying gustatory responses. J Neurophysiol 96: 3183–3193. Ghazanfar, A. and Nicolelis, M.A.L. (1997) Nonlinear processing of tactile information in the thalamocortical loop. J Neurophysiol 78: 506–510. Ghazanfar, A., Krupa, D.J., and Nicolelis, M.A.L. (2001) Role of cortical feedback in the receptive field structure and nonlinear response properties of somatosensory thalamic neurons. Exp Brain Res 141: 88–100. Ghazanfar, A.A. and Nicolelis, M.A. (1999) Spatiotemporal properties of layer V neurons of the rat primary somatosensory cortex. Cereb Cortex 9: 348–361. Ghazanfar, A.A. and Nicolelis, M.A. (2001) Feature article: the structure and function of dynamic cortical and thalamic receptive fields. Cereb Cortex 11: 183–193. Ghazanfar, A.A., Stambaugh, C.R., and Nicolelis, M.A. (2000) Encoding of tactile stimulus location by somatosensory thalamocortical ensembles. J Neurosci 20: 3761–3775. Ghazanfar, A.A., Krupa, D.J., and Nicolelis, M.A. (2001) Role of cortical feedback in the receptive field structure and nonlinear response properties of somatosensory thalamic neurons. Exp Brain Res 141: 88–100. Hartmann, M.J., Johnson, N.J., Towal, R.B., and Assad, C. (2003) Mechanical characteristics of rat vibrissae: resonant frequencies and damping in isolated whiskers and in the awake behaving animal. J Neurosci 23: 6510–6519. Harvey, M.A., Bermejo, R., and Zeigler, H.P. (2001) Discriminative whisking in the headfixed rat: optoelectronic monitoring during tactile detection and discrimination tasks. Somatosens Mot Res 18: 211–222. Knutsen, P.M., Derdikman, D., and Ahissar, E. (2005) Tracking whisker and head movements in unrestrained behaving rodents. J Neurophysiol 93: 2294–2301.



123

Kohonen, T. (1997) Self-organizing maps. New York: Springer. Krupa, D.J, Brisben, A.J., and Nicolelis, M.A.L. (2001a) A multi-channel whisker stimulator for producing spatiotemporally complex tactile stimuli. J Neurosci Methods 104: 199–208. Krupa, D.J., Matell, M.S., Brisben, A.J., Oliveira, L.M., and Nicolelis, M.A.L. (2001b) Behavioral properties of the trigeminal somatosensory system in rats performing whisker-dependent tactile discriminations. J Neurosci 21: 5752–5763. Krupa, D.J., Ghazanfar, A.A., and Nicolelis, M.A. (1999) Immediate thalamic sensory plasticity depends on corticothalamic feedback. Proc Natl Acad Sci U.S. A. 96: 8200–8205. Krupa, D.J., Wiest, M.C., Shuler, M.G., Laubach, M., and Nicolelis, M.A. (2004) Layer-specific somatosensory cortical activation during active tactile discrimination. Science 304: 1989–1992. Marescaux, C., Vergnes, M., and Depaulis, A. (1992) Genetic absence epilepsy in rats from Strasbourg--a review. J Neural Transm 35: 37–69. Markram, H. (2006) The blue brain project. Nat Rev Neurosci 7: 153–160. Martinez, W. and Martinez, A.R. (2002) Computational statistics handbook with Matlab. Boca Raton: Chapman & Hall/CRC. Moore, C. (2004) Frequency-dependent processing in the vibrissa sensory system. J Neurophysiol 91: 2390–2399. Moore, C., Nelson, S.B., and Sur, M. (1999) Dynamics of neuronal processing in rat somatosensory cortex. Trends Neurosci 22: 513–520. Nelson, R.J. (1996) Interactions between motor commands and somatic perception in sensorimotor cortex. Curr Opin Neurobiol 6: 801–810. Nicolelis, M., Dimitrov, D., Carmena, J.M., Crist, R., Lehew, G., Kralik, J.D., and Wise, S.P. (2003) Chronic, multisite, multielectrode recordings in macaque monkeys. Proc Natl Acad Sci U. S. A. 100: 11041–11046. Nicolelis, M. and Fanselow, E.E. (2002) Thalamocortical optimization of tactile processing according to behavioral state. Nat Neurosci 5: 517–523. Nicolelis, M.A. and Chapin, J.K. (1994) Spatiotemporal structure of somatosensory responses of many-neuron ensembles in the rat ventral posterior medial nucleus of the thalamus. J Neurosci 14: 3511–3532. Nicolelis, M.A., Baccala, L.A., Lin, R.C., and Chapin, J.K. (1995) Sensorimotor encoding by synchronous neural ensemble activity at multiple levels of the somatosensory system. Science 268: 1353–1358. Nicolelis, M.A., Ghazanfar, A.A., Stambaugh, C.R., Oliveira, L.M., Laubach, M., Chapin, J.K., Nelson, R.J., and Kaas, J.H. (1998) Simultaneous encoding of tactile information by three primate cortical areas. Nat Neurosci 1: 621–630. Pereira, A., Wiest, M., Thomson, E., De Araujo, I., and Nicolelis, M. (2005) Neural ensemble correlates of texture discrimination in the behaving rat’s somatosensory system. Soc Neurosci Abstr 538.13. Petersen, R.S. and Diamond, M.E. (2000) Spatial-temporal distribution of whisker-evoked activity in rat somatosensory cortex and the coding of stimulus location. J Neurosci 20: 6135–6143. Sachdev, R.N., Sellien, H., and Ebner, F.F. (2000) Direct inhibition evoked by whisker stimulation in somatic sensory (SI) barrel field cortex of the awake rat. J Neurophysiol 84: 1497–1504. Schubert, D., Staiger, J.F., Cho, N., Kotter, R., Zilles, K., and Luhmann, H.J. (2001) Layerspecific intracolumnar and transcolumnar functional connectivity of layer V pyramidal cells in rat barrel cortex. J Neurosci 21: 3580–3592. Shaw, F.Z. (2007) 7–12 Hz high-voltage rhythmic spike discharges in rats evaluated by antiepileptic drugs and flicker stimulation. J Neurophysiol 97: 238–247.


124


Shaw, F.Z., Lee, S.Y., and Chiu, T.H. (2006) Modulation of somatosensory evoked potentials during wake-sleep states and spike-wave discharges in the rat. Sleep 29: 285–293. Shimegi, S., Takafumi, A., Takehiko, I., and Sato, H. (2000) Physiological and anatomical organization of multiwhisker response interactions in the barrel cortex of rats. J Neurosci 20: 6241–6248. Shuler, M., Krupa, D.J., and Nicolelis, M.A.L. (2001) Bilateral integration of whisker information in the primary somatosensory cortex of rats. J Neurosci 21: 5251–5261. Shuler, M.G., Krupa, D.J., and Nicolelis, M.A. (2002) Integration of bilateral whisker stimuli in rats: role of the whisker barrel cortices. Cereb Cortex 12: 86–97. Thomson, E.E. and Kristan, W.B. (2005) Quantifying stimulus discriminability: A comparison of information theory and ideal observer analysis. Neural Computation 17: 741–778. Thomson, E.E., Wiest, M.C., Pereira, A., and Nicolelis, M. (2005) A behavioral paradigm for the study of category discrimination in the rat whisker system. Soc Neurosci Abstr Program #883.6. Ushiba, J., Tomita, Y., Masakado, Y., and Komune, Y. (2002) A cumulative sum test for a peri-stimulus time histogram using the Monte Carlo method. J Neurosci Methods 118: 207–214. Wiest, M.C. and Nicolelis, M.A. (2003) Behavioral detection of tactile stimuli during 7–12 Hz cortical oscillations in awake rats. Nat Neurosci 6: 913–914. Wiest, M.C., Bentley, N., and Nicolelis, M.A. (2005) Heterogeneous integration of bilateral whisker signals by neurons in primary somatosensory cortex of awake rats. J Neurophysiol 93: 2966–2973. Yuan, B., Morrow, T.J., and Casey, K.L. (1986) Corticofugal influences of S1 cortex on ventrobasal thalamic neurons in the awake rat. J Neurosci 6: 3611–3617. Zhang, M. and Alloway, K.D. (2005) Intercolumnar synchronization of neuronal activity in rat barrel cortex during patterned airjet stimulation: a laminar analysis. Exp Brain Res: 1–15. Zhu, J.J. and Connors, B.W. (1999) Intrinsic firing patterns and whisker-evoked synaptic responses of neurons in the rat barrel cortex. J Neurophysiol 81: 1171–1183.


7

Chronic Recording During Learning Aaron J. Sandler

CONTENTS Introduction............................................................................................................ 125 The Importance of Neuronal Ensembles ............................................................... 127 Using Nonhuman Primates to Study Learning Using Chronic Recording Methods....................................................................................................... 127 Analysis of Single Unit Response Properties ........................................................ 130 Quantitative Analysis of Peri-Event Time Histograms .............................. 130 Distance Index (DI) Analysis ..................................................................... 133 Analysis of Neural Ensemble Data ........................................................................ 135 Using Artificial Neural Networks (ANNs) for Single Trial Classification Based on Neural Ensemble Recordings .................... 136 LVQ-Based Population Analyses................................................................ 136 Quantification of Neuronal Responses with LVQ ...................................... 137 Population Distance Index (pDI) Analysis for Single Trial Classification Based on Neural Ensemble Recordings .................... 137 Moving-Window Analysis .......................................................................... 139 Cross Correlations to Study Neuronal Interactions .................................... 139 Partial Directed Coherence (PDC) Analysis .............................................. 139 Principle Component Analysis (PCA)........................................................ 140 Control Analyses ........................................................................................ 140 Conclusions ............................................................................................................ 141 References.............................................................................................................. 141

INTRODUCTION The study of learning has a rich tradition, going back at least to the days of Aristotle, who proposed that the formation of associations between coincident events is the way humans learn. More famously associated with learning, especially in the popular mind, is Pavlov who, in the 1920s, studied what is now known as classical conditioning or Pavlovian conditioning. In these well-known experiments, an unconditioned stimulus (food), which naturally causes an unconditioned reflex (salivation), was presented along with a neutral stimulus (a bell) with enough repetition that, eventually, the bell began to evoke the salivation even without the presence of food (conditioned 125 © 2008 by Taylor & Francis Group, LLC

126


response). At that point the bell had become a conditioned stimulus, i.e., a stimulus that, after learning a new association, evokes the conditioned response. Classical conditioning refers to an environmental stimulus that can elicit a response, but it does not address the ways in which an animal might learn how its own behavior could cause an environmental response. Work by Thorndike in the late 1890s began to address this by studying what he termed instrumental conditioning. Thorndike observed that a hungry cat could learn through trial and error that rubbing up against the side of its cage would open a latch allowing it access to food. From this he proposed his law of effect, in which he argued that the tendency to repeat a behavior is dependent on the consequences that behavior evokes. In a series of studies from the 1930s through the 1950s, Skinner followed up on Thorndike’s work, studying primarily pigeons and rats in a wide variety of conditions, including positive reinforcement (in which a behavior is more likely to recur if it is followed by a reward such as a piece of food), negative reinforcement (a behavior becoming more likely if it is followed by the removal of an aversive stimulus such as an electric shock). Skinner renamed this paradigm operant conditioning because a spontaneously emitted behavior (or operant) is what elicits the response. Although the rich psychological history of studying classical conditioning, operant conditioning, and other forms of learning has led to greater understanding of the phenomenology of learning, less is known about the neurophysiological mechanisms of these forms of learning. This is because most studies of cortical and subcortical function have involved either functional imaging, in which case the spatial resolution is too gross to permit the study of precise mechanisms of neuronal learning, or acute single-electrode recording methods, which are inherently limited in their ability to examine functional interrelationships between various cortical areas or to study any changes in neuronal firing that might occur over days or weeks. This inability to track changes across cortical areas has meant that most studies have used highly trained animals who have reached a stable level of performance on a previously learned task, rather than animals learning a new task. Multielectrode ensemble recording, however, provides the ability to record from a large number of cells simultaneously. The mammalian brain contains many millions of neurons that work in an interconnected manner to produce complex behaviors and thoughts. Understanding the interrelations between many neurons that make up a functional circuit, therefore, requires simultaneous recording from many more than one at a time. Furthermore, because the brain’s encoding of a given event (be it sensory, motor, or cognitive) relies on complex interactions between neurons, our ability to understand fundamental neural-circuit mechanisms is greatly improved when one simultaneously records the firing of many neurons, rather than just a single one at a time. Thus, multielectrode recording brings us a step closer to understanding normal brain function. Another important advantage of multielectrode recording is that it provides a more random sample of the neurons in the implanted area, obviating a priori decisions about the cell types of interest and permitting comparison of the contributions of different neurons to the encoding of, for example, a given motor action. Finally, and most important for the study of learning, chronic implantation of multielectrode arrays allows us to study ongoing processes that take more than a single


Chronic Recording During Learning

127

session to complete. Thus, multielectrode arrays may be implanted in a naïve animal and recordings made throughout the course of learning. This allows the sampling of many neurons each day of the study, even if it lasts weeks, months, or years.

THE IMPORTANCE OF NEURONAL ENSEMBLES Processing and storing information has long been suspected to require large populations of neurons with dynamic and distributed interactions (Sherrington 1906; Hebb 1949; Lashley 1950; Erickson 1968; Freeman 1975; Fuster 1995; Nicolelis, Ghazanfar et al. 1997). More recently, experimental evidence has accumulated to support this idea (Freeman 1975; Georgopoulos, Schwartz et al. 1986; Georgopoulos, Taira et al. 1993; Nicolelis and Chapin 1994; Wilson and McNaughton 1994; Fuster 1995; Nicolelis, Baccala et al. 1995; Nicolelis, Ghazanfar et al. 1997). When processing sensory stimuli and generating motor outputs, a wide variety of tasks must be accomplished: the sensory information must be processed, converted to percepts, and stored in memory; a response must be selected; and a behavior must be generated. This process is not limited to simple circuits or reflex loops, but rather requires large-scale interactions among widely distributed and interconnected populations of neurons (Nicolelis, Ghazanfar et al. 1997). Indeed, current evidence suggests that decision making based on sensory cues does involve multiple subcortical and cortical structures (Wise and Murray 2000). This evidence is most plentiful in neurophysiological studies of the oculomotor system in which visual cues instruct eye movement responses (saccades). Areas shown to be involved in oculomotor decision making in nonhuman primates include the medial temporal cortex (Salzman, Britten et al. 1990), prefrontal cortex (Kim and Shadlen 1999), areas of parietal cortex (Quintana and Fuster 1992; Platt and Glimcher 1999; Shadlen and Newsome 2001; Pesaran, Pezaris et al. 2002; McCoy, Crowley et al. 2003; McCoy and Platt 2005), the frontal eye fields (Gold and Shadlen 2003), the supplementary eye fields (Chen and Wise 1995), premotor cortex (Mitz, Godschalk et al. 1991; Boussaoud and Wise 1993; Brasted and Wise 2004), and the superior colliculus (Horwitz and Newsome 2001). Wise et al. found activity in many areas throughout the visuomotor pathway that both reflected sensory information and predicted the monkey’s choice (Wise and Murray 2000). In humans, functional magnetic resonance imaging (fMRI) studies of visuomotor tasks have found evidence of task-related activity across several cortical and subcortical areas, including frontal and parietal cortices (Thoenissen, Zilles et al. 2002).

USING NONHUMAN PRIMATES TO STUDY LEARNING USING CHRONIC RECORDING METHODS Since the mid-1990s, our laboratory has used microwire arrays to record the activity of large ensembles of single units in multiple cortical areas in behaving nonhuman primates. Nonhuman primates are attractive subjects for multisite, multineuronal recordings during learning for two main reasons: First, primates are capable of performing more sophisticated tasks than are rodents, providing expanded opportunities in conditions for study. Secondly, brain structures in primates tend to be more similar


128


to those in humans, making the findings more directly applicable to human learning. Both of these factors come into play when studying such questions as which neurons and cortical areas are recruited to solve a novel task, such as the operation of a brain–machine interface, a central goal of our lab. The two types of primates used in our studies are rhesus macaques (Macaca mulatto) and owl monkeys (Aotus trivirgatus). Regardless of the species chosen, care of nonhuman primates includes appropriate housing in a research animal facility under the direction of full time veterinary staff. The facility must be equipped to take care of any disease problems specific to the animals and general to the primate population. In learning experiments that continue over days or weeks it is also crucial to standardize day-to-day care as much as possible, to avoid introducing uncontrolled variables that may interfere with the animals’ learning. Macaques are the laboratory standard for neurophysiological studies of this type, making them an attractive choice for simple comparison to other studies. In addition, of the laboratory animals used for single unit recordings, the organization of their brain is the closest to that of humans. They are also able to learn quite complicated tasks. Owl monkeys are New World monkeys, smaller in size than macaques. They have a lissencephalic brain with a very smooth neocortex. The lack of circumvolutions and sulci allows straightforward access to distinct cortical areas that may be difficult to approach in macaque brain. Additionally, the brains of owl monkeys have been well studied in neuroanatomical and neurophysiological studies, making them an appropriate choice for chronic multielectrode implantation in multiple areas (Kaas 1994). They are also easier to maintain, and cheaper to work with, than other common models, owing to their less aggressive nature and lower level of health risks to humans. One minor complication of working with owl monkeys is due to their being nocturnal animals; this is easily solved by adjusting their light cycle to ensure that all experiments are performed during their awake time. Perhaps most importantly in the consideration of owl monkeys as experimental subjects, it is possible to obtain excellent neuronal recordings from this species. We have shown that recordings of hundreds of single units can remain viable in owl monkeys for several years, something that has not yet been achieved in any other primate species. In recent years, we have carried out very long-term studies of owl monkeys with multiple implants in primary somatosensory cortex (S1), posterior parietal cortex (PPC), dorsal premotor cortex (PMd), and primary motor cortex (M1). In one of these animals, recordings have been performed for over 5 years. A second animal has been recorded for more than 2 years. After several years, we have continued to record from 50–100 single units and over 100 multiunits per animal (Figure 7.1a-b). The quality of the recordings themselves is as important as the long duration of the implants; even after 2 and 4 years we are able to achieve high-quality recordings from these owl monkeys with good discrimination of single units (Figure 7.1c-d). Behaviorally, owl monkeys are amenable to learning tasks including somatosensory and visually cued decision-making tasks (Kaas 1994). In our lab, we have developed training methods for owl monkeys that have shown considerable success. In these tasks, our goal is to train monkeys to select reach targets based on a discrimination between several different cues. As they learn to do this, we carry out simultaneous neural ensemble recordings from multiple cortical areas. However, before the


Chronic Recording During Learning 4.0

Single units Multiunits Total units

200

Average Number of Units Per Active Wire

Total Number of Discriminable Units

250

129

150 100 50 0

3.0 2.0 1.0 0.0

0

(a)

12

24

36

48

Months After Implantation (monkey #1)

0

12 24 36 Months After Implantation (monkey #1)

48

0

12 Months After Implantation (monkey #2)

24

(c)

Average Number of Units Per Active Wire

Total Number of Discriminable Units

250 200 150 100 50

2.0

1.0 0.0

0 0 (b)

3.0

12 24 Months After Implantation (monkey #2)

(d)

FIGURE 7.1 High quality neuronal recordings were obtained 4 years after implantation in monkey #1 (a) and 2 years in monkey #2 (b). The total number of discriminable single- and multiunits and average number of single- and multiunits obtained per wire are shown for monkey #1 for 4 years after implantation (c) and monkey #2 for 2 years after implantation (d).

experimental learning can begin, the animals must undergo some initial training. We generally start this period of preimplantation training with habituation of the monkeys to the handling procedures and experimental apparatus. They are trained to sit comfortably in a specially designed primate training chair inside a soundproof recording chamber by the use of food and juice rewards over a period of 3 to 4 weeks (Kralik, Dimitrov et al. 2001). The first step in the procedure is training them to leave their home cages voluntarily, by luring them into transport boxes (Primate Products, Miami, Florida) with small pieces of fruit. Once they do this voluntarily and are comfortable enough to take food and drink in the transport box, the monkeys are then transported to the testing room, where a custom-designed restraining chair is attached to the end of the transport box, and the monkeys are again lured into the restraining chair with a small piece of fruit. The restraining chair we use was made by Crist Instrument Company out of ½ in. thick polycarbonate, and was then substantially modified in our laboratory to fit our purposes. It was designed to be initially in the form of a box that the monkeys were comfortable entering from the transport box. After moving it and the monkey into the sound-attenuating test chamber, the device can be transformed from the box form to the chair form. In this configuration, the chair allows the monkeys to sit comfortably, without being fully restrained while performing behavioral tasks. Their lower extremities are supported on a perch, their head protruding through a hole in the top, and their arms free to reach forward but prevented from reaching up to their implants and connectors. Generally, we restrain one of the monkey’s arms and leave the other one free to reach. The head is not restrained, although the front of the body harness is attached to the front of the chair to prevent swiveling of the entire body,


130


which might damage the implants. Once the monkeys are acclimated to the restraining chair, the specific experimental apparatus is introduced. For experiments studying associative learning or decision making, in which we want to avoid any simultaneous motor learning, we are careful to train the monkeys in the experimental apparatus (without using any cues) before the implantation surgery. This preexperiment training allows us to later study the formation of the association between stimulus and response without the confounding factor of the animal simultaneously having to learn a new motor output. In studies of motor learning, this pretraining is unnecessary. After a postoperative period of at least 14 days, the animals are reintroduced to the experimental apparatus. Cues are not introduced until neuronal recordings have stabilized (typically 2–3 months in owl monkeys). Monkeys generally perform 1 session per day, 5 days a week, until the predefined learning endpoint is reached.

ANALYSIS OF SINGLE UNIT RESPONSE PROPERTIES QUANTITATIVE ANALYSIS OF PERI-EVENT TIME HISTOGRAMS In analyzing any type of neuronal recording, the main goal is to determine the relationship between the firing of the neurons and some event such as a stimulus or a motor response. This can be as simple as looking for a change in firing rate that occurs simultaneously with the appearance of a stimulus. More complex neural events, however, such as the formation or enactment of a motor command, may involve variation in the patterns of neuronal firing. Rather than a simple increase or decrease in firing rate, for example, an individual neuron might show first an increase and then a decrease in firing rate (see Figure 7.2 for examples). It is therefore important to select FIGURE 7.2 (see pages 131 and 132) (See color insert following page 140.) Complex task-related neuronal responses. Eight different neuronal units that exhibit task-related modulation of neuronal firing other than a simple increase in firing rate are shown (A–H). A method capable of detecting such differences as these must be selected for the quantitative analysis. The Distance Index (DI), which preserves time-dependent information, method fits this criterion. For each unit, neuronal firing is represented both as raster plots where each trial is represented in a row with a tick mark for each action potential (top), and as peri-event time histograms where the average firing rates over trials are graphed and the thicknesses of the lines indicates the standard error (bottom). The x-axis show the time (in seconds) relative to the onset of the vibrotactile stimulus. Trials where the vibrotactile stimulus was applied contralateral to the unit are shown in red. Trials with ipsilateral stimuli are shown in blue. On the right are graphs of the units’ waveforms and interspike intervals. The contralateral and ipsilateral trials were randomly interspersed during the sessions. Here they have been sorted to allow visualization of differences between neuronal responses to the two cues. Some units responded to the onset of the task with a decrease in firing rate, as shown here in example units from M1 (A), which showed more inhibition for left reach (blue) than for right reach (red) trials, and PPC (B), which showed similar inhibition for both types of trials near the onset of the stimulus. Other tracings had even more complex patterns. The PPC unit shown in (C) exhibited an initial decrease in firing rate for both vibrotactile stimuli, followed by a rapid rise in firing rate for contralateral stimulus (red) and a slower rise for ipsilateral stimulus (blue). The PMd unit shown in (D) had an initial rise at the onset of both stimuli followed by additional peaks of firing rate approximately midway through the ipsilateral-cued (blue) but not contralateral-cued (red) trials. (Continued on p. 133.)


