U.S. flag

An official website of the United States government

NCBI Bookshelf. A service of the National Library of Medicine, National Institutes of Health.

Nicolelis MAL, editor. Methods for Neural Ensemble Recordings. 2nd edition. Boca Raton (FL): CRC Press/Taylor & Francis; 2008.

Cover of Methods for Neural Ensemble Recordings

Methods for Neural Ensemble Recordings. 2nd edition.

Show details

Chapter 9Multielectrode Recording in Behaving Monkeys

and .

INTRODUCTION

The most versatile neurophysiological paradigms for the study of cognitive function in animals are those that involve recording the activity of neurons in awake and behaving monkeys. Techniques for recording in behaving monkeys were originally developed by Herbert Jaspers and colleagues (Jasper et al., 1960) and elaborated by Edward Evarts (Evarts, 1966; Evarts, 1968) in the 1960s. Conventional recording methods, based on these early developments, employ single movable sharp electrodes to isolate single cells in regions of interest. Cells must be recorded serially over many weeks to accumulate enough data to characterize the population of cells under study. More recently, systems that permit several sharp electrodes—from approximately 2 to 16—to be independently positioned have improved the yield and allow the activity of several cells to be monitored simultaneously. Recordings of this kind, however, are generally restricted to a single cortical or subcortical site and can be maintained only for a short time.

In this chapter, we will describe a range of contemporary and emerging methods for studying the activity of large numbers of neurons, in multiple cortical and subcortical locations, for periods extending from several weeks to several years. Techniques for conducting multielectrode recordings will be described with special emphasis on issues unique to conducting these experiments in behaving monkeys* (readers interested in more detail will be directed to the more specialized chapters of this volume and other resources where appropriate). To demonstrate the value of these methods for addressing significant neuroscientific questions, an example of the use of multi-electrode recordings in behaving monkeys from our own work will be described.

Nonhuman Primate Models

One of the central goals of neuroscientific research is the elucidation of the neurophysiological mechanisms underlying normal human behavior and cognition. Progress in this area has often suffered from methodological limitations. Available methods for directly measuring activity in the central nervous system (CNS) are invasive and cannot, therefore, be used on human beings. In recent years, non-invasive methods such as electroencephalographic recording (EEG) and functional magnetic resonance imaging (fMRI), which can be used with human subjects, have been applied to study aspects of cognition ranging from simple visual perception to economic decision making. In spite of the tremendous enthusiasm generated by the results of such studies, the insights these methods provide into the actual underlying neurobiological mechanisms are clearly limited. The two most prominent and widely used methods for measuring neurophysiological activity in human beings, EEG and fMRI, have a number of important limitations.

Relating a particular behavioral or cognitive event to scalp recorded EEGs requires overcoming a formidable signal-to-noise ratio (SNR). The deviations a single event introduce into the ongoing EEG, which has an amplitude of approximately 100 μV, have an amplitude of only 1–10 μV. In combination with the contaminating ambient electrical noise inevitably introduced during the recording, this small SNR necessitates averaging large numbers of trials to detect a signal related to the event of interest. The results of such studies, therefore, provide some information about the average time at which different cognitive processes diverge. They provide little information, on the other hand, about where such differences occur. Because of the high conductivity of brain tissue, EEG signals readily propagate and are simultaneously picked up by many recording electrodes.

fMRI, on the other hand, offers better spatial resolution—approximately 1 mm—but is based on the detection of changes in blood oxygenation levels (i.e., the BOLD signal) that, in addition to being an indirect measure of neuronal activity, have severely limited temporal resolution (the BOLD signal is delayed between 1 and 5 s and peaks 4–5 s later). Therefore, to directly study the neuronal activity of the CNS that underlies perception, cognition, and action, therefore, we must rely on invasive recording methods.

Although neurophysiological recordings can be conducted in human subjects during the course of medically necessary neurosurgical procedures, ethical considerations and practical constraints are too limiting for the vast majority of questions systems cognitive neuroscientists would address. Ethical and practical considerations also prohibit the use of apes, the group of animals most closely related to human beings, for the kind of invasive procedures required to record neuronal activity in the brains of behaving animals. Old-World monkeys, the group of animals next in kinship with human beings (the evolutionary divergence of human beings and chimpanzees, the most closely related apes, occurred approximately 6–9 million years ago, whereas humans and apes diverged from Old World monkeys 25 million years ago), are, therefore, widely considered the most appropriate primate model for neurophysiological research. Macaque monkeys, the largest subspecies of Old World primates, have become the most popular primate models, with Rhesus and Long-tailed macaques being the most prominent. Because of the extensive use of Rhesus monkeys in neurobiological research, investigators can take advantage of standard neuroanatomical atlases (e.g., Paxinos et al., 1999) to guide the placement of recording electrodes. Using animals with well-characterized anatomy and heavily studied physiology offers the additional benefit of more accurate comparison with the results of other investigators. In addition, a complete sequence of the genome of the Rhesus macaque became available last year (2006)*, which will enhance the value of Rhesus macaques for neurophysiological investigations as attempts to combine electrophysiological and molecular methods progress.

Macaque monkeys can be trained to perform a variety of perceptual and cognitive tasks, ranging from simple perceptual discriminations to more elaborate tasks, such as decision making and categorization. No other animal model provides such a range of behavior. Simple and effective operant conditioning methods can be used to systematically teach monkeys to perform complex behavioral tasks. In typical paradigms, monkeys will perform such tasks from 2 to 6 h, and most available techniques allow data to be collected over several months.