Reach Right

Reach Right

Vibrotactile Stimulus

A

15

5msec

10 5 0 –1

0

1

2 3 Time (sec)

4

5

6

B

10msec

0

1

2 3 Time (sec)

4

5

6

Stim Right

Stim Right

5msec

0

1

2 3 Time (sec)


5

6

D

40 35 30 25 20 15 10 5 0 –1

5msec

0

1

2 3 Time (sec)

4

5

6

131

FIGURE 7.2

4

PMd th051304-stim-sig014c

Stim Left

PPC th051304-stim-sig036d

Stim Left Firing Rate (spikes/sec)

C

30 25 20 15 10 5 0 –1



30 25 20 15 10 5 0 –1

PPC th031504-reach-sig036d

Reach Left Firing Rate (spikes/sec)

20

Firing Rate (spikes/sec)


Reach Left

M1 th0617 04-reach-sig023b




Reach Right Reach Left

15

5msec

10 5 0

1

2 3 Time (sec)

4

5

F

6

20msec

0

1

2 3 Time (sec)

4

5

6

0

1

2 3 Time (sec)

FIGURE 7.2 (continued) © 2008 by Taylor & Francis Group, LLC

4

5

6

Reach Right

PPC po012203-reach-sig044b

Reach Left

5msec

H


Reach Right

M1 po012203-reach-sig031d


G

30 25 20 15 10 5 0 –1



40 35 30 25 20 15 10 5 0 –1


20

0 –1

M1 po012203-reach-sig024a

25 20 15 5msec

10 5 0 –1

0

1

2 3 Time (sec)

4

5

6


E


Reach Left

Reach Right

PMd th051304-reach-sig059d

132



133

a method of analysis that can detect more elaborated patterns—an analysis that is able to detect the temporally varying information in the neural spike train. Initial analysis of neuronal recordings typically consists of investigating the response properties of each individual recorded neuronal unit with respect to a particular event of interest. Raster plots and peri-event time histograms (PETHs) can provide a good visual representation of the firing patterns of individual cells around, for example, the delivery of a vibrotactile stimulus or the onset of a movement. PETH analysis involves measuring the average firing rate (spikes per second) in a series of time-sequential bins across an epoch of interest. After each session, offline computer analysis is used to construct peri-event time histograms (PETHs) of the firing of each neuron during trials. The average intensity of neuronal firing within a control background firing period is calculated, and again for each of several user-defined response epochs. The “evoked unit response” or instantaneous firing rate of the neuron can then be expressed as: (1) spikes per second, (2) average instantaneous firing rate over different peri-event epochs, (3) firing rate minus background, or (4) firing rate divided by background (i.e., signal or noise). We then use standard statistical tests to determine whether the differences in firing rate during different epochs are significant, including the Kolmogorov–Smirnov test, student’s t-test, and the cumulative frequency histogram, as appropriate.

DISTANCE INDEX (DI) ANALYSIS Despite the advances that firing rate analysis has provided, relying on average firing rate may be significantly limiting (Nicolelis, Ghazanfar et al. 1997). It assumes that the relationship between the neuronal firing rate and the event of interest is simple and unchanging (i.e., stationary), which is often not the case. Firing rate analysis ignores the temporal components of the neuronal encoding. Indeed, patterns of neuronal firing may convey more information than the average firing rate alone (Laubach, Shuler et al. 1999). These temporal patterns of firing (TPF) can be analyzed in a number of ways, including principal component analysis, discriminant analysis, and independent component analysis. To facilitate this analysis, we developed a linear technique that we refer to as the Distance Index (DI) method (Figure 7.3) (Sandler, Kralik et al. 2003). We have found the results from DI analysis to be very similar to other methods, however, the DI analysis had the advantages of being faster to perform and conceptually simpler. The DI method has since been further validated by other researchers, who came to the conclusion that it works as well as, or better than, other commonly used methods, but is more efficient to use with large data sets and provides superior

FIGURE 7.2 (continued) The firing rate of the PPC unit in (E) increased approximately half a second into both trial types, but rapidly decreased for right reach trials (red) although remaining elevated for another several seconds in left reach trials (blue). The M1 unit in (F) demonstrated elevated firing rate only during the latter half of left reach trials (blue) whereas the M1 unit in (G) demonstrated a decreased firing rate only during the latter half of the right reach trials (red). The PPC unit in (H) showed a slight difference in firing rate between left reach (blue) and right reach (red) trials only near the middle of the trial.


Right 100

150

200 Trial#2

100 150 Time (msec)

50

100

150

200 Right

0

200

50

100 150 Time (msec)

200

D

10 8 6 4 2 0

DI(shuffle001) = .035 50 [R1 R2 R3 R4 ... R20] 40 0

50 100 Time (msec)

150

200

Frequency

Average FR (All Right Trials)

14 12 10 8 6 4 2 0

Average FR (All Left Trials)

A

C

50 100 Time (msec) Distance Index =

150

FIGURE 7.3 © 2008 by Taylor & Francis Group, LLC

20

0 0.01

200

(R1–L1)2 + (R2–L2)2 + ... +(R20–L20)2

= 0.044114

30

10

[L1 L2 L3 L4 ... L20]

0

Unshuffled (True) DI (0.044114)

E

0.02 0.03 0.04 DI Histogram (1000 Shuffled Trials) 4/1000 shuffled trials lie further out then the true, unshuffled data... Significance = 0.004

0.05


50

Left 0

Left 0

B

Trial#1

Reach Direction

Firing Rate

50

Firing Rate

0

Trial #2

Firing Rate of Neuron (200ms Epoch, 10msec bins) Firing Rate

Firing Rate

Trial #1

Reach Direction

134

Firing Rate of Neuron (200ms Epoch, 10msec bins)


135

abilities to investigate time-dependent properties of the neural code (Foffani and Moxon 2004). The DI method is a linear model in which a neuron’s firing rate is calculated for many small bins (subdivisions) of an epoch of interest. The first step involves normalizing the firing rates (to avoid undue weight given to a tonically more active neuron). An individual neuron’s series of firing rates in each of these bins is then formed into a vector, and the vector formed from the average trial of a given class can be compared with that of another class. The Euclidian difference between those average vectors is calculated; the resulting number is the DI. Statistical methods are then employed to determine whether that DI is significant, i.e., the level of certainty about whether the neuron can distinguish between the classes of trials. We use a bootstrapping shuffle test to determine the significance of the DI without having to make a priori assumptions about the distribution of the data. In this method, the different classes of trials are shuffled randomly a large number of times. For each shuffle, a shuffled DI is calculated to build a distribution of possible DIs for the given data set. The significance of the true DI is then determined by locating it on the histogram of potential DIs. This analysis is performed with all neurons in the sample to determine which ones are involved in a particular function.

ANALYSIS OF NEURAL ENSEMBLE DATA Studying the development of interactions of highly distributed networks of neurons located across multiple areas of the brain requires a reliable method for quantifying the information available in the firing activity recorded from a given neuronal ensemble. Such a measure is also necessary for the purpose of studying the functional implications of neuronal plasticity over the course of learning the task. With respect to our sensory discrimination tasks, the relevant variable is the monkey’s choice of response (reach target). Thus, in a two-choice task, our methods of choice for quantifying neural information about the stimulus take the form of categorization algorithms that classify single trials as “reach to target one” or “reach to target two” on the basis of recorded neuronal firing rates. The resulting metric can be expressed as the percent of trials classified correctly and can also be further analyzed using information analysis. FIGURE 7.3 (see facing page) Schematic of Distance Index (DI) method, a linear model that is used to examine the relationship between the firing of a neuron and behavioral variables, such as reach direction. The DI method takes into account both firing rate and temporal patterns of firing during an epoch of 200 ms, by calculating the firing rate during 20 10 ms bins for each individual trial (a). These firing rates for the average trial of each class (right reach or left reach) are then expressed as a 1 × 20 vector (b). The Euclidian difference between those average vectors is calculated; the resulting number is the DI (c). A bootstrapping shuffle test is then used to determine the significance of the DI without making a priori assumptions about the distribution of the data (d). In this method, the different classes of trials are shuffled randomly a large number of times. For each shuffle, a shuffled DI is calculated to build a distribution of possible DIs for the given data set. The significance of the true DI is then determined by locating it on the histogram of potential DIs (e). By repeating this analysis with all neurons in the sample, the percentages of neurons that respond significantly different from to the two classes of the parameter can be calculated.


136


USING ARTIFICIAL NEURAL NETWORKS (ANNS) FOR SINGLE TRIAL CLASSIFICATION BASED ON NEURAL ENSEMBLE RECORDINGS Extracting information from the firing patterns of neuronal ensembles is difficult largely due to the combinatorial complexity of the problem, and the uncertainty about how information is encoded in the nervous system. Our previous studies indicated that a large number of neurons are usually active in primate S1, PPC, PMd, and M1 in order to encode tactile, decision making, and reaching information (Nicolelis, Ghazanfar et al. 1998). We also observed that there is a high degree of variability in an individual neuron’s spike trials. Nevertheless, at the level of neural ensembles, tactile information can be reliably represented on a trial-to-trial basis. That is, although the number of spikes and the temporal position of spikes of a given neuron can vary from trial to trial, the overall neuronal population response predicts multiple attributes of a tactile stimulus with great precision. Thus, statistical pattern recognition approaches, such as those based on different architectures of multilayer artificial neural networks (ANNs) can be effective tools in the investigation of putative population encoding schemes. Indeed, several groups have successfully applied ANNs to investigate potential neuronal coding schemes in the auditory (Middlebrooks, Clock et al. 1994), visual (Kjaer, Hertz et al. 1994), and somatosensory systems (Nicolelis, Ghazanfar et al. 1998).

LVQ-BASED POPULATION ANALYSES Previous studies have demonstrated that differences in neuronal activity as a function of task performance can be quantified using a nonparametric method for statistical pattern recognition called learning vector quantization (LVQ) (Ghazanfar, Stambaugh et al. 2000, Krupa, Wiest et al. 2004). This method has been used previously in our lab to study neuronal responses in the somatosensory and motor cortex (Ghazanfar, Stambaugh et al. 2000; Laubach, Wessberg et al. 2000). We have found that for neuronal ensemble data, LVQ outperforms most other modern methods for statistical pattern recognition. LVQ is carried out using Matlab (The Mathworks) and the Matlab Neural Network Toolbox in which LVQ is implemented as a two layer artificial neural network. The inputs to the network are single-trial PETHs. The first layer in the LVQ network is called the competitive layer, and its elements are equivalent to the “codebook vectors” in Kohonen’s original algorithmic implementation of LVQ (Kohonen 1997). Each competitive neuron is assigned to a class in the training data (e.g., trials with narrow or wide apertures). The LVQ network is trained such that neurons in the competitive layer become matched to typical firing patterns recorded during performance of the behavioral discrimination task. The second layer of artificial neurons essentially computes the distance between single trials in the training or testing data sets and the outputs of the neurons in the competitive layer (i.e., the product of the coefficients for a given competitive neuron and the recorded neuronal ensemble response). For our applications, we modified the code provided by The Mathworks to implement the optimized LVQ (OLVQ) learning rule (Kohonen 1997). This variant of the



137

LVQ algorithm adaptively modifies the learning rate by decreasing the learning rate for each competitive neuron if the selected example is correctly classified and otherwise increasing it. This results in rapidly decreasing learning rates for competitive neurons that are far from the decision boundaries (i.e., near the center of a subset of data from a given class) and increasing learning rates for competitive neurons near the decision boundaries.

QUANTIFICATION OF NEURONAL RESPONSES WITH LVQ Single-trial PETHs (50 ms bins) are constructed for the epoch starting 200 ms before presentation of the stimulus until after the reach is complete. The data sets are divided into training and testing subsets. The training data are used to initialize the LVQ networks by setting the coefficients for each competitive neuron dedicated to a given type of trial equal to the mean neuronal ensemble response for that type of trial (plus a small random noise term). We use as many competitive neurons as twice the number of classes of data to be classified. Leave-one-out cross-validation (i.e., iteratively use all but one trial as training data and test performance of the LVQ network on a single “hold-out” trial) is used to estimate error rates and confusion matrices for each data set. Results are quantified in terms of percentage of single trials classified correctly and the informational entropy derived from confusion matrices (Krippendorff 1986).

POPULATION DISTANCE INDEX (PDI) ANALYSIS FOR SINGLE TRIAL CLASSIFICATION BASED ON NEURAL ENSEMBLE RECORDINGS The pDI can be calculated for an ensemble of neurons (e.g., the entire set of neurons recorded from a single cortical area) in much the same way as the DI discussed above is calculated for an individual neuron. In calculating the pDI, vectors consist of appended vectors of the type calculated for the DI analysis, ending up with a vector of length n * m, where n is the number of neurons in the neuronal ensemble of interest, and m is the number of bins into which the epoch was divided. Significance is then calculated by using the same bootstrapping shuffle test as for the DI. Classification can be accomplished by using a leave-one-out cross-validation method for training and testing the model (Nicolelis 1998; Kralik, Dimitrov et al. 2001), in which one single trial is removed from the data set and the remaining trials are used to calculate the average vectors (remaining vectors) for the two classes (Figure 7.4). The left-out trial is then classified, based on the shortest distance between it and the two remaining vectors. This procedure is repeated, sequentially leaving out each trial in the sample, in order to estimate the predictive ability of the data set, and thus the potential relationship between the neuronal activity and the stimulus or behavior of interest. Results are quantified in terms of percentage of single trials classified correctly and the informational entropy derived from confusion matrices. In the latter case, information is calculated by determining how much of the uncertainty about the reach direction that exists at the beginning of the trial is reduced by the end of the trial due to the pDI calculations performed on the recorded neuronal firing. Information


Trial#6

150

200

Average FR (N-1 Right Trials) Average FR (All Left Trials)

Right 0

0

50

100

150

DI(R) =

(LO1–rR1)2 + (LO2–rR2)2 + ... +(LO20–rR20)2

DI(L) =

(LO1–rL1)2 + (LO2–rL2)2 + ... +(LO20–rL20)2

200 Right

100

0

50

150

200

0

50 100 150 Time (msec)

200

Left

DI(R) < DI(L) => Classify as Right Reach Trial

C

DI(L) < DI(R) => Classify as Left Reach Trial

Left 0

50

100

150

200 Repeat (N) Times, Sequentially Leaving Each Trial Out

Right

D 0

50

100

150

14 12 10 8 6 4 2 0

200

Construct Confusion Matrix Actual

10 8 6 4 2 0

Calculate “Remaining Vectors”

Reach Right Reach Left

[rR1 rR2 rR3 rR4 ... rR20] 0

50 100 150 Time (msec)

200

E

45

15

5

35

[rL1 rL2 rL3 rL4 ... rL20]

F 0

50 100 150 Time (msec)

FIGURE 7.4 © 2008 by Taylor & Francis Group, LLC

[LO1 LO2 LO3 LO4 ... LO20]

50 100 150 200

Left

A

B

Firing Rate

100

Predicted Reach Left Reach Right

Trial#5

Trial#1 50

200

Calculate % Correct, Information


Trial#4

Firing Rate Firing Rate Firing Rate Firing Rate

Trial#3

Firing Rate Firing Rate

Trial#2

To Which Remaning Vector Is “Left-Out” Trial Closer?

Right 0

138

Trial#1

Remove One Trial of Perievent Data Firing Rate of Neuron Reach Direction (200ms Epoch, 10msec bins)


139

is measured in bits, and this task is a 1 b problem (the monkey must either reach left or right); therefore, the maximum amount of information that could be obtained from neuronal recordings is 1 b, thus resulting in values between 0 and 1.

MOVING-WINDOW ANALYSIS A “moving-window” analysis is used to assess the time-course of information about the monkey’s decision. For this purpose, LVQ or pDI is applied sequentially to a 200– 500 ms epoch of neuronal ensemble activity that “moves” in 50–100 ms steps through the course of the trial. This analysis produces a continuous quantitative readout of the recorded population’s ability to distinguish between the possible classes of trial. Over recent years, we have acquired extensive experience with this type of analysis (Nicolelis, Ghazanfar et al. 1998; Ghazanfar, Stambaugh et al. 2000; Krupa, Wiest et al. 2004) and have been able to characterize neural ensemble responses in a way that minimized the number of a priori assumptions made about how populations of neurons encode information. The only parameters available to the LVQ ANN and pDI classifiers are the firing rate and the temporal patterning of neuronal firing within simultaneously recorded cortical ensembles.

CROSS CORRELATIONS TO STUDY NEURONAL INTERACTIONS To identify interrelationships between the different neurons (including those in separate cortical areas), we calculate pairwise crosscorrelation functions for all possible pairs of recorded cortical neurons (Carmena, Lebedev et al. 2003). This allows us to identify both time-synchronized neuronal firing and sequential (or time offset) firing.

PARTIAL DIRECTED COHERENCE (PDC) ANALYSIS To identify the direction of activity between neurons in different cortical areas at various points in the task, we use the method of partial-directed coherence (PDC) (Sameshima and Baccala 1999; Baccala and Sameshima 2001a; Baccala and Sameshima 2001b; Fanselow, Sameshima et al. 2001). Briefly, PDC is a frequency domain representation of the key concept of Granger causality, which states that an observed time series x(n) Granger-causes another series y(n), if knowledge of x(n)’s past significantly improves prediction of y(n). This relation between time series is not reciprocal, i.e., x(n) may Granger-cause y(n) without y(n) necessarily Grangercausing x(n). Nullity of PDC between two structures at a given frequency suggests FIGURE 7.4 (see facing page) Schematic of prediction model used to calculate the ability of an individual neuronal unit (or a population of neuronal units) to predict the class of a given single trial. We used the leave-one-out cross-validation method for training and testing the model, removing one single trial from the data set (a) before using the remaining trials to calculate the average vectors (remaining vectors) for the two classes (b). The left-out trial was then classified, based on the shortest distance between it and the two remaining vectors (c). This procedure was repeated, sequentially leaving out each trial in the sample (d), in order to construct a confusion matrix (e). Estimates of the predictive ability of the data set, and thus the relationship between the neuronal activity and the direction of reach, were then calculated both as percent correct prediction and by using information theory (f).


140


the lack of a direct link between those structures. We have previously adapted PDC for multivariate autoregressive modeling of signals derived from neuronal spiking (Fanselow, Sameshima et al. 2001). Because existing direct feedback relationships between each pair of channels are explicitly exposed, PDC allows the uncovering of coactivations among multichannel neuronal recordings by highlighting neuronal groups that possibly drive other neuronal groups. For the PDC analysis, we use the primary principle component of the activity collected from all of the electrodes in a given cortical area. We find the mean value of coherence across 1–50 Hz for each 2 s analysis time window throughout the recording session. These values are calculated for classical coherence and directed coherence between each pair of cortical areas. They are then correlated with the stimulus location and reach direction. Results for each coherence type and behavior can be normalized to the classical coherence value for the experiment.

PRINCIPLE COMPONENT ANALYSIS (PCA) PCA is used to reduce the large numbers of original neural signals to a smaller number of derived “components” that account for most of the variance observed in the original data set (Nicolelis, Ghazanfar et al. 1997; Nicolelis 1998; Kralik, Dimitrov et al. 2001). These components represent dimensions of information embedded in the firing pattern of the neural population, and they may reflect functional associations between the neurons in the ensemble. In our experience, neurons with similar functional characteristics, such as those related to a specific sensory input, or neurons located in the same area of the brain tend to have high coefficients on the same principal components, whereas neurons with dissimilar functional associations or from different areas of the brain tend to be clearly separated onto different principal components (Nicolelis, Ghazanfar et al. 1997; Nicolelis 1998). Each recorded neuron is treated as a separate variable in the PCA. Time series of the firing rates of each neuron (e.g., rates obtained in 10 to 25 ms bins over the trial) are correlated with those of all other neurons in the population, generating a correlation matrix of all neurons. From this correlation matrix, a series of principal components is extracted. Each of these components is formed by the weighted linear sum of the firing patterns of individual neurons, with neurons contributing differentially to the different components, as reflected in the component weights. Note that neither the response properties of the neurons nor their anatomical location are made available to the PCA algorithm; nonetheless, as described previously, these features are often pulled out by PCA. Principle components can then be used as inputs to the various analyses we have described here, including the LVQ ANN, cross-correlation, and PDC analyses discussed above.

CONTROL ANALYSES The accuracy of potential encoding mechanisms employed by cortical and thalamic neural ensembles can be investigated by several manipulations of the original data set (Ghazanfar, Stambaugh et al. 2000). These include sequentially removing neurons, one at a time, from each cortical ensemble in order to measure the variation


20 0.2

40

Window

Color Figure 1.11 8 × 8 electrode array fabricated using the discrete-wired method.

80 100

0.8

40

0.7 20 40 Window

0.15

0.1

120 140

0.05

160 180 20 40 Window

Window

Neuron

60

0.9

20

0.95

20

0.9 0.85

40 20 40 Window

0.8

Color Figure 4.12 Ensemble permanence—first principal neuron during 30 s windows along one experimental session (left). The variance carried by the first principal neurons computed at one window (vertical axis) during another window (x axis) (top right), and the cosine-angle between the first principal neurons from the two window (bottom right).


7046.indb 271

10/30/07 9:44:38 AM

A

B

Color Figure 5.2 (A) Upper-left panel. Craniotomy in the area above the dorsolateral striatum. An array of 16 microwires is ready for implantation. Lower-left panel. Awake mouse with 32 microwires inserted bilaterally in the dorsolateral striatum during a recording session. (B) Screen of our acquisition system during a recording session of the mouse presented in the left, one and a half months after surgery. The selected channel shows three clearly isolated cells recorded simultaneously from one electrode. Twenty-one other channels are also recording single-neuron activity.

Color Figure 6.1 Multiple individual whisker stimulator for (A) anesthetized and (B) head-restrained awake rats.


7046.indb 272

10/30/07 9:44:42 AM

Reach Right


Reach Right


M1

PPC

th031504-reach-sig036d


20 15

5msec

10 5 0

-1

0

1

3

2

Time (sec)

4

5

B

6

25 20 15

5msec

10 5 0

1

2

3

Time (sec)

4

5

D

6


1

2

3

Time (sec)

4

5

6

PMd

th051304-stim-sig014c

40 35 30 25 20

5msec

15 10 5 0

-1

0

1

2

3

Time (sec)

4

5

6

M1

po012203-reach-sig024a



0

Reach Right

Reach Right

PMd

th051304-reach-sig059d

30 25

15

5msec

20msec

20 15

10

10

5 -1

0

1

2

3

Time (sec)

4

5

F

6

5 0

-1

0

1

2

3

Time (sec)

4

5

6


Reach Right


Reach Right

0

-1


20

E

0

Stim Left

Stim Left Firing Rate (spikes/sec)

30

-1

5

Stim Right

Stim Right

PPC

th051304-stim-sig036d

0

10msec

15 10



C

30 25 20


A


Reach Left

th061704-reach-sig023b

M1

PPC

po012203-reach-sig031d

40 35 30 25 20 15 10 5 0


G


Reach Left

Reach Left

po012203-reach-sig044b

25 20

5msec

15

5msec

10

-1

0

1

2

3

Time (sec)

4

5

6

H

5 0

-1

0

1

2

3

Time (sec)

4

5

6

Color Figure 7.2 Complex task-related neuronal responses.


7046.indb 273

10/30/07 9:44:55 AM

t

A

B

Cx Hi Th

CP

C

Active Quiet Waking Whisker Exploration (AE) (QW) Twitching (WT)

Rapid-eyemovement (REM) Sleep

Slow-wave Sleep (SW)

Cx Th Hi CP

1 sec 1 sec 8

(Hz)

Cx 60 40 20 0 Th 60 40 20 0 Hi 60 40 20 0 CP 60 40 20 0

6

(Hz)

* *

*

*

5

*

(Hz)

*

Time State REM SWS QW AE WT

log (Power)

7

(Hz)

D

4 3600

4200

4800

Sec

Color Figure 8.1 Local field potentials and behavioral states. (A) Schematic representation of the location of the multielectrode implants on a parasagittal section of a rat brain. The primary somatosensory cortex (Cx), the hippocampus (Hi), the ventromedial posterior nucleus of the thalamus (Th), and the caudate-putamen (CP) were implanted with microelectrode arrays. (B) Photograph of a 4 × 16 tungsten microwires array used for recordings in the somatosensory cortex (barrel field). (C) Raw simultaneous LFP recordings in the four forebrain regions. Five behavioral states were coded based on behaviors and visual inspection of LFP traces. (D) LFP power spectrograms, aligned with staircase representation of behavioral states. All areas show simultaneous state-dependent variations of LFP spectral pattern. Note that AE and REM both show pronounced theta rhythm (white asterisk).


7046.indb 274

10/30/07 9:44:57 AM

0.5 0.4 0.3 0.2 0.1 0 –0.1 –0.2 –0.3 –0.4 –0.5

Amplitude Difference (arbitrary unit)

Weights (arbitrary unit)

A

PC1 PC2 0

5

4 3 2 1 0 –1

SWS REM

–2

10 15 20 25 30 35 40 Frequency (Hz)

0

5

WK WK

10 15 20 25 30 35 40 Frequency (Hz)

B

10 sec Sliding Window FFT

LFP States

0.5–20 Hz

SWS

REM

WK

Y-axis

0.5–55 Hz 0.5–4.5 Hz

X-axis

0.5–9.0 Hz

2-D Representation

tio 1

Smoothing

State-map Color Figure 8.2 Construction of state-space map. (A) Left panel, spectral features (PC1 and PC2) obtained by applying PCA to the spectrogram of individual forebrain LFPs in one rat. Right panel, difference of the average amplitude spectrum in different behavioral states. Each trace represents the result calculated from one forebrain LFP. Notice both methods obtained similar spectral features. (B) Construction of state-space map: A sliding window Fourier transform was applied to each LFP signal to calculate two spectral amplitude ratios at 1-sec temporal resolution. The spectral ratios obtained from individual LFPs were combined using PCA and represented by the first PC. The two PCs (blue traces) were further smoothed with a 20-bin Hanning window (red traces). The two smoothed spectral ratios define the 2-D state-space. Note that clear cluster structures emerged in the 2-D state space after smoothing, and that either feature alone (1-D histogram) was insufficient for clear cluster separation.


7046.indb 275

10/30/07 9:45:02 AM

A

Scatter Plots

B

Density Plots

C

Speed Plots

IS SWS QW REM

AE

IS SWS QW

REM

AE

IS

SWS

REM

QW WT AE

QW

AE

SWS

REM

WT

Lo w

High

Min

Max

Color Figure 8.3 Global brain states. (A) Scatter plots of the 2-D state space, color-coded for behaviorally coded states. Each dot corresponds to a 1 s window from which the amplitude ratios were calculated. For all animals, 48 h of recording is displayed; to avoid graphic saturation, only 20% of the data points were evenly sampled and plotted. Note that clusters correspond to behavioral states. (B) Density plots, calculated from the scatter plots, show the conserved cluster topography and the relative abundance of various states. (C) Speed plots representing the average velocity of spontaneous trajectories within the 2-D state-space. Stationarity (low speed) can be observed within the three main clusters, whereas the maximum speed is reached during transitions from one cluster to another, i.e., between brain states. IS, intermediate sleep.


7046.indb 276

10/30/07 9:45:09 AM

v

7046.indb 277

Behavioral States

REM

WT

Fr.(Hz)

55

8 7 6

0

55

Auto

Auto Visual

2700

3200

3700

4200

AE + QW

SWS

REM

Acc.

0.93 (0.83)

0.93 (0.91)

1.00 (0.96)

Sen.

0.83 (0.62)

1.00 (0.92)

0.98 (0.63)

Spe.

1.00 (1.00)

0.89 (0.89)

1.00 (1.00)

4700 Sec

8 7 6

0

Visual Time

Behavioral States

2000 Time

2500

3000

3500

AE + QW

SWS

REM

Acc.

0.95 (0.87)

0.95 (0.94)

0.99 (0.96)

Sen.

0.91 (0.77)

1.00 (0.95)

0.94 (0.52)

Spe.

0.99 (0.99)

0.93 (0.93)

1.00 (1.00)

4000 Sec

Color Figure 8.4 Automatic state-coding algorithm. Comparison of automatic state-coding with behavioral states in two rats. 10/30/07 9:45:11 AM


log (Power)

SWS

Fr.(Hz)

AE + QW

Automatic Algorithm

Contour Maps

log (Power)

Automatic Algorithm

Contour Maps

0.65

B 0.65

0.6

0.6

0.55

0.55

0.5

0.5

0.45

0.45

Y

A

0.4

Scatter Plot 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65

X

AE + QW

SWS

0.4

70 60 50 40 30 20 Density Plot

10

0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65

REM

Three-dimensional Map

C

10 3

3

10

10×

10×

EMG

10 3

EMG

3

10

15×

15×

5×1

03

03

5×1

–0 0.7 0.70 0.65 0.6 0.60 0.5 0.55 0.50 0.4 0.45 0.3 0.40 0.35 0.2

Y

X

0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70

X

Color Figure 8.5 Global brain states in mice. Scatter plot (A) and density plot (B) of the 2-D state space generated using hippocampal LFPs recorded in one mouse. Similar cluster structures as those described in rats (Figure 8.3A) were found. (C) Neck EMG amplitude was plotted as the third dimension (arbitrary unit), showing that the REM cluster had low or absent EMG activity, whereas the adjacent WK cluster showed maximal EMG activity. Two views represent two different projections of the same 3-D scatter plot.


7046.indb 278

10/30/07 9:45:12 AM

Relative Power

Coherence

All Areas Combined 2–4 Hz

Cortex - Thalamus

Cortex - Putamen

2–4 Hz

2–4 Hz

0.8

IS SWS

REM

QW AE 5–8 Hz

0.4 0 –0.4 –0.8 log P

5–8 Hz

0.4 0 –0.4 –0.8

0.8 0.4 0 –0.4 –0.8

0.4 0.2 0 –0.2 –0.4 25–55 Hz

0.3 0.2 0.1 0 –0.1 –0.2 –0.3

0.4

0.4

0.3

0.3

0.2

0.2

0.6

25–55 Hz

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.6

0.6 0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.6 0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.2 0.15 0.1

8–14 Hz

0.1 14–22 Hz

0.5

0.25

5–8 Hz

0.1 8–14 Hz

0.5

0.6

2–4 Hz

0.1 5–8 Hz

0.1 14–22 Hz

14–22 Hz

0.5

0.1 8–14 Hz

0.6

0.5

0.1

0.8

8–14 Hz

0.6

Cortex - Hippocampus

25–55 Hz

0.3

14–22 Hz

25–55 Hz

0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.5

0.25

0.4

0.2

0.3

0.15

0.2

0.1

0.1

Color Figure 8.6 Gradients within global brain states. Overlay of spectral power and coherence in different frequency bands on the 2-D state space of one rat. Spectral power was calculated on the log scale, averaged across all forebrain regions, with the mean log power subtracted. Thus, the amplitude represents the relative power above or below the mean value at each frequency band. Coherence was calculated for pairs of LFP channels as indicated. Several gradients were observed within global brain states: Note the high delta power and coherence during deep SWS (upper right end of the SWS cluster), while spindle power is higher at the other end of SWS cluster (shallow SWS). Also, within the WK cluster, active exploration (left end) and quiet awake (right end) showed prominent differences in the power and coherence at the theta and gamma frequency bands. In addition, the IS state was characterized by high power and coherence over a very broad frequency range.


7046.indb 279

10/30/07 9:45:14 AM

WK–SWS

A

WK

SWS

REM

WK

Cx Th Hi CP

REM–WK Cx Th Hi CP

IS

SWS–IS–REM Cx Th Hi CP

IS

SWS–IS–QW–SWS Cx Th Hi CP

REM–IS–QW–SWS IS Cx Th Hi CP

Pairwise Coherence Spectra

1.0

50 40 30 20 10

State

Coherence

Cx - Hi Frequency (Hz)

B

Time

WK

5600

SWS

REM

5800

6000

6200

6400

6600

sec

0.8 0.6 0.4 0.2 0.0

WT

Color Figure 8.7 Trajectory analysis.


7046.indb 280

10/30/07 9:45:15 AM

Color Figure 9.1 Experimental setup, behavioral tasks, changes in performance with training, EMG records during pole and brain control, and stability of model predictions.


7046.indb 281

10/30/07 9:45:17 AM

Color Figure 9.2 Performance of linear models in predicting multiple parameters of arm movements, gripping force, and EMG from the activity frontoparietal neuronal ensembles recorded in pole control.


7046.indb 282

10/30/07 9:45:18 AM

A

B Acids

Taste bud

PKD2L1

Sweet, Bitter, Umami

Salt

T1Rs or T2Rs

ENaCs αβγ

G Protein

PLCβ2 PIP2 DAG TRPM5 Channel Endoplasmic Reticulum Neuro– transmitters

IP3 Taste Receptor Cell

Ca2+

IP3R3

Na+ Channel

Depolarization

Na+

K+ Ca2+

K+ Channel

GPCRs Peptides, Hormones

rNTS

Ca2+ Channel

ATP

Gustatory Cranial Nerves VII, IX, X

P2X Receptor

Action Potential

Color Figure 10.1 Illustration of a taste bud, taste receptor cell, and associated neurons.


7046.indb 283

10/30/07 9:45:22 AM

Ventroposterior Medial Nucleus of the Thalamus Insular Cortex

To Orbitofrontal Cortex Parabrachial Nucleus

Glossopharyngeal Nerve IX Chorda Tympani Nerve VII

Nucleus of the Solitary Tract (Rostral Division)

Lingual Tonsilar Nerve X

COLOR Figure 10.2 Anatomy of the principal central gustatory pathways. Taste-specific information is conveyed by cranial nerves VII, IX, and X (blue lines) to the rNTS (rostral division of the nucleus tractus solitarius) in the medulla. In primates, fibers (red lines) from second-order taste neurons in the rNTS project ipsilaterally to the VPMpc (parvicellular part of the ventroposterior medial) nucleus of the thalamus. Shown in orange is the PBN (parabrachial nucleus) of the pons that, in rodents, is a relay for rNTS afferents and projects third-order fibers to the VPMpc. In both cases, thalamic efferents (green lines) project to the insula, defining the primary gustatory cortex (GC) which, in turn, projects (black lines) to the orbitofrontal cortex (OFC), sometimes defined as a secondary cortical taste area. Not shown are descending projections from the GC and OFC to subcortical structures in the ventral forebrain that, in rodents, also receive ascending projections from PBN and rNTS (see text). The insula projects to the amygdala (Shi and Cassell 1998), which in turn projects to the basal forebrain, lateral hypothalamus, substantia nigra pars compacta, and ventral tegmental area (Scott and PlataSalaman 1999; Fudge and Haber 2000), the latter being the origin of the mesolimbic dopaminergic projection to the nucleus accumbens, a part of the ventral striatum. In turn, the OFC projects to the ventral striatum, lateral hypothalamus, and amygdala (Öngür and Price 1998; Shi and Cassell 1998; Scott and Plata-Salaman 1999), and these subcortical structures are also mostly interconnected. (Adapted from Simon, S.A. and de Araujo, I.E. et al. (2006). The neural mechanisms of gustation: a distributed processing code. Nat Rev Neurosci 7(11): 890–901). © 2008 by Taylor & Francis Group, LLC

7046.indb 284

10/30/07 9:45:24 AM

Licking

A Sucrose 1

Neurons

2 3 4 5 6 7 8 –0.5

0.0

Time (s)

B Water

0.5

1.0

Licking

1

Neurons

2 3 4 5 6 7 8 –0.5

Correct Discrimination (%)

C

66 64

0.0

52 50 –0.5

0.5

Anticipation

1.0

Licking

* a

62 60 58 56 54

Time (s)

* *

d

d

b

–0.25

0.0 Time (s)

0.25

0.5

Color Figure 10.8 Ensemble activity of OFC neurons discriminates and anticipates natural rewards in the rat.


7046.indb 285

10/30/07 9:45:25 AM

7046.indb 286

B

0.42

0

20

40

60 Trials

80

100

600

D

400 300 200 100 0

0

20

40

60 Trials

0.5

1.0

1 2 3 4 5 6 7

500 ITI (s)

Satiety-Sensitive Neurons

0.05

0.46

0.38

Normalized Firing Rate

C

0.50

8 Normalized Firing Rate

Normalized Firing Rate

A

80

100

0.5 0.4 0.3 0.2

Mean Neurons 1-7

0.1 25

45

65

Trials

Color Figure 10.11 Coding of hunger and satiety states by neuronal ensembles and single cells. 10/30/07 9:45:26 AM


85

105


141

in discrimination capability as a function of the ensemble size. A power function (Carpenter, Georgopoulos et al. 1999) can be used to estimate the number of neurons needed to achieve a certain level of neuronal ensemble performance (e.g., 99% correct discrimination of single trials) because information capacity changes nonlinearly with an increasing number of neurons. The integration time used to describe each neuron’s sensory response (bin size) can also be varied, a procedure designed to alter the temporal structure of each neuron’s sensory response and to test the interaction between rate and temporal coding. Procedures for decorrelating the neuronal data are also used, by either shifting the timing of individual spikes or shuffling the trial sequence for each neuron randomly. Finally, S1, PPC, PMd, and M1 ensembles can be analyzed in isolation or together in order to investigate the potential role of corticocortical interactions.

CONCLUSIONS New advances in multielectrode ensemble recording, allowing simultaneous recordings from large numbers of neurons, now enable us to study the changing functional interrelationships between various cortical areas over the course of learning. Ongoing learning can now be studied longitudinally, with sufficient data obtained from multiple cortical areas during each session, and across sessions, over a time course that can be as long as years. We suggest that the techniques of ensemble neuronal recording and advances in analysis of the large amounts of data that can be obtained will facilitate a broader, system-wide understanding of information processing in the brain which will, in turn, aid in the future study of learning and cognition in both the normal and the pathological states.

REFERENCES Baccala, L.A. and Sameshima, K. (2001a). Overcoming the limitations of correlation analysis for many simultaneously processed neural structures. Prog Brain Res 130: 33–47. Baccala, L.A. and Sameshima, K. (2001b). Partial directed coherence: a new concept in neural structure determination. Biol Cybern 84(6): 463–74. Boussaoud, D. and Wise, S.P. (1993). Primate frontal-cortex—effects of stimulus and movement. Exp Brain Res 95(1): 28–40. Brasted, P.J. and Wise, S.P. (2004). Comparison of learning-related neuronal activity in the dorsal premotor cortex and striatum. Eur J Neurosci 19(3): 721–740. Carmena, J.M. and Lebedev, M.A. et al. (2003). Learning to control a brain-machine interface for reaching and grasping by primates. PLoS Biol 1(2): E42. Carpenter, A.F. and Georgopoulos, A.P. et al. (1999). Motor cortical encoding of serial order in a context-recall task. Science 283(5408): 1752–7. Chen, L.L. and Wise, S.P. (1995). Supplementary eye field contrasted with the frontal eye field during acquisition of conditional oculomotor associations. J Neurophysiol 73(3): 1122–34. Erickson, R.P. (1968). Stimulus coding in topographic and nontopographic afferent modalities: on the significance of the activity of individual sensory neurons. Psychol Rev 75(6): 447–65. Fanselow, E.E. and Sameshima, K. et al. (2001). Thalamic bursting in rats during different awake behavioral states. Proc Natl Acad Sci U.S.A 98(26): 15330–5.


142


Foffani, G. and Moxon, K.A. (2004). PSTH-based classification of sensory stimuli using ensembles of single neurons. J Neurosci Methods 135(1–2): 107–20. Freeman, W.J. (1975). Mass action in the nervous system. New York, Academic Press. Fuster, J.M. (1995). Memory in the cortex of the primate. Biol Res 28(1): 59–72. Georgopoulos, A.P. and Schwartz, A.B. et al. (1986). Neuronal population coding of movement direction. Science 233(4771): 1416–9. Georgopoulos, A.P. and Taira, M. et al. (1993). Cognitive Neurophysiology of the Motor Cortex. Science 260(5104): 47–52. Ghazanfar, A.A. and Stambaugh, C.R. et al. (2000). Encoding of tactile stimulus location by somatosensory thalamocortical ensembles. J Neurosci 20(10): 3761–75. Gold, J.I. and Shadlen, M.N. (2003). The influence of behavioral context on the representation of a perceptual decision in developing oculomotor commands. J Neurosci 23(2): 632–51. Hebb, D.O. (1949). The organization of behavior: A neuropsychol theory. New York, Wiley. Horwitz, G.D. and Newsome, W.T. (2001). Target selection for saccadic eye movements: direction-selective visual responses in the superior colliculus. J Neurophysiol 86(5): 2527–42. Kaas, J.H. (1994). Aotus: The owl monkey. Baer, J.F., Weller, R.E., and Kakoma, I. San Diego, Academic. Kim, J.N. and Shadlen, M.N. (1999). Neural correlates of a decision in the dorsolateral prefrontal cortex of the macaque. Nat Neurosci 2(2): 176–85. Kjaer, T.W. and Hertz, J.A., et al. (1994). Decoding cortical neuronal signals: network models, information estimation and spatial tuning. J Comput Neurosci 1(1–2): 109–39. Kohonen, T. (1997). Self-organizing maps. New York, Springer. Kralik, J.D. and Dimitrov, D.F., et al. (2001). Techniques for long-term multisite neuronal ensemble recordings in behaving animals. Methods 25(2): 121–50. Krippendorff, K. (1986). Info theory: Structural models of qualitative data, Thousand Oaks, CA, Sage Publications. Krupa, D.J. and Wiest, M.C. et al. (2004). Layer-specific somatosensory cortical activation during active tactile discrimination. Science 304(5679): 1989–1992. Lashley, K.S. (1950). In search of the engram. Symp. Soc. Exp. Biol., Cambridge, Cambridge University Press. Laubach, M. and Shuler, M. et al. (1999). Independent component analyses for quantifying neuronal ensemble interactions. J Neurosci Methods 94(1): 141–54. Laubach, M. and Wessberg, J. et al. (2000). Cortical ensemble activity increasingly predicts behaviour outcomes during learning of a motor task. Nature 405(6786): 567–71. McCoy, A.N. and Crowley, J.C. et al. (2003). Saccade reward signals in posterior cingulate cortex. Neuron 40(5): 1031–40. McCoy, A.N. and Platt, M.L. (2005). Expectations and outcomes: Decision-making in the primate brain. J Comp Physiol 191(3): 201–11. Middlebrooks, J.C. and Clock, A.E. et al. (1994). A panoramic code for sound location by cortical neurons. Science 264(5160): 842–4. Mitz, A.R. and Godschalk, M. et al. (1991). Learning-dependent neuronal activity in the premotor cortex: activity during the acquisition of conditional motor associations. J Neurosci 11(6): 1855–72. Nicolelis, M. (1998). Methods for simultaneous neuronal ensemble recordings. Methods for neural ensemble recordings. Nicolelis, M. Ed. Boca Raton, FL, CRC Press. Nicolelis, M.A. and Chapin, J.K. (1994). Spatiotemporal structure of somatosensory responses of many-neuron ensembles in the rat ventral posterior medial nucleus of the thalamus. J Neurosci 14(6): 3511–32.



143

Nicolelis, M.A.L. and Baccala, L.A. et al. (1995). Sensorimotor encoding by synchronous neural ensemble activity at multiple levels of the somatosensory system. Science 268(5215): 1353–1358. Nicolelis, M.A.L. and Ghazanfar, A.A. et al. (1997). Reconstructing the engram: Simultaneous, multisite, many single neuron recordings. Neuron 18(4): 529–537. Nicolelis, M.A.L. and Ghazanfar, A.A. et al. (1998). Simultaneous encoding of tactile information by three primate cortical areas. Nat Neurosci 1(7): 621–630. Pesaran, B. and Pezaris, J.S. et al. (2002). Temporal structure in neuronal activity during working memory in macaque parietal cortex. Nat Neurosci 5(8): 805–11. Platt, M.L. and Glimcher, P.W. (1999). Neural correlates of decision variables in parietal cortex. Nature 400(6741): 233–8. Quintana, J. and Fuster, J.M. (1992). Mnemonic and predictive functions of cortical neurons in a memory task. Neuroreport 3(8): 721–4. Salzman, C.D. and Britten, K.H. et al. (1990). Cortical microstimulation influences perceptual judgements of motion direction. Nature 346(6280): 174–7. Sameshima, K. and Baccala, L.A. (1999). Using partial directed coherence to describe neuronal ensemble interactions. J Neurosci Methods 94(1): 93–103. Sandler, A.J. and Kralik, J.D. et al. (2003). Neuronal correlates of primate somatosensorimotor learning. Society for Neuroscience Annual Meeting, New Orleans, LA. Shadlen, M.N. and Newsome, W.T. (2001). Neural basis of a perceptual decision in the parietal cortex (area LIP) of the rhesus monkey. J Neurophysiol 86(4): 1916–36. Sherrington, C.S. (1906). Integrated action of the nervous system. Cambridge, Cambridge University Press. Thoenissen, D. and Zilles, K. et al. (2002). Differential involvement of parietal and precentral regions in movement preparation and motor intention. J Neurosci 22(20): 9024–34. Wilson, M.A. and McNaughton, B.L. (1994). Reactivation of hippocampal ensemble memories during sleep. Science 265(5172): 676–9. Wise, S.P. and Murray, E.A. (2000). Arbitrary associations between antecedents and actions. Trends Neurosci 23(6): 271–276.


8

Defining Global Brain States Using Multielectrode Field Potential Recordings Shih-Chieh Lin and Damien Gervasoni

CONTENTS Introduction............................................................................................................ 145 Forebrain Dynamics In Different Wake-Sleep States ........................................... 146 Limitations of Existing State Identification Algorithms ............................ 148 State-Space Framework Reveals Global Brain States ........................................... 149 Data Collection ........................................................................................... 149 2-D State Space........................................................................................... 150 Global Brain States..................................................................................... 153 Global Brain States: The Neural Correlates of Wake-Sleep States ....................... 154 State-Coding Algorithm ............................................................................. 154 Comparison against Behavioral State Coding............................................ 155 Validation against EMG Activity in Mice.................................................. 155 Gradients and Functional Subdivisions Within Global Brain States..................... 157 Forebrain Dynamics During State Transitions ...................................................... 159 General Discussion ................................................................................................ 161 Advantages and Limitations of the State-Space Framework...................... 161 The Driving Forces: Neuromodulatory Systems ........................................ 163 State-Dependent Information Processing and Memory Formation ........... 164 Conclusions ............................................................................................................ 165 References.............................................................................................................. 165

INTRODUCTION Electrical activity is essential for neuronal communication. Over the years, in vivo multielectrode recordings have revealed that the electrical activities of individual neurons are not independent of each other. Instead, neurons tend to fire in a coordinated way within a given neural network. When measured as the electroencephalogram (EEG) or local field potential (LFP) signals, this neural coordination results in complex oscillatory activity patterns, which reflect synchronous synaptic potentials 145 © 2008 by Taylor & Francis Group, LLC

146


in a local network (Lopes da Silva 1991). Thus, unveiling the physiological mechanisms generating such complex oscillatory neural activity patterns is key to achieving a better understanding of how the brain operates in behaving animals. The dynamics of the forebrain is not random. Ever since the initial discovery of cerebral electrical activity by Caton (Caton 1875) in rabbits and monkeys, and later in humans by Berger (Berger 1929), different patterns of forebrain activity have been tightly linked to various behavioral and wake-sleep states. Indeed, these distinct patterns of neural activity have become incorporated as part of the criteria of wakesleep states (Green and Arduini 1954; Rechtschaffen and Kales 1968; Lopes da Silva and van Leeuwen 1969; Timo-Iaria et al. 1970; Moruzzi 1972; Winson 1972; Winson 1974; Gottesmann 1992; Steriade et al. 1993), suggesting that forebrain dynamics fall into several different regimes. This observation is intriguing because the same neural circuit can support several different dynamic regimes, which likely serve distinct roles in information processing and storage. Therefore, a quantitative description of its network dynamics can further reveal how the forebrain underlies so many fundamental functions in mammals. In this chapter, we first describe the forebrain oscillatory activity patterns associated with different wake-sleep states, and highlight limitations of existing state identification methods. Then, we introduce a novel state-space framework (Gervasoni et al. 2004) that we have employed to quantitatively describe global forebrain dynamics in rodents. Such an analysis revealed several distinct regimes in which the forebrain can operate. These regimes correspond to distinct global brain states and are correlated with the occurrence of major wake-sleep states observed in both rats and mice. In addition, the state-space framework proposed here has allowed us to characterize the gradient dynamics within global brain states, providing a quantitative description of state transition dynamics in rodents. We end this chapter by discussing the underlying driving forces and potential functional roles of global brain states.

FOREBRAIN DYNAMICS IN DIFFERENT WAKE-SLEEP STATES In mammals, the wake–sleep cycle consists of periodic alternation of three major behavioral states: waking (WK), slow-wave sleep (SWS), and rapid-eye-movement (REM) sleep. In conjunction with prominent changes in the behavior of the animal, these behavioral states are associated with remarkably different forebrain dynamics (Figure 8.1C). During WK, animals interact with their environment either actively or passively to acquire information about their immediate surrounding space. At the same time, cortical activity is dominated by low-amplitude fast oscillations (beta and gamma frequency bands, >20Hz) (Figure 8.1D) (Murthy and Fetz 1992; Steriade et al. 1993; Maloney et al. 1997; Rols et al. 2001). As the animal actively explores the environment, LFP activity displays prominent theta oscillations (5–7 Hz), especially in the hippocampus (Winson 1974; Buzsaki et al. 1990). These oscillations are modulated according to the level of arousal, attention, motor activity, and the presence or absence of incoming stimuli (Murthy and Fetz 1996; Fries et al. 2001; Fell et al. 2003; Buzsaki and Draguhn 2004). In contrast to WK, sleep appears as a periodical state of quiescence, in which there is minimal processing of incoming sensory information. The behavioral hallmark of


Defining Global Brain States Using Multielectrode Field Potential Recordings 147

Active Quiet Waking Whisker Exploration (AE) (QW) Twitching (WT)

Rapid-eyemovement (REM) Sleep

Slow-wave Sleep (SW)

&

)

)

)

# ''

"$!$( %

)

FIGURE 8.1 Local field potentials and behavioral states. (A) Schematic representation of the location of the multielectrode implants on a parasagittal section of a rat brain. The primary somatosensory cortex (Cx), the hippocampus (Hi), the ventromedial posterior nucleus of the thalamus (Th), and the caudate-putamen (CP) were implanted with microelectrode arrays. (B) Photograph of a 4 × 16 tungsten microwires array used for recordings in the somatosensory cortex (barrel field). (C) Raw simultaneous LFP recordings in the four forebrain regions. Five behavioral states were coded based on behaviors and visual inspection of LFP traces. (D) LFP power spectrograms, aligned with staircase representation of behavioral states. All areas show simultaneous state-dependent variations of LFP spectral pattern. Note that AE and REM both show pronounced theta rhythm (white asterisk).


148


this state could thus be summarized as a suspension of the activities of the waking state. However, sleep is not defined by a homogenous state. At the very least, it can be divided into two distinct states, each named after its main neurophysiological features. As animals fall asleep, slow coherent oscillations emerge in the cerebral cortex in different frequency bands (delta waves 1–4 Hz and spindles 7–14 Hz) (Steriade and McCarley 1990; Steriade et al. 1993) characterizing a first sleep state called slowwave sleep, or non-REM sleep. These oscillations are organized into complex wave sequences by a very slow oscillation (usually 0.6–1 Hz) (Amzica and Steriade 1995; Contreras and Steriade 1995), and are concomitant with a progressive sensory disconnection (Moruzzi 1972; Steriade and McCarley 1990). As SWS deepens, spindling activities tend to decrease while delta waves become more prominent, with an intensity that correlates with the duration of the preceding awake state (Dijk et al. 1990). After a certain time spent in SWS, a second sleep state ensues, characterized in humans by dreaming and conspicuous REM. Appropriately, this state has been named REM sleep (Aserinsky and Kleitman 1953; Dement and Kleitman 1957). During REM sleep, the cortical activity is quite similar to that observed during WK (Winson 1974; Borbely et al. 1984). That means that, for instance, in rats prominent gamma oscillations (Maloney et al. 1997; Gross and Gotman 1999) and theta oscillations (Vanderwolf 1969) are observed during this sleep state. Although cortical activity of REM sleep is remarkably similar to that of WK, the functional sensory disconnection of the cortex reaches a maximum and motor output is actively suppressed (Dement and Kleitman 1957; Jouvet et al. 1959; Jouvet 1962). This has motivated some authors to propose an alternative nomenclature for REM sleep, calling it either activated (Aserinsky and Kleitman 1953) or paradoxical sleep (Jouvet 1962).

LIMITATIONS OF EXISTING STATE IDENTIFICATION ALGORITHMS The various behavioral and neurophysiological features associated with wake-sleep states have led many to propose objective criteria to classify such states (Rechtschaffen and Kales 1968; Datta and Hobson 2000). Thus, the sleep research community has adopted polysomnographic criteria to identify wake-sleep states based on the activity patterns of EEG/LFP, electromyograms (EMGs), and electro-oculogram (EOG), in addition to behavioral criteria. EMG activity is maximal during WK, when the animal actively explores the environment and decreases as the animal enters SWS. During REM, the EMG activity is absent and the animal becomes atonic. EOG activity, on the other hand, is lowest during SWS and becomes prominent during REM. In rodents (Timo-Iaria et al. 1970; Datta and Hobson 2000), REM sleep is characterized by occasional whisker movements instead of eye movements, as well as particular sleep postures resulting from atonia. Although wake-sleep states have traditionally been coded by trained sleep researchers or clinicians based on polysomnographic criteria, many automatic state identification algorithms have been developed to provide objective and consistent coding of wake-sleep states, based on quantitative measures of EEG and EMG features (Robert et al. 1999). Despite these efforts to identify wake-sleep states based on physiological features, most state-coding algorithms, both manual and automatic, face several important



limitations. First, most algorithms implicitly assume that the wake–sleep cycle consists of several categorically different and predefined stable states. This approach tends to characterize the wake–sleep cycle as a stair case process, jumping back and forth between a set of states (Figure 8.1D, lower panel). Yet, this assumption is unlikely to be true when different stages of SWS (e.g., stage 1 through 4 non-REM sleep) appear as different as two categorically different states (e.g., WK versus SWS) in the stair case representation. Second, the stair case representation of states further promotes the unrealistic view that state transitions occur instantaneously, with no intermediate periods between them, even when the dynamics of the system does not clearly resemble any predefined states. Third, different state-coding algorithms have used different physiological measures and criteria to define states, making comparisons across algorithms difficult. To make matters worse, the number of states identified in different algorithms ranges from two to as many as eight (Kleinlogel 1990; Gottesmann 1992). Fourth, because wake-sleep states result from broad changes in neuronal activity, ultimately, distinct internal brain states should be differentiated based on neuronal signals alone, without the need to rely on EMG signals. Yet, few algorithms have succeeded in that regard (Robert et al. 1996). These limitations and inconsistencies among algorithms signify the lack of a coherent framework to quantitatively describe the forebrain dynamics in different wake-sleep states.

STATE-SPACE FRAMEWORK REVEALS GLOBAL BRAIN STATES Five years ago, we set out to develop a novel framework to quantitatively describe the type of global forebrain dynamics that defines distinct awake-sleep states. This novel approach was based primarily on analyzing the spectral information of LFPs. To avoid the limitations of the state-coding algorithms originating from predefining a set of categorically different states, we took the alternative approach. This involved representing the LFP spectral information with continuous-scale spectral features. Our assumption at the time was that if appropriate spectral features were chosen, one would be able to not only reveal global dynamic states generated by the forebrain, but also provide highly quantitative descriptions of the transition processes involved in switching from one state to another. Furthermore, because behavioral states are known to be associated with different forebrain dynamics, the spectral features we chose should also be able to separate the major behavioral states generated by our subjects (rats and mice). Such a framework could further address questions such as: how many distinct brain states really exist? Are these dynamic states different categorically or part of a continuous spectrum? Do state transitions follow stereotypical sequences? In the next section we describe the approach that allowed us to implement the novel state-space framework required to address these broad issues (Gervasoni et al. 2004).

DATA COLLECTION Initially, LFP signals were collected in adult rats through chronically implanted microelectrode arrays in several forebrain regions, including the primary


150


somatosensory “barrel” cortex, hippocampus, dorsal striatum, and the ventral posterior medial nucleus of the thalamus (Gervasoni et al. 2004). Rats were habituated to the recording chamber prior to the recording sessions. Brain activity was then recorded continuously for up to 120 h. LFPs were preamplified (500X), filtered (0.3–400 Hz), and digitized at 500 Hz using a digital acquisition card (National Instrument, Austin) and a Multi-Neuron Acquisition Processor (Plexon Inc., Dallas). Behaviors were videotaped and synchronized with LFP recordings. Wake-sleep states were coded by two trained experimenters on the basis of commonly adopted polysomnographic criteria (behaviors and associated LFP activity) (Timo-Iaria et al. 1970; Kleinlogel 1990; Gottesmann 1992; Maloney et al. 1997; Fanselow and Nicolelis 1999). This coding, referred to as behavioral state coding, was used later for comparison with results obtained from our novel state-space method. Five behavioral states were coded (Figure 8.1C): Active exploration (AE): Animal engaged in exploratory behavior (locomotion, whisking, sniffing), with low-amplitude cortical LFP, and high theta (5–9 Hz) and gamma (30–55 Hz) power. Quiet waking (QW): Animal immobile (standing or sitting quietly) or engaged in “automatic” stereotyped behaviors (eating, drinking, grooming), with low-amplitude cortical LFP and relatively high theta and gamma activity but less than during AE. Whisker twitching (WT): Animal immobile and standing, with rhythmic whisker movements (twitching) at the same frequency of underlying corticalthalamic oscillations (7–12 Hz) (Fanselow and Nicolelis 1999). WT is a physiological state that occurs with variable prevalence in rats, ranging from 0 to 9% over a 24 h recording period (Nicolelis et al. 1995; Fanselow and Nicolelis 1999; Wiest and Nicolelis 2003; Gervasoni et al. 2004). Slow-wave sleep (SWS): Animal lying immobile with eyes closed and slow regular respiratory movements. SWS begins with sleep spindles (10–14 Hz) superimposed to delta waves (1–4 Hz). As SWS deepens, delta oscillations become predominant, although isolated spindles can still be observed. Rapid-eye-movement (REM): Animal immobile and atonic except for intermittent whisker and ear twitches, with low cortical LFP amplitude and very high theta and gamma power. Epochs containing spindles associated with hippocampal theta rhythm (intermediate sleep) were at that point scored as part of REM episodes (Gottesmann 1973; Mandile et al. 1996).

2-D STATE SPACE To reveal the stable neuronal dynamic regimes of forebrain networks, we focused on investigating the spectral content of forebrain LFP activity. LFPs were represented in the frequency domain as power spectrogram (Figure 8.1D), with the amplitude of each frequency bin calculated every second of the data record. Consistent with our understanding of spectral patterns in different wake-sleep states, the spectrogram showed more theta and gamma oscillation power during WK and REM, and more delta and spindle oscillations during SWS.



The grouping of frequency bins into frequency bands indicated that different frequency bins were not independent of each other. The implication is that if covarying frequency bins were appropriately combined, the spectrogram could be represented with much fewer variables, while still preserving most of the information. In the literature of statistical pattern recognition (Webb 2002), this is equivalent to a dimension reduction problem, i.e., representing a high-dimensional data set (in our case, many frequency bins) with only a few variables (spectral features). One commonly used dimension reduction technique is principal component analysis (PCA), which searches for linear combinations of original variables in the high-dimensional data set, called principal components (PC), that explain the maximal amount of variability of the original data. Successive PCs contain orthogonal combinations of original variables, and are sorted by the amount of variability each PC explains. Thus, most of the information of the high-dimensional data set could be represented by the first few PCs that account for most of the variability. When PCA was applied to the spectrogram, the first two PCs (Figure 8.2A, left panel) typically explained 70–80% of the total variability of the LFP spectrogram.* The spectral combinations of the first two PCs were similar in shape when calculated from different forebrain LFPs. A different dimension reduction strategy, such as linear discriminant analysis, relies on the knowledge of underlying categories and searches for features that best differentiate these classes. In our case, we were interested in finding spectral features that at least can separate the three major wake-sleep states. Assuming multidimensional normal distribution of spectral amplitudes within each class, the discriminant features correspond to the difference of the average spectrum between wake-sleep states. As illustrated in Figure 8.2A (right panel), the main spectral difference between SWS and WK states is in the slow frequency oscillations (25 kg/m 2) adults, 30% of these being considered clinically obese (BMI > 30 kg/m2). In the United States alone, a staggering 30% of all adults are obese (Stein and Colditz 2004). Increase in adiposity leads to significant metabolic disregulation (Muoio and Newgard 2006), with important health and economic consequences (Stein and Colditz 2004). Understanding the central mechanisms of food reward and appetite regulation is therefore an important objective in neuroscience research, hopefully allowing a deeper comprehension of obesity and other eating disorders such as anorexia nervosa and bulimia nervosa.

METHODS FOR MEASURING PREFERENCE IN FREELY LICKING RODENTS Neural recordings have been performed from animals engaged in brief access tests, conducted as described in the available literature (Glendinning, Gresack et al. 2002). Briefly, the animal has access to only one sipper, to which it is given intermittent access in sequential trials. Animals start each trial voluntarily, the structure of which is as follows. Following the animal’s first lick to a solution, the sipper delivers one tastant aliquot for each detected lick response for a brief period (5 s), after which access to the sipper is blocked for 7 s. After this intertrial period, animals are allowed to initiate a new trial. A computer-controlled multiple-valve system (ALA Scientific, Westbury, New York) will allow for water and several different solutions to be presented randomly within blocks of as many different stimuli being tested, one tastant per trial. The cumulative number of licks for all trials of each tastant is recorded and used to calculate the respective lick ratio, i.e., the amount of that tastant consumed with respect to water: LickRatio = n(tastant)/n(water), where n(.) denotes the total


206

Methods for Neural Ensemble Recordings, Second Edition Licking

A Sucrose 1 2 Neurons

3 4 5 6 7 8 –0.5

0.0

0.5

1.0

Time (s) Licking

B Water 1 2 Neurons

3 4 5 6 7 8 –0.5

0.0

0.5

Time (s)

Correct Discrimination (%)

C 66 64

Anticipation

52 50 –0.5

Licking

* a

62 60 58 56 54

1.0

* *

d

d b

–0.25

0.0 Time (s)

0.25

0.5

FIGURE 10.8 (See color insert following page 140.) Ensemble activity of OFC neurons discriminates and anticipates natural rewards in the rat. A and B. Color-coded peristimulus time histograms of activity from eight simultaneously recorded orbitofrontal cortex (OFC) neurons, showing responses to sucrose (A) and water (B). Zero on the x-axis denotes the time of stimulus delivery. Red colors represent maximal activity and blue the minimum activity of each single unit. Although similarities can be observed in the activation pattern of this neuronal ensemble during intake of water and sucrose, differences are also evident, suggesting that OFC neuronal ensembles discriminate between gustatory stimuli. C. Graph showing average correct discrimination between water and sucrose by 16 recorded neuronal ensembles. Discrimination was quantified at four distinct periods, two before (Continued on next page.)


Neural Ensemble Recordings from Central Gustatory-Reward Pathways

207

number of licks for a given stimulus during a session. These values are compared against 1.0, which is the reference value meaning indifference with respect to water. We also use conventional two-bottle preference tests, usually to compare consumption of a given tastant against water using the preference ratio: PreferenceRatio = n(tastant)/(n(tastant) + n(water)). This value is compared against 0.5, which is the reference value meaning indifference with respect to water. When measuring preference in two-bottle tests, side bias must be taken into account. Another test, where the animal has access to a single tastant through a single sipper (one-bottle forced-choice test) measures acceptance, i.e., total consumption in number of licks, and allows for comparisons between different tastants presented across several sessions (Costa, Gutierrez et al. 2006) or the same tastant presented in different conditions across multiple sessions. Two-bottle preference and one-bottle forced-choice tests have the advantage of allowing the manifestation of associations between taste and postingestive effects, giving the experimenter an opportunity to investigate them. The possibility of such associations occurring is reduced in brief access paradigms given that limited amounts of each tastant are delivered through the same sipper in multiple and sequential trials of all tastants. However, these tests reduce the number of tastants that may be presented in the same session and, in the case of the two-bottle test, increase the potential for ambiguous somatosensory effects.

NEURAL ENSEMBLE RECORDINGS IN TASTE-REWARD PATHWAYS When a rodent licks for a natural reward, such as water or a palatable and nutritive solution, the licking activity is not continuous, resulting in a temporal distribution of licks into clusters (at least three rhythmic licks with interlick interval 30 Hz), which are not easily detectable by the EEGs. It was reported that ECoG-based BCIs are more accurate than the EEGbased systems, and the subjects train faster (Leuthardt, Schalk et al. 2004). EEG-based BCIs have been tested in both normal subjects and patients. These tests demonstrated that such systems work in providing simple communication channels, but at the same time exposed their limitations in more complex operations. Normal subjects and neurologically disabled patients were able to control computer cursors using the EEG-based BCIs, and in some cases they enacted movements in artificial devices. The cursors were often used to indicate the users’ selections, for example, selections of letters in a spelling device. A spelling device based on slow cortical potentials was the first successful application of this type (Birbaumer, Ghanayim et al. 1999). BCIs that use slow potentials as their input continue to be developed (Hinterberger, Kubler et al. 2003). In addition to slow potentials, BCIs based on M and B rhythms were tested in severely paralyzed patients (Kubler, Nijboer et al. 2005). In one successful study, a tetraplegic patient gained the ability to grasp objects using a motor-imagery BCI that recognized B waves in his sensorimotor cortex and put in action a functional electrical stimulator connected to his paralyzed hand (Pfurtscheller, Muller et al. 2003). In another clinical trial, a partially paralyzed patient controlled a motor-imagery-based BCI (Kubler, Nijboer et al. 2005) that was


Building Brain–Machine Interfaces to Restore Neurological Functions

223

coupled to an implanted neuroprosthesis system (Freehand ©; Keith, Peckham et al. 1989). P300-based BCIs were implemented in tetraplegic patients (Piccione, Giorgi et al. 2006) and in patients with amyotrophic lateral sclerosis (Sellers and Donchin 2006). These experiments demonstrated that patients with communication and motor deficits can restore some of the lost functions using EEG-based systems.

INVASIVE BMIS Invasive BMIs have been tested mostly in laboratory animals. Only a few studies in human subjects have been conducted (Patil, Carmena et al. 2004; Hochberg, Serruya et al. 2006). This approach is based on the recordings of high-quality signals from the ensembles of single neurons. Pioneering studies on biofeedback based on singlecell activity were conducted by Fetz et al. in the 60s and 70s (Fetz 1969; Fetz and Finocchio 1971; Fetz and Finocchio 1972; Fetz and Baker 1973; Fetz and Finocchio 1975; Fetz 1992). They showed that monkeys can learn to voluntarily control the activity of their cortical neurons if they are provided with a biofeedback that indicates the level of neuronal activity. A few years after these studies, Edward Schmidt proposed that cortical neural activity can be used as a source of voluntary motor commands to a prosthetic device for restoration of motor functions in patients with paralysis (Schmidt 1980). Experimental testing of Schmidt’s proposal took almost 20 years to come true because of the technical difficulties associated with the complex task of recording from large ensembles of cortical cells and analyzing these vast amounts of information in real time. Solving these problems was possible because of a series of breakthroughs in technology and experimental approaches that led to an advanced methodology for chronic, multisite, multielectrode recordings from large populations of single neurons (Nicolelis, Baccala et al. 1995; Nicolelis, Ghazanfar et al. 1997; Nicolelis 2001; Nicolelis and Ribeiro 2002; Nicolelis, Dimitrov et al. 2003). In the first experimental demonstration of a BMI, neuronal population activity recorded in multiple brain areas of behaving rats controlled the one-dimensional movements of a robotic device (Chapin, Moxon et al. 1999). Importantly, it was shown that the rats could control the robot without performing any overt movements with their body parts. This first demonstration triggered a large number of follow-up studies. BMI approaches based on single-neuron recordings were soon implemented in primates: first in New World (Wessberg, Stambaugh et al. 2000), and then in rhesus monkeys (Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). As a result of these initial experiments, several laboratories developed BMIs that reproduced arm-reaching trajectories (Chapin, Moxon et al. 1999; Wessberg, Stambaugh et al. 2000; Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005) and the combination of reaching and grasping movements (Carmena, Lebedev et al. 2003), using either computer cursors or robotic manipulators as actuators. Up to date, the majority of the studies on the invasive BMIs have been conducted in rhesus monkeys. The proposed BMIs can be distinguished by several features: the number of electrode arrays implanted in monkey cortex; cortical sites implanted;


224


characteristics of neural signal recorded, and the size of the neuronal population sampled. The approach of the Duke University team is based on multisite recordings from large neuronal ensembles. Other groups suggested a single cortical site recordings from small neuronal samples (less than 30 neurons) (Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Tillery and Taylor 2004). Some of these groups have focused on the recordings from the primary motor cortex (Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002), whereas one group has chosen posterior parietal cortex as the primary source for inputs for their decoding system (Musallam, Corneil et al. 2004). The design of the Duke University BMI, based on long-term recordings from large populations of neurons (100–400 units), has its origin in the early studies of our laboratory that were carried out in 1995. In these experiments, chronic, multisite, multielectrode recordings in freely behaving rats were pioneered, which allowed extracellular recordings of all major processing levels of the somatosensory system (Nicolelis, Baccala et al. 1995). A series of studies followed this demonstration in which ensemble encoding of tactile stimuli in the somatosensory system of the rat was uncovered using computational algorithms for pattern recognition, such as artificial neural networks (Ghazanfar, Stambaugh et al. 2000; Krupa, Wiest et al. 2004). This design was then moved to BMI experiments in rats (Chapin, Moxon et al. 1999) and monkeys (Wessberg, Stambaugh et al. 2000; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). Currently, we are investigating the clinical relevance of this approach. We have already completed one clinical study in which multielectrode recordings were conducted during the intraoperative placement of deep brain stimulators in Parkinsonian patients (Patil, Carmena et al. 2004). We found that the decoding algorithms as employed in our previous studies of the BMIs in monkeys can successfully extract task-related signals from neural activity recorded in the subthalamic nucleus and the thalamus.

PRINCIPLES OF BMI OPERATION BMIs that are based on neuronal ensemble recordings convert the activity of many cortical neurons into output signals that are useful for controlling artificial actuators. Such decoding makes use of the property of individual neurons located in motor areas to modulate their activity in relationship to movements. In addition, task-related modulations in other cortical and subcortical areas can be used to control BMIs. Task-related modulations of the firing rate of single neurons in monkey motor cortex were first described by Evarts and colleagues four decades ago (Evarts 1966; Evarts 1968a; Evarts 1968b). These initial studies and the studies that followed discovered that firing modulation in single cells were variable from trial to trial (Cohen and Nicolelis 2004; Wessberg and Nicolelis 2004; Carmena, Lebedev et al. 2005; Stein, Gossen et al. 2005). The modulation of a single neuron firing rate can change substantially from one trial to the next even if the overt movements remain virtually identical. However, fairly consistent firing patterns are revealed by averaging the activity of large populations of single neurons across many trials. Discharges of individual neurons fluctuate around these average values. The main idea of the



225

BMI based on large neuronal ensembles is that averaging across large populations of neurons significantly reduces the variability of the population signal (Wessberg, Stambaugh et al. 2000; Carmena, Lebedev et al. 2005). Processing of neural signals by the BMIs can be divided into several stages. First, high-quality signals emanating from many neurons have to be recorded by an implant placed in the brain area of interest. Second, motor control signals have to be extracted from the recorded firing patterns of neuronal populations. Third, the extracted control signals have to enact motor behaviors in an artificial actuator, for example, a prosthetic device. Finally, the information about the actuator performance has to be delivered back to the subjects in some sort of feedback so that errors in performance could be corrected. The subject’s brain is an important component of these control loops not only because it provides the signals that control the actuator movements and analyzes the feedback information but also because its neural circuits can undergo plastic changes to improve the performance. An invasive BMI based on the decoding of neuronal ensemble activity can be accepted by patients only if it is capable of reliable performance. It is also important that the prosthetic devices aided by BMIs feel and act as the subjects’ own limbs. Our recent results suggest that advanced neuroprostheses of the limbs can be incorporated in the internal representation of the body maintained by the brain by creating conditions under which the brain can undergoes experience-dependent plasticity. In the majority of experiments, plastic adaptation of the brain was achieved with the help of visual feedback. In our opinion, an even more efficient way to incorporate the prosthesis in the brain representation of the body could be to use multiple feedback signals derived from pressure and position sensors placed on the prosthetic limb and delivered back to the brain through the somatosensory modality. Such feedback signals are expected to create a realistic perception of the ownership of the prosthetic limb. Moreover, we predict that somatosensory feedback can be delivered to the brain using microstimulation of cortical somatosensory areas. Long-term operation of a sensorized prosthesis will then evoke cortical plasticity resulting in the dedication of sensorimotor and associative areas of the brain to representing the prosthetic device as if it were a natural body appendage.

PROBLEMS FACED BY INVASIVE BMIS In order to fulfill the ambitious goal of creating a clinically useful invasive BMI for restoring upper limb mobility, the BMI research has to resolve several key bottlenecks. First, the stability of neuronal recordings has to be improved. In the perspective, BMIs intended for clinical applications have to allow recordings of large populations of neurons for many years without deterioration of recording quality. The number of recorded neurons has to be in the thousands. To achieve this task, a new generation of biocompatible 3-D electrode matrices will have to be developed. Such electrode arrays will have thousands of recording channels while producing minimal inflammatory reaction and brain tissue damage. Second, next-generation computational algorithms have to be developed. Such algorithms will flexibly reconstruct the activity of large neuronal populations under a range of conditions and


226


allow stable and accurate performance of the neuroprostheses. One challenging task is the development of computational algorithms for controlling multiple-degreeof-freedom artificial actuators. Third, the property of the brain to undergo plastic changes has to be fully utilized in the design of prosthetic devices. Fourth, a new generation of upper limb prosthetics will have to be engineered that will allow light and convenient-to-use prosthetic limbs to operate under different load conditions.

IMPLANTABLE DEVICES FOR RECORDINGS FROM LARGE NEURONAL ENSEMBLES Experiments in the monkeys showed that high-yield recordings from large populations of neurons can be achieved using implantable microwire arrays. Such implants are characterized by excellent recording quality from several months to several years, depending on the monkey species involved in the study. Traditionally, neurophysiologists prefer to record from single neurons. However, our analyses showed that multiunit signals that comprise activity of several neurons can provide a highly efficient information channel as well (Carmena, Lebedev et al. 2003). In addition to neuronal activity, local field potentials (LFPs) can be used as the source of neural information (Pesaran, Pezaris et al. 2002; Mehring, Rickert et al. 2003; Rickert, Oliveira et al. 2005; Scherberger, Jarvis et al. 2005). We expect that advanced neuroprosthetic devices will record and process different types of signals. Here, we discuss the issues related to using single- and multiunit signals as the BMI input. One fundamental question in the BMI design concerns the number of neurons that can efficiently control a neuroprosthetic device for restoration of motor control. Different opinions on this matter have been expressed. Some groups (Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Tillery and Taylor 2004) have strongly claimed that recordings from a small number of neurons can be sufficient for good performance of a BMI. In their experiments, rhesus monkeys controlled relatively simple movements of computer cursors. Therefore, these results cannot be generalized to include more complex movements, especially those approaching the multiple-degree-of-freedom functionality of the human arm. In addition, the claim that small neuronal samples are sufficient for good BMI performance are evidently related to technical difficulties of recordings from large populations of neurons. In our offline analysis of large populations of neurons recorded in the cortex of rhesus monkeys that operated a BMI for reaching and grasping movements, selected populations of highly tuned neurons could predict movement parameters (Sanchez, Carmena et al. 2004). However, those were the neurons specifically selected to maximize the signal-to-noise ratio (SNR) of the population. In a typical experimental situation, highly tuned neurons constitute only a small proportion in a random sample of cortical cells detected by the implanted electrodes. Thus, it would be unrealistic to expect that, in a neuronal population recorded by a small number of implanted electrodes, a large fraction of cells could be highly tuned to a particular motor parameter. It would be even more unrealistic to expect that small numbers of neurons would be useful for decoding several parameters simultaneously. From these basic considerations we conclude that large neuronal samples are always preferable. Large samples of brain cells are needed, at the very least, to produce a neuronal pool from which



227

sufficient number of highly tuned neurons can be drawn. In addition to this obvious point, there are additional reasons for relying on large neuronal populations. One advantage that the large samples of neurons have compared to small samples is their high reliability, which improves considerably with the number of simultaneously recorded neurons (Carmena, Lebedev et al. 2005). The predictions of behavioral variables obtained from small ensembles suffer from instability and drifts, whereas in the large ensembles this problem is minimized. A less understood issue is the adaptation of cortical neurons during a long-term utilization of a BMI. Depending on the decoding algorithms, selected subsets of recorded neurons can exhibit specific adaptations. In our opinion, using large neuronal samples in BMI control provides excellent opportunity to detect the neurons with particular plastic properties and selectively adapt the prediction models to accommodate these neurons. In addition, we have frequently observed that the “best neuron” of today’s session might not be as good for BMI control the day after. Because of this effect, the BMI needs to have access to as many neurons as possible. We conclude that it is highly unlikely that a small group of neurons will be sufficient to control advanced neuroprostheses developed for human use. One important challenge that the BMI faces is the problem of biocompatibility (Schultz and Willey 1976; Dodson, Chu et al. 1978; Landis 1994; Berry M 1999; Tresco, Biran et al. 2000; Polikov, Tresco et al. 2005). Currently, chronically implanted microwire arrays allow good quality of recordings in experimental animals (Wessberg, Stambaugh et al. 2000; Nicolelis 2001; Carmena, Lebedev et al. 2003; Nicolelis 2003; Nicolelis, Dimitrov et al. 2003; Lebedev, Carmena et al. 2005). Whereas a typical monkey experiment runs for several months (in some cases year), BMIs intended for human use have to operate reliably for many years. First and foremost, the broad and challenging issue of biological compatibility has to be properly addressed and solved. In rhesus monkey experiments, recording quality often deteriorates, which likely happens because of electrode encapsulation by fibrous tissue and cell death in the vicinity of the electrode (Polikov, Tresco et al. 2005). Several measures have been suggested to cope with encapsulation. One way to improve biocompatibility is using cone electrodes that contain neurotrophic medium (Kennedy 1989; Kennedy, Mirra et al. 1992; Kennedy and Bakay 1998; Kennedy, Bakay et al. 2000). In addition, coating of the electrodes with biochemical factors that promote neuronal growth (nerve growth factor (NGF), brain-derived neurotrophic (BDNF), laminin) and various antiinflammatory compounds (e.g., dexamethazone) (Ignatius, Sawhney et al. 1998; Cui, Lee et al. 2001; Rahimi and Juliano 2001; Kam, Shain et al. 2002; Biran, Noble et al. 2003; Cui, Wiler et al. 2003; Polikov, Tresco et al. 2005) have been suggested. These promising approaches will need much testing and development before the optimal biocompatible electrode is designed. We expect that, in parallel with the efforts to resolve the biocompatibility issues, new 3-D electrode matrices will be developed. These advanced recording systems will increase the number of simultaneously recorded neurons from hundreds to thousands. Human safety is the primary issue that should be taken into account in these developments. Traditional arrays of rigid, single-ended electrodes may not be adequate for clinical applications for several reasons. Such arrays can sample neuronal activity from flat surfaces of the cortex, but are not well suited for the recordings


228


from cortical sulci. In addition, their reliability is jeopardized by the routine that involves plugging and unplugging of external headstages and the use of cables. Such operations are common in monkey experiments. However, they carry a significant risk of damaging the implant, which is unacceptable for clinical applications. We envision a neuroprosthetic of the future as a fully implantable electronic device for amplification of a large number of neuronal signals equipped with a wireless link that subserves its communication with the computing devices and external actuators. This goal is the major technological challenge that will determine the success of invasive BMIs in clinical applications intended for restoration of neurological functions in patients. The development of telemetry (Claude, Knutti et al. 1979; Knutti, Allen et al. 1979; Mackay 1998; Chien and Jaw 2005; Mohseni, Najafi et al. 2005) intended for multichannel transmission (Bossetti, Carmena et al. 2004; Morizio et al. 2005) is underway and will need thorough testing in animal experiments followed by clinical trials in humans. In addition to traditional approaches, many novel ideas of how to improve recording techniques have been proposed. These ideas consider ceramic-based multielectrode arrays (Moxon, Leiser et al. 2004) and even nanotechnology probes that access the brain through the vascular system (Llinas, Walton et al. 2005). Much research will be needed to select the viable ones from these proposals.

ALGORITHMS FOR PROCESSING NEURONAL ACTIVITY The neuronal mechanisms through which motor and cognitive information is processed in the mammalian brain are far from being completely understood. Rate encoding, temporal encoding, and population encoding principles have been suggested in the literature as ways for information to be represented by neural circuits. Many of these ideas continue to be tested in neurophysiological experiments. Overall, neurophysiological results provide a wealth of information for BMI implementations. Thorough knowledge of the computations carried out by populations of neurons, however, is not totally critical to the design of clinically useful BMIs. That is because BMI systems typically take advantage of a correlation between the discharges of cortical neurons and motor parameters of interest. Thus, BMI computational algorithms perform a reverse operation: they predict motor parameters from the patterns of neuronal firing. Although motor parameters are being predicted, this does not necessarily mean that the neurons from which such predictions are derived normally have a causal relation with the generation of movements. Indeed, such a relation can be quite complex. However, chronic utilization of BMIs in animals seem to be capable of inducing a causal relationship between the neuronal firing and the actuator movements. This new relationship emerges as a result of a process of experience-dependent neuronal plasticity that accompanies BMI operation. One type of correlational relationship between neuronal activity and movement utilized in BMI design is directional tuning (Georgopoulos, Schwartz et al. 1986; Georgopoulos, Kettner et al. 1988). In addition, BMI algorithms have also taken advantage of more general spatiotemporal correlation of neuronal activity with kinematic parameters (Ashe and Georgopoulos 1994; Moran and Schwartz 1999; Averbeck, Chafee et al. 2005) and kinetic (Sergio and Kalaska 1998; Todorov 2000; Sergio, Hamel-Paquet et al. 2005).



229

Since the first experimental demonstration in rats, a great number of linear and nonlinear algorithms for translating neuronal activity into commands to move artificial actuators have been suggested (Chapin, Moxon et al. 1999; Wessberg, Stambaugh et al. 2000; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Kim, Sanchez et al. 2003; Brockwell, Rojas et al. 2004; Brown, Kass et al. 2004; Kemere, Shenoy et al. 2004; Wessberg and Nicolelis 2004; Wu, Black et al. 2004; Hu, Si et al. 2005; Truccolo, Eden et al. 2005). Interestingly, rather simple multiple linear regression models appear to be very efficient in many BMI designs and often outperform more complicated methods (Wessberg, Stambaugh et al. 2000; Serruya, Hatsopoulos et al. 2002; Carmena, Lebedev et al. 2003; Patil, Carmena et al. 2004; Lebedev, Carmena et al. 2005; Santucci, Kralik et al. 2005). A linear model represents predicted parameters (limb position and velocity, muscle force, etc.) as weighted sums of neuronal rates. Neuronal rates are measured at different time points in the past (typically, up to 1s in the past). The number of input variables in a linear model and the size of the time window used for predictions can be optimized for each parameter (Wessberg and Nicolelis 2004; Santucci, Kralik et al. 2005). To cope with overfitting, it is often advantageous to remove noisy inputs from the model and use only the neurons that are highly correlated with the parameter of interest (or task-related, as neurophysiologists usually say) (Sanchez, Carmena et al. 2004). Linear methods can continuously adapt as long as the subjects train with the BMI (Taylor, Tillery et al. 2002). To simultaneously predict multiple motor parameters, several linear models are set to run in parallel. Using this algorithm, we predicted arm position and velocity and hand-gripping force while the monkey operated a BMI that controlled reaching and grasping movements in a robotic manipulator (Carmena, Lebedev et al. 2003). Importantly, the predictions of multiple parameters were obtained from the same population of neurons. This finding supports the theory of parallel, distributed processing that postulates simultaneous processing of multiple variables by the overlapping neuronal ensembles. In such processing scheme, one population of neurons can predict several parameters, and a single cortical neuron contributes to several predictions. Using large-scale recordings from neuronal ensembles located in multiple cortical areas, a wide range of information can be extracted from neural circuits—both the information that the subject is consciously aware of and the information processed implicitly and subconsciously. As such, we expect that, in the future, neuron ensemble-based BMIs will by far outperform the ones based on EEG recordings. The choice of parameters extracted in each future clinical implementation of BMIs will depend on the rehabilitation or therapeutic goals of these applications. For example, our experimental BMI for reaching and grasping (Carmena, Lebedev et al. 2003) can be used as a prototype of human neural prosthesis intended for restoration of the basic capacity of reaching and grasping in paralyzed patients. Such devices can be helpful for quadriplegic or “locked-in” patients to perform basic exploration and manipulation of objects in their surrounding space. However, the tasks of future neuroprostheses will not be limited to limb motor control. We envision the development of BMIs that one day could synthesize speech in aphasic patients using neuronal signals recorded in intact regions related to speech generation. Moreover, BMIs could conceivably be used to restore communication between parts of the central and peripheral neural system affected by neurological diseases.


230


One promising system is the BMI that predicts EMG-like signals (Santucci, Kralik et al. 2005). This design has the benefit of being able to control biologically inspired devices. Such devices can produce a whole range of actuator stiffness. Stiffness of a robotic apparatus is an important property needed for a future generation of practical prostheses of the limbs intended for manipulation of very different objects. For instance, such a device will be able to throw a ball and use a pen to write a signature—very different mechanical operations that require specific control modes. Another intriguing application for BMIs that decode muscle activation-like signals is implementing direct stimulation of muscles that paralyzed patients cannot activate voluntarily (Degnan, Wind et al. 2002; Navarro, Krueger et al. 2005; Peckham and Knutson 2005). Such BMIs are likely to be of great benefit to patients, especially because they can be entirely encased in the patient’s body. In addition to decoding motor parameters (Chapin, Moxon et al. 1999; Wessberg, Stambaugh et al. 2000; Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Santucci, Kralik et al. 2005), decoding cognitive signals and using them to control BMIs is a promising idea for future developments. BMIs that decode intended reach direction during the delay periods preceding movement execution have been proposed (Shenoy, Meeker et al. 2003; Hatsopoulos, Joshi et al. 2004; Musallam, Corneil et al. 2004; Rizzuto, Mamelak et al. 2005). It should be noted that BMIs based exclusively on cognitive signals cannot execute continuous control of movement parameters. Instead, they decode higher-order parameters, such as prepared reach direction or characteristics of objects that potentially can be grasped. For such decoding algorithm to work, higher-order details of motor execution have to be delegated to the actuator controller. Recently, we have examined a mode of operation that we called a shared control mode (Kim, Biggs et al. 2005). Our analyses showed that shared control can improve the accuracy with which the robotic device implements reaching and grasping movements. In an advanced neural prosthesis, a shared-control mode of operation would be achieved by an algorithm that extracts high-order signals from the brain and sends them to a low-level controller that uses artificial “reflex-like” circuits to improve the precision of the prosthetic limb movements. Cognitive variables that advanced BMIs can utilize remain largely unexplored. We predict that new BMI designs will incorporate such higher-order neuronal representations of movements as encoding of movement sequences (Hoshi and Tanji 2004; Lu and Ashe 2005), reference frames (Graziano and Gross 1998; Batista, Buneo et al. 1999; Battaglia-Mayer, Ferraina et al. 2000; Olson 2003), and potential movement targets (Cisek and Kalaska 2002). We expect that BMIs will learn to simultaneously manipulate multiple spatial variables such as movement direction, orientation of selective spatial attention, and gaze angle (Boussaoud, Barth et al. 1993; Lebedev and Wise 2001). One unexplored area is the implementation of BMIs for the encoding of temporal characteristics of movements. Neurophysiological, neuropsychological, and imaging studies (Ivry 1996; Leon and Shadlen 2003; Matell, Meck et al. 2003; Roux, Coulmance et al. 2003) point to a rather distributed representation of temporal information in the brain that includes cortical, thalamic, and striatal circuitry. Recently, we approached this issue by examining neuronal recordings obtained from



231

primary motor and premotor cortex while rhesus monkeys performed self-timed button presses (O’Doherty, Lebedev et al. 2005). Neuronal ensembles accurately predicted both the time that elapsed since the monkey pressed the button and the time until the button was released. Currently, we are adding this type of information into our newest BMI systems. We expect that such BMIs will be able to mimic the temporal structure in which episodes of movement execution are intermingled with periods of immobility.

BRAIN PLASTICITY INVOLVED IN BMI OPERATIONS Ideally, a prosthetic device directly controlled by brain activity should look and feel as the subject’s own limb. In theory, this ambitious goal can be attained thanks to the remarkable ability of the brain to plastically adapt to new motor and sensory requirements. It is currently believed that the brain contains a supramodal representation of the body, which is often termed the “body schema.” Almost 100 years ago, Head and Holmes suggested that the “body schema” could extend itself to include a wielded tool (Head and Holmes 1911). Controlling an artificial actuator through a BMI is a process somewhat similar to the operation required by subjects to manipulate tools—a capacity that endows only higher primates, such as chimpanzees and humans (Breuer, Ndoundou-Hockemba et al. 2005). Brain plasticity that occurs during tool usage was demonstrated by the experiments in which monkeys used a rake to retrieve distant objects (Iriki, Tanaka et al. 1996). As the monkey practiced, cortical neurons extended their visual receptive fields along the length of the rake. Remapping of the “body schema” during tool usage has been also demonstrated in psychophysics experiments in humans (Gurfinkel, Levick et al. 1991; Maravita, Spence et al. 2003). Additionally, in a neuroimaging study (Maruishi, Tanaka et al. 2004) specific activations of the right ventral premotor cortex occurred during manipulation of a myoelectric prosthetic hand. These results strongly suggest that long-term usage of an actuator, which is directly controlled by brain activity, may evoke substantial cortical and subcortical remapping. Likely, during this process of remapping a new vivid perceptual experience will emerge: the subjects will start to perceive the artificial actuator as if it belongs to the subject’s own body. In support of this suggestion, primary sensorimotor cortex activation is observed during perceived voluntary movements of phantom limbs in amputees (Roux, Lotterie et al. 2003). Recently, a series of remarkable demonstrations of tool assimilation was obtained when experimental animals learned to operate an actuator through BMIs. At the peak of their performance, these animals were capable of operating such interfaces without the need of moving their own limbs (Chapin, Moxon et al. 1999; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). Even if decoding algorithms were initially trained to predict overt movements of the animals’ limbs, after animals started to control the actuator through the BMI and stopped moving their own limbs, they still continued to control the actuator through modulations of their neurons’ firing rate. Our analysis showed that, after the mode of operation was switched to direct BMI control, even when the animals still continued to move their own arms, neuronal tuning of the movements of the subject’s


232


own limbs decreased (Lebedev, Carmena et al. 2005). This finding indicated that the neuronal activations became dedicated to controlling the actuator movements. One interpretation of this result is that the brain gradually assimilated the actuator within the same maps that represented the animal’s own body (Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). In agreement with this hypothesis, the several studies of continuous BMI operations in primates have now reported physiological changes in neuronal tuning, which include changes in a neuron’s preferred direction and direction tuning strength (Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). In addition, broad changes in pairwise neuronal correlation occur under brain-control mode (Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). Although neuronal plasticity associated with BMI usage needs to be studied in much more detail, it is clear that it can underlie incorporation of neural prostheses into the internal representation of the body during prolonged BMI operations in human users.

SOMATOSENSORY FEEDBACK IN BMI DESIGN Normally, the sense of ownership of the limb is facilitated by the continuous peripheral tactile and proprioceptive signals that occur when the limb moves and interacts with external objects (Gurfinkel, Levick et al. 1991; Maravita, Spence et al. 2003). Looking into the future, we envision that limb prostheses will be equipped with a variety of sensors providing multiple channels of artificial somatosensory feedback signals, which can better inform the subject’s brain about the continuous functions of these actuators. Such sensory feedback will be different from the one provided in current BMI designs, in which animals receive sensory information from the actuator through visual feedback (Serruya, Hatsopoulos et al. 2002; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Lebedev, Carmena et al. 2005). Until recently, the use of tactile and proprioceptive feedback in BMI research had remained untouched. During the last two years, we have started to investigate the possibility of providing sensory feedback information from a robotic actuator to the brain using multichannel microstimulation of cortical somatosensory areas. The idea of providing sensory feedback using microstimulation is supported by previous studies that demonstrated the ability of monkeys in sensing microstimulation patterns and using this information to guide their behavioral responses (Romo, Hernandez et al. 2000; Cohen and Newsome 2004). In our laboratory, we carried out a long-term study in which owl monkeys learned to guide their reaching movements by decoding vibratory stimuli applied to their arms (Fitzsimmons, Drake et al. 2007; see chapter 7 in this volume). In the following study, instead of vibratory stimulation, patterns of microstimulation pulses were applied through the electrodes implanted in the primary somatosensory cortex (see chapter 3 in this volume). The same monkeys were still able to correctly interpret the instructions provided by cortical microstimulation. Moreover, the monkeys’ performance guided by microstimulation was so good that they eventually surpassed the level of performance observed in the experiments with vibrostimulation applied to the skin. These results suggest that cortical microstimulation and other sensory inputs to the brain may become useful to “sensorize” prosthetic limbs controlled by a BMI.



233

CONCLUSION We envision that future neuroprosthesis will include a fully implantable recording system that wirelessly transmits multiple neuronal ensemble signals to a BMI module that decodes motor commands and cognitive characteristics of the action the subject intends to perform. Such BMI will analyze both high-order commands, derived from the brain activity, and peripheral feedback signals involved in artificial reflex-like control loops. This shared-control mode of operation will ensure high accuracy of movements performed by a multiple-degree-of-freedom prosthetic device or by the subject’s own limbs activated by functional electrical stimulation of peripheral nerves and muscles. Moreover, arrays of touch and position sensors will produce multiple feedback signals that could be delivered to brain somatosensory areas using multichannel microstimulation. We hope that such hybrid BMIs may one day restore motor functions in patients suffering from devastating consequences of multiple neurological disorders.

REFERENCES Ashe, J. and Georgopoulos, A.P. (1994). Movement parameters and neural activity in motor cortex and area 5. Cereb Cortex 4(6): 590–600. Averbeck, B.B. and Chafee, M.V. et al. (2005). Parietal representation of hand velocity in a copy task. J Neurophysiol 93(1): 508–18. Batista, A.P. and Buneo, C.A. et al. (1999). Reach plans in eye-centered coordinates. Science 285(5425): 257–60. Battaglia-Mayer, A. and Ferraina, S. et al. (2000). Early coding of reaching in the parietooccipital cortex. J Neurophysiol 83(4): 2374–91. Bayliss, J.D. and Ballard, D.H. (2000). A virtual reality testbed for brain-computer interface research. IEEE Trans Rehabil Eng 8(2): 188–90. Berry M., Butt. A., and Logan, A. (1999). Cellular responses to penetrating CNS injury. Boca Raton, CRC Press. Biran, R. and Noble, M.D. et al. (2003). Directed nerve outgrowth is enhanced by engineered glial substrates. Exp Neurol 184(1): 141–52. Birbaumer, N. (2006). Brain-computer-interface research: coming of age. Clin Neurophysiol 117(3): 479–83. Birbaumer, N. and Ghanayim, N. et al. (1999). A spelling device for the paralysed. Nature 398(6725): 297–8. Birbaumer, N. and Kubler, A. et al. (2000). The thought translation device (TTD) for completely paralyzed patients. IEEE Trans Rehabil Eng 8(2): 190–3. Bossetti, C.A. and Carmena, J.M. et al. (2004). Transmission latencies in a telemetry-linked brain-machine interface. IEEE Trans Biomed Eng 51(6): 919–24. Boussaoud, D. and Barth, T.M. et al. (1993). Effects of gaze on apparent visual responses of frontal cortex neurons. Exp Brain Res 93(3): 423–34. Breuer, T. and Ndoundou-Hockemba, M. et al. (2005). First observation of tool use in wild gorillas. PLoS Biol 3(11): e380. Brockwell, A.E. and Rojas, A.L. et al. (2004). Recursive Bayesian decoding of motor cortical signals by particle filtering. J Neurophysiol 91(4): 1899–907. Brown, E.N. and Kass, R.E. et al. (2004). Multiple neural spike train data analysis: state-ofthe-art and future challenges. Nat Neurosci 7(5): 456–61. Carmena, J.M. and Lebedev, M.A. et al. (2003). Learning to control a brain-machine interface for reaching and grasping by primates. PLoS Biol 1(2): E42.


234


Carmena, J.M. and Lebedev, M.A. et al. (2005). Stable ensemble performance with single-neuron variability during reaching movements in primates. J Neurosci 25(46): 10712–6. Chapin, J.K. and Moxon, K.A. et al. (1999). Real-time control of a robot arm using simultaneously recorded neurons in the motor cortex. Nat Neurosci 2(7): 664–70. Chien, C.N. and Jaw, F.S. (2005). Miniature telemetry system for the recording of action and field potentials. J Neurosci Methods 147(1): 68–73. Cisek, P. and Kalaska, J.F. (2002). Simultaneous encoding of multiple potential reach directions in dorsal premotor cortex. J Neurophysiol 87(2): 1149–54. Claude, J.P. and Knutti, J.W. et al. (1979). Applications of totally implantable telemetry systems to chronic medical research. Biotelem Patient Monit 6(3): 160–71. Cohen, D. and Nicolelis, M.A. (2004). Reduction of single-neuron firing uncertainty by cortical ensembles during motor skill learning. J Neurosci 24(14): 3574–82. Cohen, M.R. and Newsome, W.T. (2004). What electrical microstimulation has revealed about the neural basis of cognition. Curr Opin Neurobiol 14(2): 169–77. Cui, X. and Lee, V.A. et al. (2001). Surface modification of neural recording electrodes with conducting polymer/biomolecule blends. J Biomed Mater Res 56(2): 261–72. Cui, X. and Wiler, J. et al. (2003). In vivo studies of polypyrrole/peptide coated neural probes. Biomaterials 24(5): 777–87. Degnan, G.G. and Wind, T.C. et al. (2002). Functional electrical stimulation in tetraplegic patients to restore hand function. J Long Term Eff Med Implants 12(3): 175–88. Dodson, R.F. and Chu, L.W. et al. (1978). Cerebral tissue response to electrode implantation. Can J Neurol Sci 5(4): 443–6. Donchin, E. and Spencer, K.M. et al. (2000). The mental prosthesis: assessing the speed of a P300-based brain-computer interface. IEEE Trans Rehabil Eng 8(2): 174–9. Evarts, E.V. (1966). Pyramidal tract activity associated with a conditioned hand movement in the monkey. J Neurophysiol 29(6): 1011–27. Evarts, E.V. (1968a). Relation of pyramidal tract activity to force exerted during voluntary movement. J Neurophysiol 31(1): 14–27. Evarts, E.V. (1968b). A technique for recording activity of subcortical neurons in moving animals. Electroencephalogr Clin Neurophysiol 24(1): 83–6. Fetz, E.E. (1969). Operant conditioning of cortical unit activity. Science 163(870): 955–8. Fetz, E.E. (1992). Are movement parameters recognizably coded in activity of single neurons? Behav Brain Sci 15: 679–690. Fetz, E.E. and Baker, M.A. (1973). Operantly conditioned patterns on precentral unit activity and correlated responses in adjacent cells and contralateral muscles. J Neurophysiol 36(2): 179–204. Fetz, E.E. and Finocchio, D.V. (1971). Operant conditioning of specific patterns of neural and muscular activity. Science 174(7): 431–5. Fetz, E.E. and Finocchio, D.V. (1972). Operant conditioning of isolated activity in specific muscles and precentral cells. Brain Res 40(1): 19–23. Fetz, E.E. and Finocchio, D.V. (1975). Correlations between activity of motor cortex cells and arm muscles during operantly conditioned response patterns. Exp Brain Res 23(3): 217–40. Fitzsimmons, N. and Drake, W. et al. (2007). Primate reaching cued by multichannel spatiotemporal cortical microstimulation. J Neurosci 27(21): 5593–602. Georgopoulos, A.P. and Kettner, R.E. et al. (1988). Primate motor cortex and free arm movements to visual targets in three-dimensional space. II. Coding of the direction of movement by a neuronal population. J Neurosci 8(8): 2928–37. Georgopoulos, A.P. and Schwartz, A.B. et al. (1986). Neuronal population coding of movement direction. Science 233(4771): 1416–9. Ghazanfar, A.A. and Stambaugh, C.R. et al. (2000). Encoding of tactile stimulus location by somatosensory thalamocortical ensembles. J Neurosci 20(10): 3761–75.



235

Graziano, M.S. and Gross, C.G. (1998). Spatial maps for the control of movement. Curr Opin Neurobiol 8(2): 195–201. Gurfinkel, V.S. and Levick, Y.S. et al. (1991). Body scheme concept and motor control. Body scheme in the postural automatisms regulation. Intellectual processes and their modelling. Moscow, Nauka: pp 24–53. Hatsopoulos, N. and Joshi, J. et al. (2004). Decoding continuous and discrete motor behaviors using motor and premotor cortical ensembles. J Neurophysiol 92(2): 1165–74. Head, H. and Holmes, G. (1911). Sensory disturbances from cerebral lesion. Brain 34: 102–254. Hinterberger, T. and Kubler, A. et al. (2003). A brain-computer interface (BCI) for the locked-in: comparison of different EEG classifications for the thought translation device. Clin Neurophysiol 114(3): 416–25. Hinterberger, T. and Veit, R. et al. (2005). Neuronal mechanisms underlying control of a brain-computer interface. Eur J Neurosci 21(11): 3169–81. Hochberg, L.R. and Serruya, M.D. et al. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 442(7099): 164–71. Hoshi, E. and Tanji, J. (2004). Differential roles of neuronal activity in the supplementary and presupplementary motor areas: from information retrieval to motor planning and execution. J Neurophysiol 92(6): 3482–99. Hu, J. and Si, J. et al. (2005). Feature detection in motor cortical spikes by principal component analysis. IEEE Trans Neural Syst Rehabil Eng 13(3): 256–62. Ignatius, M.J. and Sawhney, N. et al. (1998). Bioactive surface coatings for nanoscale instruments: effects on CNS neurons. J Biomed Mater Res 40(2): 264–74. Iriki, A. and Tanaka, M. et al. (1996). Coding of modified body schema during tool use by macaque postcentral neurones. Neuroreport 7(14): 2325–30. Ivry, R.B. (1996). The representation of temporal information in perception and motor control. Curr Opin Neurobiol 6(6): 851–7. Kam, L. and Shain, W. et al. (2002). Selective adhesion of astrocytes to surfaces modified with immobilized peptides. Biomaterials 23(2): 511–5. Keith, M.W. and Peckham, P.H. et al. (1989). Implantable functional neuromuscular stimulation in the tetraplegic hand. J Hand Surg [Am] 14(3): 524–30. Kelly, S.P. and Lalor, E.C. et al. (2005). Visual spatial attention control in an independent brain-computer interface. IEEE Trans Biomed Eng 52(9): 1588–96. Kemere, C. and Shenoy, K.V. et al. (2004). Model-based neural decoding of reaching movements: a maximum likelihood approach. IEEE Trans Biomed Eng 51(6): 925–32. Kennedy, P.R. (1989). The cone electrode: a long-term electrode that records from neurites grown onto its recording surface. J Neurosci Methods 29(3): 181–93. Kennedy, P.R. and Bakay, R.A. (1998). Restoration of neural output from a paralyzed patient by a direct brain connection. Neuroreport 9(8): 1707–11. Kennedy, P.R. and Bakay, R.A. et al. (2000). Direct control of a computer from the human central nervous system. IEEE Trans Rehabil Eng 8(2): 198–202. Kennedy, P.R. and Mirra, S.S. et al. (1992). The cone electrode: ultrastructural studies following long-term recording in rat and monkey cortex. Neurosci Lett 142(1): 89–94. Kim, H.K. and Biggs, S.J. et al. (2005). Continuous shared control stabilizes reach and grasping with brain-machine interfaces. IEEE Trans Biomed Eng 53: 1164–73. Kim, S.P. and Sanchez, J.C. et al. (2003). Divide-and-conquer approach for brain machine interfaces: nonlinear mixture of competitive linear models. Neural Networks 16(5–6): 865–71. Knutti, J.W. and Allen, H.V. et al. (1979). An integrated circuit approach to totally implantable telemetry systems. Biotelem Patient Monit 6(3): 95–106. Krupa, D.J. and Wiest, M.C. et al. (2004). Layer-specific somatosensory cortical activation during active tactile discrimination. Science 304(5679): 1989–92.


236


Kubler, A. and Kotchoubey, B. et al. (2001). Brain-computer communication: unlocking the locked in. Psychol Bull 127(3): 358–75. Kubler, A. and Neumann, N. et al. (2001). Brain-computer communication: self-regulation of slow cortical potentials for verbal communication. Arch Phys Med Rehabil 82(11): 1533–9. Kubler, A. and Nijboer, F. et al. (2005). Patients with ALS can use sensorimotor rhythms to operate a brain-computer interface. Neurology 64(10): 1775–7. Landis, D.M. (1994). The early reactions of non-neuronal cells to brain injury. Annu Rev Neurosci 17: 133–51. Lebedev, M.A. and Carmena, J.M. et al. (2005). Cortical ensemble adaptation to represent velocity of an artificial actuator controlled by a brain-machine interface. J Neurosci 25(19): 4681–93. Lebedev, M.A. and Wise, S.P. (2001). Tuning for the orientation of spatial attention in dorsal premotor cortex. Eur J Neurosci 13(5): 1002–8. Leon, M.I. and Shadlen, M.N. (2003). Representation of time by neurons in the posterior parietal cortex of the macaque. Neuron 38(2): 317–27. Leuthardt, E.C. and Schalk, G. et al. (2004). A brain-computer interface using electrocorticographic signals in humans. J Neural Eng 1(2): 63–71. Llinas, R.R. and Walton, K.D. et al. (2005). Neuro-vascular central nervous recording/ stimulating system: Using nanotechnology probes. J Nanoparticle Res 7: 111–127. Lu, X. and Ashe, J. (2005). Anticipatory activity in primary motor cortex codes memorized movement sequences. Neuron 45(6): 967–73. Mackay, S. (1998). Bio-medical telemetry: Sensing and transmitting biological information from animals and man, Wiley-IEEE Press. Maravita, A. and Spence, C. et al. (2003). Multisensory integration and the body schema: close to hand and within reach. Curr Biol 13(13): R531–9. Maruishi, M. and Tanaka, Y. et al. (2004). Brain activation during manipulation of the myoelectric prosthetic hand: a functional magnetic resonance imaging study. Neuroimage 21(4): 1604–11. Matell, M.S. and Meck, W.H. et al. (2003). Interval timing and the encoding of signal duration by ensembles of cortical and striatal neurons. Behav Neurosci 117(4): 760–73. Mehring, C. and Rickert, J. et al. (2003). Inference of hand movements from local field potentials in monkey motor cortex. Nat Neurosci 6(12): 1253–4. Middendorf, M. and McMillan, G. et al. (2000). Brain-computer interfaces based on the steady-state visual-evoked response. IEEE Trans Rehabil Eng 8(2): 211–4. Mohseni, P. and Najafi, K. et al. (2005). Wireless multichannel biopotential recording using an integrated FM telemetry circuit. IEEE Trans Neural Syst Rehabil Eng 13(3): 263–71. Moran, D.W. and Schwartz, A.B. (1999). Motor cortical representation of speed and direction during reaching. J Neurophysiol 82(5): 2676–92. Morizio, J., Go, V., Bender, J., Makgill, S., Santos, L., Lehew, G., and Nicolelis, M. (2005). Fifteen-channel wireless headstage system for single-unit rat recordings. Society for Neuroscience Annual Meeting, 31: 68–4., Washington D.C. Moxon, K.A. and Leiser, S.C. et al. (2004). Ceramic-based multisite electrode arrays for chronic single-neuron recording. IEEE Trans Biomed Eng 51(4): 647–56. Musallam, S. and Corneil, B.D. et al. (2004). Cognitive control signals for neural prosthetics. Science 305(5681): 258–62. Navarro, X. and Krueger, T.B. et al. (2005). A critical review of interfaces with the peripheral nervous system for the control of neuroprostheses and hybrid bionic systems. J Peripher Nerv Syst 10(3): 229–58. Nicolelis, M.A. (2001). Actions from thoughts. Nature 409(6818): 403–7. Nicolelis, M.A. (2003). Brain-machine interfaces to restore motor function and probe neural circuits. Nat Rev Neurosci 4(5): 417–22.



237

Nicolelis, M.A. and Baccala, L.A. et al. (1995). Sensorimotor encoding by synchronous neural ensemble activity at multiple levels of the somatosensory system. Science 268(5215): 1353–8. Nicolelis, M.A. and Dimitrov, D. et al. (2003). Chronic, multisite, multielectrode recordings in macaque monkeys. Proc Natl Acad Sci U.S.A. 100(19): 11041–6. Nicolelis, M.A. and Ghazanfar, A.A. et al. (1997). Reconstructing the engram: simultaneous, multisite, many single neuron recordings. Neuron 18(4): 529–37. Nicolelis, M.A. and Ribeiro, S. (2002). Multielectrode recordings: the next steps. Curr Opin Neurobiol 12(5): 602–6. Nowlis, D.P. and Kamiya, J. (1970). The control of electroencephalographic alpha rhythms through auditory feedback and the associated mental activity. Psychophysiol 6(4): 476–84. O’Doherty, J. and Lebedev, M. et al. (2005). Ensemble Representation of Time: Interhemispheric Communication Involved? Society for Neuroscience Annual Meeting, 31: 402–5., Washington, D.C. Obermaier, B. and Muller, G.R. et al. (2003). Virtual keyboard controlled by spontaneous EEG activity. IEEE Trans Neural Syst Rehabil Eng 11(4): 422–6. Obermaier, B. and Neuper, C. et al. (2001). Information transfer rate in a five-classes braincomputer interface. IEEE Trans Neural Syst Rehabil Eng 9(3): 283–8. Olson, C.R. (2003). Brain representation of object-centered space in monkeys and humans. Annu Rev Neurosci 26: 331–54. Patil, P.G. and Carmena, J.M. et al. (2004). Ensemble recordings of human subcortical neurons as a source of motor control signals for a brain-machine interface. Neurosurgery 55(1): 27–35. Peckham, P.H. and Knutson, J.S. (2005). Functional electrical stimulation for neuromuscular applications. Annu Rev Biomed Eng 7: 327–60. Pesaran, B. and Pezaris, J.S. et al. (2002). Temporal structure in neuronal activity during working memory in macaque parietal cortex. Nat Neurosci 5(8): 805–11. Pfurtscheller, G. and Brunner, C. et al. (2006). Mu rhythm (de)synchronization and EEG single-trial classification of different motor imagery tasks. Neuroimage 31(1): 153–59. Pfurtscheller, G. and Leeb, R. et al. (2006). Walking from thought. Brain Res 1071(1): 145–52. Pfurtscheller, G. and Lopes da Silva, F.H. (1999). Event-related EEG/MEG synchronization and desynchronization: basic principles. Clin Neurophysiol 110(11): 1842–57. Pfurtscheller, G. and Muller, G.R. et al. (2003). ‘Thought’--control of functional electrical stimulation to restore hand grasp in a patient with tetraplegia. Neurosci Lett 351(1): 33–6. Pfurtscheller, G. and Neuper, C. et al. (2003). Graz-BCI: state of the art and clinical applications. IEEE Trans Neural Syst Rehabil Eng 11(2): 177–80. Piccione, F. and Giorgi, F. et al. (2006). P300-based brain computer interface: reliability and performance in healthy and paralysed participants. Clin Neurophysiol 117(3): 531–7. Polikov, V.S. and Tresco, P.A. et al. (2005). Response of brain tissue to chronically implanted neural electrodes. J Neurosci Methods 148(1): 1–18. Rahimi, O. and Juliano, S.L. (2001). Transplants of NGF-secreting fibroblasts restore stimulus-evoked activity in barrel cortex of basal-forebrain-lesioned rats. J Neurophysiol 86(4): 2081–96. Rickert, J. and Oliveira, S.C. et al. (2005). Encoding of movement direction in different frequency ranges of motor cortical local field potentials. J Neurosci 25(39): 8815–24. Rizzuto, D.S. and Mamelak, A.N. et al. (2005). Spatial selectivity in human ventrolateral prefrontal cortex. Nat Neurosci 8(4): 415–7. Romo, R. and Hernandez, A. et al. (2000). Sensing without touching: psychophysical performance based on cortical microstimulation. Neuron 26(1): 273–8.


238


Roux, F.E. and Lotterie, J.A. et al. (2003). Cortical areas involved in virtual movement of phantom limbs: comparison with normal subjects. Neurosurgery 53(6): 1342–52; discussion 1352–3. Roux, S. and Coulmance, M. et al. (2003). Context-related representation of timing processes in monkey motor cortex. Eur J Neurosci 18(4): 1011–6. Sanchez, J.C. and Carmena, J.M. et al. (2004). Ascertaining the importance of neurons to develop better brain-machine interfaces. IEEE Trans Biomed Eng 51(6): 943–53. Santucci, D.M. and Kralik, J.D. et al. (2005). Frontal and parietal cortical ensembles predict single-trial muscle activity during reaching movements in primates. Eur J Neurosci 22(6): 1529–40. Scherberger, H. and Jarvis, M.R. et al. (2005). Cortical local field potential encodes movement intentions in the posterior parietal cortex. Neuron 46(2): 347–54. Schmidt, E.M. (1980). Single neuron recording from motor cortex as a possible source of signals for control of external devices. Ann Biomed Eng 8(4–6): 339–49. Schultz, R.L. and Willey, T.J. (1976). The ultrastructure of the sheath around chronically implanted electrodes in brain. J Neurocytol 5(6): 621–42. Sellers, E.W. and Donchin, E. (2006). A P300-based brain-computer interface: initial tests by ALS patients. Clin Neurophysiol 117(3): 538–48. Sergio, L.E. and Hamel-Paquet, C. et al. (2005). Motor cortex neural correlates of output kinematics and kinetics during isometric-force and arm-reaching tasks. J Neurophysiol 94(4): 2353–78. Sergio, L.E. and Kalaska, J.F. (1998). Changes in the temporal pattern of primary motor cortex activity in a directional isometric force versus limb movement task. J Neurophysiol 80(3): 1577–83. Serruya, M.D. and Hatsopoulos, N.G. et al. (2002). Instant neural control of a movement signal. Nature 416(6877): 141–2. Sheikh, H. and McFarland, D.J. et al. (2003). Electroencephalographic(EEG)-based communication: EEG control versus system performance in humans. Neurosci Lett 345(2): 89–92. Shenoy, K.V. and Meeker, D. et al. (2003). Neural prosthetic control signals from plan activity. Neuroreport 14(4): 591–6. Stein, R.B. and Gossen, E.R. et al. (2005). Neuronal variability: noise or part of the signal? Nat Rev Neurosci 6(5): 389–97. Sterman, M.B. and Macdonald, L.R. et al. (1974). Biofeedback training of the sensorimotor electroencephalogram rhythm in man: effects on epilepsy. Epilepsia 15(3): 395–416. Sutter, E.E. and Tran, D. (1992). The field topography of ERG components in man--I. The photopic luminance response. Vision Res 32(3): 433–46. Taylor, D.M. and Tillery, S.I. et al. (2002). Direct cortical control of 3D neuroprosthetic devices. Science 296(5574): 1829–32. Tillery, S.I. and Taylor, D.M. (2004). Signal acquisition and analysis for cortical control of neuroprosthetics. Curr Opin Neurobiol 14(6): 758–62. Todorov, E. (2000). Direct cortical control of muscle activation in voluntary arm movements: a model. Nat Neurosci 3(4): 391–8. Tresco, P.A. and Biran, R. et al. (2000). Cellular transplants as sources for therapeutic agents. Adv Drug Deliv Rev 42(1–2): 3–27. Truccolo, W. and Eden, U.T. et al. (2005). A point process framework for relating neural spiking activity to spiking history, neural ensemble, and extrinsic covariate effects. J Neurophysiol 93(2): 1074–89. Wessberg, J. and Nicolelis, M.A. (2004). Optimizing a linear algorithm for real-time robotic control using chronic cortical ensemble recordings in monkeys. J Cogn Neurosci 16(6): 1022–35.



239

Wessberg, J. and Stambaugh, C.R. et al. (2000). Real-time prediction of hand trajectory by ensembles of cortical neurons in primates. Nature 408(6810): 361–5. Wolpaw, J.R. (2004). Brain-computer interfaces (BCIs) for communication and control: a mini-review. Suppl Clin Neurophysiol 57: 607–13. Wolpaw, J.R. and Birbaumer, N. et al. (2002). Brain-computer interfaces for communication and control. Clin Neurophysiol 113(6): 767–91. Wolpaw, J.R. and McFarland, D.J. (2004). Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans. Proc Natl Acad Sci U.S.A 101(51): 17849–54. Wu, W. and Black, M.J. et al. (2004). Modeling and decoding motor cortical activity using a switching Kalman filter. IEEE Trans Biomed Eng 51(6): 933–42. Wyricka, W. and Sterman, M. (1968). Instrumental conditioning of sensorimotor cortex EEG spindles in the waking cat. Psychol Behav 3: 703–707.


12

Conceptual and Technical Approaches to Human Neural Ensemble Recordings Dennis A. Turner, Parag G. Patil, and Miguel A.L. Nicolelis

CONTENTS Introduction............................................................................................................ 242 Relevant Preclinical Studies for Human Neural Ensemble Recordings ................244 Requirements for Human Ensemble Recording and Neuroprosthetics Devices ... 245 Appropriate Human Medical Conditions ................................................... 247 Human Technical Considerations: Engineering Translation ......................248 Preliminary Human Study Experience.................................................................. 249 Microwire Array Experience...................................................................... 249 Cyberkinetics and the Utah Electrode Experience..................................... 250 Cortical Electrocorticogram (ECoG) and Brain Control ........................... 250 Future Human Neural Ensemble Recordings ........................................................ 251 Types of Electrodes..................................................................................... 251 Superficial Cortical Electrodes ........................................................ 251 Deep Electrode Arrays..................................................................... 253 Microwire Recording Electrodes..................................................... 253 Silicon and Other Recording Arrays .......................................................... 253 Electronics .................................................................................................. 253 Intervention and Training Methods ............................................................ 254 Sensory Feedback—Stimulation Electrodes and Multineuron Field Stimulation....................................................................................... 254 Conclusions ............................................................................................................ 255 Acknowledgments.................................................................................................. 255 References.............................................................................................................. 255


242


INTRODUCTION The ability to perform either multineuron or local field/EEG recordings from the nervous system is a critical requirement to develop a new generation of neuroprosthetics that can sense the brain’s intent for action (Nicolelis 2001, 2003). This form of sensing neuroprosthesis builds upon the concept of current neuroprosthetic devices, which are primarily for macrostimulation of neural elements, such as deep brain stimulation (DBS); (Abosch, Hutchison et al. 2002; Rodriguez-Oroz, Obeso et al. 2005). A key aspect of this evolving technology is the translation of preclinical multineuron recording and analysis technology into the clinical arena (Donoghue 2002; Carmena, Lebedev et al. 2003; Mussa-Ivaldi, Miller et al. 2003). This translation requires the use of medical-grade components at all levels of electrodes, connections, and electronics, and the stabilization of technology and software for the long process of Food and Drug Administration (FDA) approval. Human sensing neuroprosthetic devices currently depend upon control signals from residual nerve or muscle activity to restore motor functions lost due to disease or trauma. It has been proposed that these devices could be significantly improved by directly harnessing brain activity from central motor-related regions to drive artificial actuators (Chapin 2000; Nicolelis 2001; Caves, Shane et al. 2002; Nicolelis, Chapin et al. 2002; Chapin and Chapin 2004). Recently, laboratory studies involving nonhuman primates have made considerable advances toward the development of such devices. For example, neuronal ensemble recordings from motor areas of cerebral cortex in nonhuman primates have been demonstrated to accurately predict three-dimensional arm movements (Chapin, Moxon et al. 1999; Wessberg, Stambaugh et al. 2000; Taylor, Tillery et al. 2002; Carmena, Lebedev et al. 2003; Nicolelis, Dimitrov et al. 2003) and to successfully control a robotic arm neuroprosthetic device. Despite these interesting advances, primate studies have yet to address the fundamental question as to whether current brain–machine interface (BMI) technology and approaches may be successfully applied to human patients, in particular, those who are naïve regarding the eventual tasks (Wolpaw, Birbaumer et al. 2002; Patil, Carmena et al. 2004; Gage, Ludwig et al. 2005). Nonhuman primate BMI studies suggest that multineuronal recordings are critical for neuroprosthetic applications, and may require a minimum of 50–100 recorded neurons to drive a real-time neuroprosthesis (Nicolelis 2001, 2003; Sanchez, Carmena et al. 2004). In addition to cortical motor regions, subcortical regions, such as the motor thalamus and subthalamic nucleus, are also involved in motor planning and execution, and could serve as alternative multineuron recording sites (Lenz, Kwan et al. 1990; Cheruel, Dormont et al. 1996; Abosch, Hutchison et al. 2002; Guillery, Sherman et al. 2002; MacMillan, Dostrovsky et al. 2004; Patil, Carmena et al. 2004). Devices utilizing control signals from the nervous system have also been developed recently to enhance functional independence, using external reflections of brain events rather than direct neuronal recordings, such as electroencephalogram (EEG), direct cortical surface recordings (ECoG), or evoked potentials (Kubler, Kotchoubey et al. 1999; Donchin, Spencer et al. 2000; Pfurtscheller, Guger et al. 2000; Birch, Bozorgzadeh et al. 2002; Wolpaw, Birbaumer et al. 2002; Scherberger, Jarvis et al.


Approaches to Human Neural Ensemble Recordings

243

2005). These external signals suffer considerable information loss, because the control signal is derived from thousands or millions of neurons averaged across time and space. For example, scalp EEG signals can enable the control of approximately 6–7 characters per minute on an optimized keyboard, for a short period, but this is very limited for most purposes (Wolpaw, Birbaumer et al. 2002). Although a large variety of devices and approaches to neuroprosthetics are currently available, there is not at present a robust control signal that can be derived directly from the brain using noninvasive methods and that leads to fast, reliable conversion of thoughts into actions (Nicolelis 2001, 2003). This challenge leads to two problems. The first is that a high-throughput, reliable control signal is needed to directly link the brain with external devices, for translation of thought into action (Figure 12.1). The second is the inherent understanding of what packets of action potentials mean to the brain, and how sensorimotor information is concurrently processed by multiple subcortical and cortical structures that define a neural circuit. This challenge is thus posed from two different angles—the clinical treatment domain of using a control signal (regardless of its meaning if it works) to actuate an external event, and the research domain of interpreting information generated by networks of neurons involved in coding, ultimately leading to a better understanding of brain function. For both of these challenging goals, the concept and practical achievement of neural ensemble recordings are critical; the various types of available and envisioned devices will be discussed in detail.

Implanted Brain Electrodes

Spike Processing and Telemetry

Prediction Algorithm to Reconstruct Future Action from Past Spike Events Visual and Sensory Feedback

Real (Robot) or Virtual (Computer) Action for Motor or Communication Augmentation

FIGURE 12.1 Concept of BMI neuroprosthesis. The goal of a “sensing” neuroprosthetic device requires some form of recording interface with the nervous system, usually including implanted brain electrodes, and a technique to amplify and process these electrical signals. The signals can then be used to form predictions of future activity for device control functions, assisted with visual and/or somatosensory feedback.


244


RELEVANT PRECLINICAL STUDIES FOR HUMAN NEURAL ENSEMBLE RECORDINGS Recent work using nonhuman primates has demonstrated the feasibility of using multiple cortical arrays in a BMI format for eventual patient use, focusing on predicting motor function during a task, whereas this is only one aspect of typical human tasks (which include communication). Typically, a nonhuman primate has been implanted with a large number of cortical electrodes (128–1000), and of these an active population has been selected (Wessberg, Stambaugh et al. 2000; Carmena, Lebedev et al. 2003; Sanchez, Carmena et al. 2004). Nonhuman primates typically can rapidly learn a virtual task, or one in which the reward is derived from moving an external device rather than their own extremity. The presence of visual feedback provides a more rapid training path toward such a virtual task, to allow the primate direct feedback on performance. Direct somatosensory feedback can also facilitate learning and usefulness of such devices. First, techniques for electrode array (32–64 microwire electrodes per array) implantation into the cortex have been devised, so that the electrode recordings are stable for up to 2 to 6 years, depending on which primate species is used as a subject (Kralik, Dimitrov et al. 2001; Nicolelis, Dimitrov et al. 2003). A human equivalent of a subcortical microwire bundle is shown in Figure 12.2. These microwire arrays provide excellent long-term, multiunit neuronal recordings, with a typical extracellular triphasic profile (Figure 12.3). Second, real-time multichannel systems have been devised for recording from a large number of channels (up to 512 currently; Plexon, Inc., Dallas, Texas). The timing of sorted neuron spike waveforms then can be transmitted real time (40 KHz sampling rate) to a processing computer that uses past and current neuronal behavior to predict the position of the arm in space. The prediction is a control signal (in real time) that can drive an external actuator, such as a small robotic arm device (Sensable 1.5, Phantom) or a large, robust robotic arm for heavier tasks (Chapin, Moxon et al. 1999; Wessberg, Stambaugh et al. 2000; Serruya, Hatsopoulos et al. 2002; Carmena, Lebedev et al. 2003). The accuracy of the

FIGURE 12.2 32 Channel microwire array. Devices for even temporary use in the human brain require either FDA approval or an investigational device exemption (IDE). At present few recording electrodes fall into this category, except for single-unit metal electrodes (i.e., tungsten or Pt/Ir; F. Haer Company) and bundles of microwire electrodes (AdTech Medical). The figure shows a picture of a bundle of 32 of such Pt/Ir microwires, used for acute and subacute (i.e., < 30 days) human recordings. The scale bar represents 1 cm, and once introduced into the brain, the microwires diverge as shown.



245

Ch4

Ch12

250 µV

Ch13

Ch19 6.4

6.6 6.8 7.0 Time (sec)

7.2

0 800 Time (µs)

FIGURE 12.3 Multiunit data. The following data show an epoch of recording from multiple channels of a microwire electrode (Patil, Carmena et al. 2004). The varying single-unit action potentials are sorted and demonstrated to the right, for each channel. Note the different waveforms, and occasional multiple units, in a single channel (i.e., chapter 12).

predictions using a linear algorithm has been excellent, ranging up to 90% after a training period. In practice, the nonhuman primate can rapidly learn to control the external device using the brain interface directly, if there is sufficient visual feedback provided to properly clue the animal. These ground-breaking experiments have led the way to consider an equivalent device for humans, for sensing the brain’s intent and control of an external device (Turner, Dimitrov et al. 2004; Ojakangas, Shaikhouni et al. 2006).

REQUIREMENTS FOR HUMAN ENSEMBLE RECORDING AND NEUROPROSTHETICS DEVICES The translation from a nonhuman primate device into a human clinical BMI has a number of requirements, to be demonstrated prior to consideration for FDA approval. First, a reliable stable, electrode system is needed: (1) for detection of either inherent nervous system signals (neuronal action potentials) or an averaged or derived signal (such as EEG); or (2) microstimulation of small groups of nerve cells or axons. The solution has generally been platinum/iridium (Pt/Ir) contacts, which show minimal long-term alterations, such as in DBS electrodes. However, a long-term neuronal recording device has not so easily been developed. There are many version of silicon (Figure 12.4) and biologically inspired electrodes for detecting output signals from multiple neurons (Kennedy, Bakay et al. 1998, 2000), but the stability over time of the resulting signals, and whether local brain damage occurs, have been questioned. However, simple microwires have been used for chronic recordings in both humans (Cameron, Yashar et al. 2001; Staba, Wilson et al. 2002; Ekstrom, Kahana et al. 2003) and nonhuman primates for many years. In addition, Pt/Ir microwires show inherently minimal damage at the electrode–brain interface, even with constant


246


FIGURE 12.4 Utah microarray. This silicon array has been available for several years, and consists of multiple silicon spikes attached to a base (courtesy, Cyberkinetics Web site). The depth of the spikes is either 1.0 or 1.5 mm. A pressure applicator is used to insert the silicon array into the cortex, to cross the pia.

stimulation (Figure 12.2). The typical size of these microwires is 30–40 µm, the end or cut surface representing the electrode surface. Second, a complex, multichannel stimulation and/or recording system is critical to impart specificity to the channels communicating with the nervous system. For stimulation, this may require multiple, separately controllable microstimulation channels, each spatially organized for optimal stimulation that can be combined to form a pattern recognizable to the patient in terms of a specific sound, in the case of the cochlear prosthesis. For recording, multiple electrodes (as in the case of microwire arrays) require a large number of amplifier channels and suitable computer software for interpretation of the singles, and transformation of these signals into an appropriate control signal. Typically, microwire arrays of anywhere between 32 channels and 1000 channels have been utilized, and only recently have complex, multichannel amplifier and spike detection systems become available (Plexon Inc., Dallas, Texas). Third, a method of interpreting and decoding the neuronal signals is required. For a motor neuroprosthesis, this requires combining or translating the multineuron signals into a robust motor control signal, such as three-dimensional arm movement in space, which can be replicated on a robot (Lauer, Peckham et al. 2000; Gage, Ludwig et al. 2005; Sanchez, Erdogmus et al. 2005). Various approaches to such processing of the spike data into a desired target-control data set have included a linear combination algorithm, and recurrent neural networks (Gage, Ludwig et al. 2005; Sanchez, Erdogmus et al. 2005). However, this final control signal must be translated into a sufficiently detailed signal that a peripheral device can understand and act on. Fourth, communication with an external device is required for either sensory or motor neuroprostheses. For sensory prostheses, a detector relevant to the sensory modality is critical, such as an acoustic discriminator for a cochlear prosthesis, or a visual imaging device for a visual prosthesis. For motor prostheses, an appropriate motor control device is needed, such as a robot arm with a gripper, for performing tasks such as eating, to be controlled by the brain–computer interface. Many virtual tasks are also of great relevance, such as an optimized keyboard to allow rapid transmission of characters for communication (Reed 1998; Kubler, Kotchoubey et al. 1999, 2001; Caves, Shane et al. 2002; Wolpaw, Birbaumer et al. 2002; Kubler, Neumann et al. 2005).



247

What would a motor neuroprosthetics device look like? A simple example is electric wheelchair control (Nicolelis 2001; Nicolelis, Chapin et al. 2002). A wheelchair could be controlled to move forward or in other directions, at a certain rate, mimicking an external joystick with direct brain control. Another example is a robotic arm for enhancing independent eating, in a patient with quadriplegia and minimal hand or arm function. The common aspects of such a device include a brain electrode array, implanted in one or more areas of the brain (Figure 12.1), and detection and sorting of the action potential data. These data can then be combined (in a linear, nonlinear, or other optimized format) into a device output stream, for example a set of (X,Y,Z) coordinates for delivery to the robotic arm to control motion. The size and shape of a final, implanted product would likely closely resemble the current version of the DBS electrode. The DBS electrode currently includes a brain electrode, an electrical extension, and a control unit (implanted in the chest wall, similar to a pacemaker). It is likely that the neuroprosthetics system would include multiple arrays of microelectrodes with independent microwires, implanted into either the cortex or subcortical structures. An implanted system would then connect to preamplifiers and spike-sorting processing, then the signals transferred via radio telemetry to an external device for actuation. An initial system would include visual feedback to identify the accuracy of the intended response, and to allow correction on subsequent trials. A more sophisticated system could include direct sensory feedback, to allow, for example, detection of weight and other properties and more sensitive tasks to be performed (Nicolelis 2001, 2003). This envisioned system would be close to the size, degree of invasiveness, and compassionate use of the DBS system already in common clinical practice, which is well accepted by patients, and entails a relatively low risk in terms of untoward brain complications.

APPROPRIATE HUMAN MEDICAL CONDITIONS Conditions of the nervous system occur in which output and the human interface with the external environment becomes impaired (Chapin 2000; Caves, Shane et al. 2002; Wolpaw, Birbaumer et al. 2002; Caplan 2003). These conditions can be roughly categorized into two groups. The first group includes situations in which the supratentorial central nervous system (CNS) is intact, but there is damage at either the brainstem or spinal level. In these conditions the cerebral cortex and cognition are functioning (as in a quadriplegic or a patient with spinal paralysis secondary to amyotrophic lateral sclerosis [ALS]), but the central representation of the periphery is altered, due to the drastically changed motor output and sensory input. The most severe type of patient in this group is the “locked-in” patient, in whom a brainstem stroke or damage has left normal cortical functioning, but in whom there is virtually no residual interface with the environment, except for perhaps eye movements (Kennedy, Bakay et al. 1998; Kubler, Kotchoubey et al. 2001, 2005). The second group includes patients whose supratentorial CNS has suffered the primary damage, as in the case of a stroke accompanied by aphasia or hemiparesis, for example. In both of these groups of patients there is a profound need for enhanced communication, interaction with the environment, and control of external devices, for quality of life, independence, activities of daily living, and output of creative thought (Caves, Shane


248


et al. 2002). Most current prostheses depend upon residual peripheral control, for example, eye movements, to activate an external device. These are highly limited in bandwidth, in terms of the ability to transmit effective information back and forth between the brain and the environment (Caves, Shane et al. 2002). For this reason, considerable interest has developed in a direct brain–computer interface, in which direct brain control of either external devices or a natural limb can be achieved. The potential for this type of interface include a higher bandwidth and more natural device control, by using signals that the brain itself generates, for the purpose of interactions with the environment. However, there have been significant limitations in almost every aspect of the development of such interfaces.

HUMAN TECHNICAL CONSIDERATIONS: ENGINEERING TRANSLATION The design questions to be resolved before a permanent, implantable system can be specified include: (1) the stability of the neuronal signals over time, and optimal neuronal/field potential/EEG recording electrodes and interface; (2) the number of effective channels needed for device control; (3) whether spike sorting and spike detection are needed; (4) the brain location for optimal control for various devices; and (5) the training period and methods required to achieve useful device control, to establish proof of principle for patient functional enhancement. These design questions assume that the external, peripheral devices to be controlled and the algorithms to optimally convert a neural signal into action are similar to those either in common clinical use, or already have been developed in preclinical studies. Once these needs are resolved, a permanent system can be engineered, tested in nonhuman primates for FDA approval studies, and eventually permanently implanted in patients. Several types of electrodes have now been in use or considered for multineuron recording. These include microwires (metal insulated wires, exposed at the tip, 15–40 µm in diameter), silicon devices with one or multiple contacts per structured array (Figure 12.4 as an example), and the Kennedy neurotrophic electrode (to attract ingrowth of axons) (Kennedy, Bakay et al. 1998, 2000). Sharp electrodes (i.e., tungsten or Pt/Ir) with a single contact are excellent for short-term physiology but may be limited in durability for prolonged recordings. A significant roadblock to many of these electrode types is human pia, which is thick and rubbery. Sharp tungsten electrodes may penetrate the pia and have been used for single-unit recording in humans for years, but have not been maintained long-term to assess for durability. Microwires in general cannot breach the pia, so surgically or chemically opening the pia is critical for their insertion (Nicolelis, Dimitrov et al. 2003). The silicon-based electrodes (such as the Utah electrode; Cyberkinetics Web site [Maynard, Nordhausen et al. 1997]) require a pneumatic insertion device because of the multiple points, and are limited to a depth of 1.5 mm. Various versions are available for restricted human use, including the Kennedy neurotrophic electrode (on an IDE), the Cyberkinetics Utah silicon array (on an IDE), and the AdTech microwire bundle (FDA approved for surgical use only). It is not clear if any of these devices meet FDA criteria for chronic, permanent implantation, and currently there are no approved internal implantable electronics (i.e., amplifiers, digitizers, telemetry devices, etc.), for direct connection to these



249

multineuron arrays. In contrast, current nonhuman primate systems depend upon large head-mounted connectors that could not be tolerated in a long-term human setting. Thus, there is considerable impetus to work toward a fully implantable system, for possible eventual human application (Nicolelis 2001; Turner 2004). The timeline for such an implantable system, however, may require years of FDA approval studies prior to human application. We next consider what has been done so far in humans, and electrode functioning and constraints.

PRELIMINARY HUMAN STUDY EXPERIENCE MICROWIRE ARRAY EXPERIENCE The AdTech Medical microwires (Figure 12.2) have been used for over 15 years, as an adjunct microelectrode inserted into the brain, along with depth electrodes (Cameron, Yashar et al. 2001; Staba, Wilson et al. 2002; Ekstrom, Kahana et al. 2003). These single-unit studies have analyzed a number of individual neuronal characteristics. We have adapted a bundle of these microwires (32; Figure 12.2) for insertion into subcortical areas during placement of DBS electrodes (Patil, Carmena et al. 2004). Short-term (intraoperative) human testing has been performed from an initial cohort of 11 patients, using a concurrent motor task. These electrode signals were then amplified in a 32-channel Plexon system and stored. We have recorded 32-channel human single-unit data in subthalamic nucleus and motor thalamus during DBS placement. Simple upper-extremity motor tasks were developed and optimized for the intraoperative testing situation, namely a squeezeball to assess grip pressure or a virtual reality glove to measure hand and finger position. The studies performed so far have focused on subthalamic nucleus (STN) and motor thalamus (VIM/VOP), during placement of DBS. The protocol used has included routine single-unit recordings in these structures for localization of the DBS, then placement of the 32-channel electrode (Figure 12.2) at an optimal location. Multiple channels have been recorded (up to 55), and the signal-to-noise ratio has averaged 5.5 for the 32-wire channels. Various single units are then sorted by the Plexon software (Figure 12.2). There was noted to be a correlation between movement and activity in many recordings, though there was considerable instability of the multineuron data even over 5 min. Using recurrent neural network algorithms, predictions for future motion patterns with R values as high as 0.82 have been calculated, particularly for VIM/VOP thalamic units. These preliminary intraoperative data clearly show the feasibility of using even a limited number of human cortical or subcortical neurons (ranging from 8 to 55) for predicting the performance of a hand task, from a single trial. Despite severe constraints due to limited time, small neuronal ensembles, as well as operative and patient considerations, the recorded neuronal signals were sufficiently information rich that the trained models were able to predict hand gripping force during the test period quite accurately, as shown for four sessions (R2 = 0.38 to 0.68). Overall prediction performance, across 20 recording sessions was significant, though modest (R2 = 0.26 ± 0.04; mean ± SEM). This predictive value may be sufficient for a simple task control, for example. In general, the ANN, least mean square (LMS), or Kalman filter decoding algorithms outperformed the simpler Wiener filter


250


algorithm. Furthermore, similar to findings in nonhuman primate studies, prediction accuracy increased with both training period duration and the number of neurons within the ensemble, suggesting that the longer training intervals and the larger ensemble sizes possible with chronic implantation of a human BMI would likely result in improved BMI performance. These findings demonstrate the feasibility of the hypothesis that chronically recorded neuronal units in subcortical motor centers may provide an effective motor control signal for task prediction. In nonhuman primate studies, model-training periods of 10 min are common, and then animal training periods of several days may be needed for accurate task prediction. Our findings resemble the results of nonhuman primate studies (Wessberg, Stambaugh et al. 2000; Carmena, Lebedev et al. 2003), because increasing correlations between hand gripping force and model predictions were obtained by adding neurons and lengthening the training period. Thus, multichannel human data are highly promising for development of a chronic, human, multichannel BMI, if the recordings can be extended to larger groups of neurons, and the models remain useful for these chronic recordings. Further information on long-term stability and usefulness of microwires for neuronal recordings will await the development of an implantable neuroprosthetic device.

CYBERKINETICS AND THE UTAH ELECTRODE EXPERIENCE Cyberkinetics Co. has recently attempted to leverage nonhuman primate findings using the Utah silicon array (Figure 12.4), to perform initial clinical trials. This array has been used in rats and nonhuman primates, though the longevity of singleunit recordings and their quality continues to be controversial (Maynard, Nordhausen et al. 1997; Donoghue 2002; Serruya, Hatsopoulos et al. 2002). Cyberkinetics recently published data (Hochberg, Serruya et al. 2006) from a preliminary feasibility clinical trial, with long-term externalized implants of the Utah array, placed in the motor cortex of patients with severe disabilities. Despite the high expectations raised by Cyberkinetics, the initial results were mixed. Overall, only one out of four implanted patients was capable of utilizing the Cyberkinetics system to control either a computer cursor or a one-degree-of-freedom robotic hand. Further testing is planned to enhance these initial clinical results.

CORTICAL ELECTROCORTICOGRAM (ECOG) AND BRAIN CONTROL Surface EEG systems have been used for several years as initial brain–computer interfaces, with the EEG signal as the primary brain signal (Wolpaw, Birbaumer et al. 2002). However, scalp surface EEG is bandwidth limited to approximately 20 Hz and demonstrates poor spatial and temporal characteristics. It has recently been noted that cortical surface electrical recordings (ECoG) provide much better signal-to-noise resolution than scalp surface EEG, and much higher bandwidth, potentially to ~500 Hz signal content, and increased spatial resolution. Because cortical surface recordings are invasive (and not FDA approved past the perioperative recording stage), these signals have primarily been recorded during invasive epilepsy evaluation, when a grid or series of macro EEG electrodes are temporarily placed on



251

the surface of the brain, for diagnostic reasons. Recently, using bandwidth partitions of the ECoG signal, it has been possible to perform binary decisions in real time, indicating that perhaps cortical surface EEG may inherently contain information for at least simple task prediction (Leuthardt, Schalk et al. 2004; Leuthardt, Schalk et al. 2006). An extension of cortical surface EEG is local field potential within the cortical layers, which has also been used as a control signal when multineuron units are not available (Scherberger, Jarvis et al. 2005). Cortical surface ECoG technology exists in terms of electrodes, because Medtronics stimulating electrodes are also excellent EEG recording electrodes (and FDA approved for epidural use), but not FDA approved for cortical surface use at this point.

FUTURE HUMAN NEURAL ENSEMBLE RECORDINGS Several issues will require resolution for a clinical brain–machine implant, including development of stable (and nonreactive) sensing, medical-grade electrodes for some level of brain functioning (i.e., multineuron, field potential, or ECoG), implantable electronics for amplifying, digitizing, and externalization of the brain signals, external electronics for task prediction and realization on a virtual or realistic device, learning algorithms for naïve tasks, and augmentation of sensory feedback for sensing the external world and device capabilities.

TYPES OF ELECTRODES Superficial Cortical Electrodes As discussed earlier, the Cyberkinetics Utah silicon array is one prototype of a fixedspace recording array, though the depth is limited to 1.5 mm below the pial surface due to silicon manufacturing constraints. Microwires, particularly those constructed of Pt/Ir, tend to be insufficiently rigid for pial penetration and hence require some form of pial opening for placement. One option is stripping back the pia, but unfortunately this will lead to devascularization of that region of cortex (and may possibly enhance seizure generation), because the blood vessels to brain gyri are on the undersurface of the pia. Another option is to use a sharper electrode, such as a comb silicon electrode with less penetration points, or a Pt/Ir stiffer microelectrode design. The Bak electrode (Figure 12.5) is an example of an array developed for pial penetration, using thicker Pt/Ir or Ir microwires/microelectrodes, though still, similar to microwires, limited to a single end contact per electrode. The Bak electrode has a thin connecting wire, which may enhance flexibility, to allow the electrode to float with the brain, and to avoid anchoring to the skull. Long-term degradation of the multineuron signal may occur with micromotion around the electrode, and if anchored only to the cortex this may be minimized, compared to skull anchoring. Many of the comb silicon electrodes offer a higher density of electrode contacts spaced along the shaft, allowing recording at multiple depths. To understand the proper depth, we obtained human motor cortex from a fresh cadaver specimen, and then stained the neurons using a Nissl stain, demonstrated in Figure 12.6. The nominal depth to layer V (with Betz cells, large cortical


252


FIGURE 12.5 Microprobes (Bak) microarray. Pt/Ir is the preferred material of choice for human brain implants, based on the long history of metals safe for implant (particularly for stimulation). The Bak microarray is a series of tungsten microwires, sharpened at the tips for insertion past the pia into the cortex. The example shown has an electrode array bonded to a ceramic plate, with fixed spacing, and microwires 4 mm in length. The flexible, sealed side-mounted connecting cables would allow placement beneath the dura and tunneling out from the dura and skull at a distance from the electrode array. This is one alternative design to tungsten fixed-density microwire arrays, constructed from medical-grade components (courtesy Microprobes Web site).

FIGURE 12.6 Human motor cortex. Although the depth of the layers in mammalian motor cortex is well known, the specific measurements of human motor cortex are not commonly reported. To analyze the depth of various cortical layers, human motor cortex was identified from a fresh cadaveric brain, then processed for Nissl stain, and sectioned perpendicular to the cortical surface. The various layers are marked, including the presence of Betz cells in layer V. The horizontal bar is 0.5 mm. The nominal depth to layer V is 1.8–2.0 mm, but when corrected for shrinkage (~40%, considering the ethanol dehydration used after fixation), this value corrects to ~3.0 mm, indicating the depth to which electrodes must penetrate to reach layer V.

motoneurons) is about 1.8–2.0 mm. However, Nissl preparations use ethanol tissue dehydration, leading to at least 40% linear tissue shrinkage. After correcting for shrinkage, this suggests appropriate electrode depth to reach layer V is 2.5–3 mm at least. A staggered array may be optimal, to allow different depths to be sampled.



253

Deep Electrode Arrays The placement of microwire arrays into deep structures (such as motor thalamus, for example) depends upon the electrodes appropriately fanning out. Although this was demonstrated with intraoperative fluoroscopy (Patil, Carmena et al. 2004), it is not clear if the electrodes fan out to a degree that the various microwires are actually in independent regions of the thalamus. There may be increased stability with such deep placement compared to cerebral cortex, because the electrode is in effect anchored for several additional centimeters within brain substance, decreasing brain pulsation and movement with respect to the skull. New versions of deep electrodes include partially flexible electrodes with multiple contacts along the sides, allow multiple recording points per electrode. Microwire Recording Electrodes Many different electrode designs have been suggested for the chronic, ensemble recordings required of a neuroprosthetic BMI (Moxon, Leiser et al. 2004). Among these, microwire arrays have been demonstrated to record chronically from multiple neurons for periods up to 2 years (Nicolelis, Dimitrov et al. 2003). Although these microwire arrays have been used in diagnostic invasive epilepsy recordings for up to a few days, there is no experience on their longevity in vivo in nonhuman primate models.

SILICON AND OTHER RECORDING ARRAYS Silicon electrodes have been developed partly for their ease of manufacture, and the possible addition of recording electronics directly upon the silicon chip that provides the electrode base. However, there remain concerns about the long-term biocompatibility with silicon substrates in the brain, as well as long-term recording capabilities.

ELECTRONICS Coupled with multineuron arrays there is a need for implanted electronics to match. One approach used in research at Duke University has been to place the amplifiers directly on the same silicon chip as the silicon electrodes. However, it is not clear if this may result in more noise or not. Another approach has been to locate the electronics a small distance away from the electrodes (i.e., a few millimeters), so that the electrode can float with the brain, rather than being directly skull attached. In both instances, there is a need for at least ~100 channel amplifiers, which leads to considerable problems with information transfer unless there is information reduction at some level. Although a digital spike detector may be the most efficient technique for spike detection, the analog signal is not then available for later spike sorting, which is the usual approach to discriminating multiple neurons per channel. However, analog detection is more difficult and may require automatic spike sorting approaches, which are not yet well implemented. The approach has been to filter the raw signal for each channel, digitize and then perform an initial threshold spike detection, and then keep the analog data for all channels just around the time of the detection (i.e., similar to the 0.8–1.0 ms Plexon wave storage). A separate option may be to scrutinize a single analog channel at a


254


time, as opposed to keeping the short waveforms for all of the channels temporarily, to allow verification that the channel is still working and has clear neuronal-related signals. The telemetry out to a real-time system will clearly be bandwidth limited and, thus, the critical data (such as the ISI or spike occurrence, or waveforms) should be limited, for transfer out to the real-time actuation system (such as a robotic arm). The development and packaging of these electronics will be critical to implementation of a human neuroprosthetic, which will be far more complicated than any current device.

INTERVENTION AND TRAINING METHODS In most cases the BMI will be most useful in individuals with poor control of their extremities or in communication enhancement (Chapin 2000; Caves, Shane et al. 2002; Chapin and Chapin 2004). In such patients, many desired tasks to enhance activities of daily living and independence will be novel, and there will not be training data on how the brain can carry out the task, particularly with respect to multineuron recording. The process of acquiring mastery over a naïve task using brain control has not been well mapped out, but several potential strategies exist. For example, fMRI studies have shown that the performance of imaginary tasks, using an imaginary extremity, are similar to those of real tasks in terms of cortical activation, except that the cerebellum is not engaged (presumably due to its error correction role) (Lotze, Flor et al. 2001; Roux, Ibarrola et al. 2001; Stippich, Ochmann et al. 2002; Giraux, Sirigu et al. 2003; Nair, Purcott et al. 2003). In microelectrode mapping of the motor thalamus in humans, MacMillan et al (MacMillan, Dostrovsky et al. 2004) also noted that thalamic neurons respond to imaginary tasks (similar to fMRI activation of thalamus). Imaginary extremity movement can be entrained to synchronize with an external movement, and likely neuronal firing follows this entrainment (Giraux, Sirigu et al. 2003). Thus, one approach to training for a naïve task is to entrain the imaginary or illusory percept of movement (of a missing extremity, for example) with an external percept, such as a predictable, moving object. Additionally, having electrodes located in the brain region normally subserving the desired task may facilitate rapid training.

SENSORY FEEDBACK—STIMULATION ELECTRODES AND MULTINEURON FIELD STIMULATION Visual feedback of a task being trained under brain control is critical, so that the desired task goals can be assessed on an ongoing basis. External sensory feedback can also be applied, for example, as a pressure on the skin, with the pressure applied being proportional, for example, to a robot gripper. Thalamic microstimulation can also lead to the perception of pressure, particularly in ventral intermediate thalamus (VIM), the region for proprioceptive perception (Hua, Garonzik et al. 2000); hence, graded microstimulation of sensory thalamus may provide a highly stable, long-term form of sensory feedback, particularly if a spatial pattern of microstimulation can be applied to coordinate with the external task. For example, a robot gripper with force sensation in the grippers could be used to trigger thalamic microstimulation in implanted electrodes, such as DBS electrodes, and then the patient trained to use this alternative sensation as a cue. If a DBS electrode is used, then the Medtronics



255

stimulator could be reprogrammed to induce appropriate stimuli to the sensory thalamus in coordination with external events.

CONCLUSIONS The translation of preclinical multineuron recording applications to the clinical level remains challenging. Although there is preliminary data available from short-term testing with microwires, the two devices taken into longer-term human trials have both shown limited capabilities, primarily, cursor movement and binary choices (such as the Kennedy single-unit device [Kennedy, Bakay et al. 1998]). Scalp EEG recordings provide similar binary information (Wolpaw, Birbaumer et al. 2002), so the worth of an implanted device has not yet been demonstrated, in spite of the considerable promise. However, the ability to hold information-rich cortical or subcortical neuronal units for a prolonged period of time is a critical aspect of such a planned multineuron array, and this maintenance clearly has yet to be shown. Other challenges include implanted multichannel electronics and telemetry, how to train for a novel task when there is no or limited training data, and the ability to provide afferent information to the CNS via a technique such as microstimulation to alert a patient as to location and grip of a potential prosthesis. Thus, in spite of the proof of principle having been shown by many nonhuman primate groups for cortical control of a motor prosthesis, the translation into clinical applicability and the clinical proof of principle remain critical challenges.

ACKNOWLEDGMENTS This work was supported by a DARPA grant to M.A.L. Nicolelis.

REFERENCES Abosch, A. and Hutchison, W.D. et al. (2002). Movement-related neurons of the subthalamic nucleus in patients with Parkinson disease. J. Neurosurg. 97(5): 1167–72. Birch, G.E. and Bozorgzadeh, Z. et al. (2002). Initial on-line evaluations of the LF-ASD brain-computer interface with able-bodied and spinal-cord subjects using imagined voluntary motor potentials. IEEE Trans. Neural Sys. Rehabil. Eng. 10(4): 219–24. Cameron, K.A. and Yashar, S. et al. (2001). Human hippocampal neurons predict how well word pairs will be remembered. Neuron 30(1): 289–98. Caplan, A.L. (2003). Is better best? A noted ethicist argues in favor of brain enhancement. Sci. Am. 289(3): 104–5. Carmena, J.M. and Lebedev, M.A. et al. (2003). Learning to control a brain-machine interface for reaching and grasping by primates. Plos Biol. 1(2): E42. Carmena, J.M. and Lebedev, M.A. et al. (2003). Learning to control a brain-machine interface for reaching and grasping by primates. Public Libr. Sci. 1(2): 193–208. Caves, K. and Shane, H.C. et al. (2002). Connecting AAC devices to the world of information technology. Assistive Technol. 14(1): 81–9. Chapin, J.K. (2000). Neural prosthetic devices for quadriplegia. Curr. Opin. Neurol. 13(6): 671–5. Chapin, J.K. (2004). Using multi-neuron population recordings for neural prosthetics. Nat. Neurosci. 7(5): 452–5.


256


Chapin, J.K. and Moxon, K.A. et al. (1999). Real-time control of a robot arm using simultaneously recorded neurons in the motor cortex. [see comment]. Nat. Neurosci. 2(7): 664–70. Cheruel, F. and Dormont, J.F. et al. (1996). Activity of neurons of the subthalamic nucleus in relation to motor performance in the cat. Exp. Brain Res. 108(2): 206–20. Donchin, E. and Spencer, K.M. et al. (2000). The mental prosthesis: assessing the speed of a P300-based brain-computer interface. IEEE Trans. Rehabil. Eng. 8(2): 174–9. Donoghue, J.P. (2002). Connecting cortex to machines: recent advances in brain interfaces. Nat Neurosci 5 Suppl: 1085–8. Ekstrom, A.D. and Kahana, M.J. et al. (2003). Cellular networks underlying human spatial navigation. Nature 425(6954): 184–8. Gage, G.J. and Ludwig, K.A. et al. (2005). Naive coadaptive cortical control. J. Neural Eng. 2(2): 52–63. Giraux, P. and Sirigu, A. et al. (2003). Illusory movements of the paralyzed limb restore motor cortex activity. Neuroimage 20 Suppl. 1: S107–11. Guillery, R.W. and Sherman, S.M. et al. (2002). The thalamus as a monitor of motor outputs. Philos. Trans. R Soc London—Series B: Biol. Sci. 357(1428): 1809–21. Hochberg, L.R. and Serruya, M.D. et al. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia.[see comment]. Nature 442(7099): 164–71. Hua, S.E. and Garonzik, I.M. et al. (2000). Microelectrode studies of normal organization and plasticity of human somatosensory thalamus. J. Clin. Neurophysiol. 17(6): 559–74. Kennedy, P.R. and Bakay, R.A. et al. (1998). Restoration of neural output from a paralyzed patient by a direct brain connection. Neuroreport 9(8): 1707–11. Kennedy, P.R. and Bakay, R.A. et al. (2000). Direct control of a computer from the human central nervous system. IEEE Trans. Rehabil. Eng. 8(2): 198–202. Kralik, J.D. and Dimitrov, D.F. et al. (2001). Techniques for long-term multisite neuronal ensemble recordings in behaving animals. Methods (Duluth) 25(2): 121–50. Kubler, A. and Kotchoubey, B. et al. (1999). The thought translation device: a neurophysiological approach to communication in total motor paralysis. Exp. Brain Res. 124(2): 223–32. Kubler, A. and Kotchoubey, B. et al. (2001). Brain-computer communication: unlocking the locked in. Psychol. Bull. 127(3): 358–75. Kubler, A. and Neumann, N. et al. (2005). Brain-computer interfaces--the key for the conscious brain locked into a paralyzed body. Prog. Brain Res. 150: 513–25. Lauer, R.T. and Peckham, P.H. et al. (2000). Applications of cortical signals to neuroprosthetic control: a critical review. IEEE Trans. Rehabil. Eng. 8(2): 205–8. Lenz, F.A. and Kwan, H.C. et al. (1990). Single unit analysis of the human ventral thalamic nuclear group. Activity correlated with movement. Brain 113(Pt 6): 1795–821. Leuthardt, E.C. and Schalk, G. et al. (2006). The emerging world of motor neuroprosthetics: a neurosurgical perspective. Neurosurgery 59(1): 1–14; discussion 1–14. Leuthardt, E.C. and Schalk, G. et al. (2004). A brain-computer interface using electrocorticographic signals in humans. J. Neural Eng. 1(2): 63–71. Lotze, M. and Flor, H. et al. (2001). Phantom movements and pain. An fMRI study in upper limb amputees. Brain 124(Pt 11): 2268–77. MacMillan, M.L. and Dostrovsky, J.O. et al. (2004). Involvement of human thalamic neurons in internally and externally generated movements. J. Neurophysiol. 91(2): 1085–90. Maynard, E.M. and Nordhausen, C.T. et al. (1997). The Utah intracortical Electrode Array: a recording structure for potential brain-computer interfaces. Electroencephalogr. Clin. Neurophysiol. 102(3): 228–39. Moxon, K.A. and Leiser, S.C. et al. (2004). Ceramic-based multisite electrode arrays for chronic single-neuron recording. IEEE Trans. Biomed. Eng. 51(4): 647–56.



257

Mussa-Ivaldi, F.A. and Miller, L.E. et al. (2003). Brain-machine interfaces: computational demands and clinical needs meet basic neuroscience. Trends Neurosci. 26(6): 329–34. Nair, D.G. and Purcott, K.L. et al. (2003). Cortical and cerebellar activity of the human brain during imagined and executed unimanual and bimanual action sequences: a functional MRI study. Cogn Brain Res. 15(3): 250–60. Nicolelis, M.A. (2001). Actions from Thoughts. Nature 409: 403–407. Nicolelis, M.A. (2003). Brain–Machine Interfaces to Restore Motor Function and Probe Neural Circuits. Nat Rev Neurosci. 4: 417–422. Nicolelis, M.A. and Chapin, J.K. et al. (2002). Controlling robots with the mind. Sci. Am. 287(4): 46–53. Nicolelis, M.A. and Dimitrov, D. et al. (2003). Chronic, multisite, multielectrode recordings in macaque monkeys. Proc. Natl. Acad. Sci. U.S.A. 100(19): 11041–6. Ojakangas, C.L. and Shaikhouni, A. et al. (2006). Decoding movement intent from human premotor cortex neurons for neural prosthetic applications. J. Clin. Neurophysiol. 23(6): 577–84. Patil, P.G. and Carmena, J.M. et al. (2004). Ensemble recordings of human subcortical neurons as a source of motor control signals for a brain–machine interface. Neurosurgery 55(1): 27–35; discussion 35–8. Pfurtscheller, G. and Guger, C. et al. (2000). Brain oscillations control hand orthosis in a tetraplegic. Neurosci. Lett. 292(3): 211–4. Reed, C.M. and Durlach, N.I. (1998). Note on information transfer rates in human communication. Presence 7: 509–518. Rodriguez-Oroz, M.C. and Obeso, J.A. et al. (2005). Bilateral deep brain stimulation in Parkinson’s disease: a multicentre study with 4 years follow-up.[see comment]. Brain 128(Pt 10): 2240–9. Roux, F.E. and Ibarrola, D. et al. (2001). Virtual movements activate primary sensorimotor areas in amputees: report of three cases. Neurosurgery 49(3): 736–41; discussion 741–2. Sanchez, J.C. and Carmena, J.M. et al. (2004). Ascertaining the importance of neurons to develop better brain-machine interfaces. IEEE Trans. Biomed. Eng. 51(6): 943–53. Sanchez, J.C. and Erdogmus, D. et al. (2005). Interpreting spatial and temporal neural activity through a recurrent neural network brain-machine interface. IEEE Trans. Neural Syst. Rehabil. Eng. 13(2): 213–9. Scherberger, H. and Jarvis, M.R. et al. (2005). Cortical local field potential encodes movement intentions in the posterior parietal cortex. Neuron 46(2): 347–54. Serruya, M.D. and Hatsopoulos, N.G. et al. (2002). Instant neural control of a movement signal. Nature 416(6877): 141–2. Staba, R.J. and Wilson, C.L. et al. (2002). Sleep states differentiate single neuron activity recorded from human epileptic hippocampus, entorhinal cortex, and subiculum. J. Neurosci. 22(13): 5694–704. Stippich, C. and Ochmann, H. et al. (2002). Somatotopic mapping of the human primary sensorimotor cortex during motor imagery and motor execution by functional magnetic resonance imaging. Neurosci. Lett. 331(1): 50–4. Taylor, D.M. and Tillery, S.I. et al. (2002). Direct cortical control of 3D neuroprosthetic devices.[see comment]. Science 296(5574): 1829–32. Turner, D.A., Dimitrov, D.F., and Nicolelis, M.A.L. (2004). Neuroprosthetics and clinical realization of brain-machine interfaces. Modern Neurosurgery: Clinical Translation of Neuroscience Advances. D.A. Turner. New York, CRC Press. Wessberg, J. and Stambaugh, C.R. et al. (2000). Real-time prediction of hand trajectory by ensembles of cortical neurons in primates.[comment]. Nature. 408(6810): 361–5. Wolpaw, J.R. and Birbaumer, N. et al. (2002). Brain-computer interfaces for communication and control. Clin. Neurophysiol. 113(6): 767–91.


Methods for Neural Ensemble Recordings,

Pattern classification using ensemble methods

Neural Systems for Control

Neural Systems for Control

Neural Systems for Control

Ensemble, c’est tout

Ensemble, c’est tout

Ensemble, c’est tout

Neural Stem Cells: Methods and Protocols

Neural Stem Cells (Methods in Molecular Biology)

Neural Stem Cells. Methods and Protocols

Neural Cell Transplantation. Methods and Protocols

Neural Stem Cells. Methods and Protocols

Neural Cell Transplantation Methods and Protocols

Neural Solutions for Information Security

Neural Networks for Pattern Recognition

Sensitivity analysis for neural networks

Recurrent neural networks for prediction

Neural Networks for Hydrological Modelling

Sensitivity Analysis for Neural Networks

Neural Networks for Pattern Recognition

Protocol for Neural Cell Culture

Feature Selection and Ensemble Methods for Bioinformatics: Algorithmic Classification and Implementations

Neural Representation and Neural Computation

Neural Engineering

Metaheuristic procedures for training neural networks

Data Assimilation The Ensemble Kalman Filter

Data Assimilation: The Ensemble Kalman Filter

Data Assimilation: The Ensemble Kalman Filter

Metaheuristic procedures for training neural networks

Methods for Neural Ensemble Recordings,

Pattern classification using ensemble methods

Neural Systems for Control

Neural Systems for Control

Neural Systems for Control

Ensemble, c’est tout

Ensemble, c’est tout

Ensemble, c’est tout

Neural Stem Cells: Methods and Protocols

Neural Stem Cells (Methods in Molecular Biology)

Neural Stem Cells. Methods and Protocols

Neural Cell Transplantation. Methods and Protocols

Neural Stem Cells. Methods and Protocols

Neural Cell Transplantation Methods and Protocols

Neural Solutions for Information Security

Neural Networks for Pattern Recognition

Sensitivity analysis for neural networks

Recurrent neural networks for prediction

Neural Networks for Hydrological Modelling

Sensitivity Analysis for Neural Networks

Neural Networks for Pattern Recognition

Protocol for Neural Cell Culture

Feature Selection and Ensemble Methods for Bioinformatics: Algorithmic Classification and Implementations

Neural Representation and Neural Computation

Neural Engineering

Metaheuristic procedures for training neural networks

Data Assimilation The Ensemble Kalman Filter

Data Assimilation: The Ensemble Kalman Filter

Data Assimilation: The Ensemble Kalman Filter

Metaheuristic procedures for training neural networks

Recommend Documents