There are a number of reasons to believe that extrapolation from neurophysiological results obtained in macaque monkeys is generally valid. First, the functional neuroanatomy of macaque monkeys, although not identical, is similar to that of humans (Orban et al., 2004; Van Essen et al., 2001). Second, the sensory systems of macaques appear to be quite similar to those of human beings. For example, well-trained monkeys appear to have similar visual discrimination thresholds to those observed in humans (Crist et al., 2001). Color vision in macaques appears, furthermore, to permit the same range of color distinctions that humans can make (Loop and Crossman, 2000; Sperling and Harwerth, 1971). Finally, the similarity of the body plan of monkeys and human beings makes designing very similar motor paradigms possible. Recently, macaques have even been used to model human locomotor behavior (Nakatsukasa et al., 2006; Ogihara et al., 2005).

Multielectrode Recording Methodology

Whatever functional role one attributes to individual neurons, it is clear that even the simplest primate perceptions or behaviors involve the coordinated activity of a large population of neurons distributed in multiple cortical and subcortical parts of the brain. What individual neurons contribute to the production of particular cognitive events, however, is an issue of some debate. A popular view, which has guided a great deal of neurophysiological investigation over the last 40 years, is that each neuron (at least in sensory and motor systems) represents a single basic feature at a particular level of the appropriate representational hierarchy. For example, the process of vision is widely believed to begin with the deconstruction of the visual scene into a set of fundamental elements, such as form, texture, motion, color and depth. Individual neurons in the early stages of cortical visual processing, accordingly, are said to be tuned to one of these attributes. In this scheme, the firing of one of these neurons represents an estimation of the probability that its preferred stimulus (e.g., an oblique contour) has been encountered. Thus, the functional role of an individual cell is to signal whether or not a particular feature is present. At subsequent stages of the processing hierarchy, inputs from lower levels converge onto individual cells and confer upon them the ability to signal the presence of an appropriate higher-order feature (e.g., surfaces, objects). This view has an attractive simplicity and has undoubtedly motivated a great deal of informative research. One important consequence of this view is that elaborate neuronal circuits can be adequately characterized by serially recording their individual neuronal constituents. Nevertheless, sufficient evidence has accumulated over the years that many in the field now believe that this conception of the role of individual neurons is inadequate (Nicolelis et al., 1997).

One problem that raises difficulties for many models of cortical processing is the variability of neuronal responses. Neurons produce different responses in (apparently) identical circumstances. The traditional approach is to treat such variability as noise. In order for a physiologist to characterize the response properties of a particular cell, therefore, it is necessary to average a large number of responses to the same event. Of course, this cannot be the way the brain works, and to account for the ability of an organism to react in a timely fashion to real-world events, it has been proposed that the brain averages the activity of many neurons whose role in the system is redundant—they all provide the same information.

An alternative view is that the functional role played by individual neurons is much more dynamic than the traditional view implies. In this view, individual neurons are sensitive to multiple aspects of any given cognitive event (e.g., multiple features of a stimulus to which the animal was exposed; Schiller, 1996), and the response produced by the neuron depends on both the external and internal context in which the event takes place (e.g., surrounding features of the environment, the current interests of the animal, prior experience with the stimulus, etc). In this view, then, the variability of neuronal responses is not noise but reflects the role the neuron is playing within the ensemble of neurons coordinating the response of the animal. To understand the response generated by a particular neuron to a singular event (e.g., the presentation of a stimulus in an experiment), it is necessary to adequately sample the activity of other constituents of the network of responding cells. Doing so, however, requires the ability to simultaneously record the activity of large numbers of neurons residing in differing cortical and subcortical locations.

BEHAVIORAL TRAINING

Macaque monkeys can be trained to perform a wide range of sensory discrimination and motor control tasks. Training a monkey to perform a complex behavioral task, however, often requires substantial investment of time and effort. Careful consideration and planning of the methods used to train the monkey can be of considerable benefit.*

Preliminary Training

The training of a naïve monkey begins by acclimation of the monkey to the presence of the investigator. Over the course of one to two weeks, the investigator visits the monkey in its home cage and spends 30–45 minutes encouraging the monkey to approach the front of the cage and take treats (e.g., pieces of fruit) from the investigator’s hand. The familiarity of the monkey with the investigator and the positive association of the investigator with reward developed during this period will reduce the amount of anxiety the monkey experiences during subsequent training and facilitate the training process.

Once the monkey is comfortable in the presence of the investigator (e.g., will remain in the front of the cage near the investigator and wait to receive additional treats), training the monkey to accept the handling and restraint procedures that will be used during the experiment can begin. The pole-and-collar system is a widely used method of handling macaque monkeys in the research laboratory setting. A specially designed collar, typically made of nylon or aluminum, is placed around the monkey’s neck under light anesthesia and the monkey is allowed several days to acclimate to the presence of the collar. A rigid pole, typically about 60 cm in length, with a clasp on one end for attachment to the collar, is used to remove the monkey from the cage. Naïve monkeys are often highly agitated during their initial experiences of being restrained and, therefore, it is best to proceed slowly and allow the animal to become familiar with the pole-and-collar system. A common strategy is to capture the monkey, restrain them for several minutes without removing them from the cage while providing frequent food rewards, and release them. Doing this several times a day for 1 to 2 weeks will generally eliminate the resistance of the monkey to restraint.

The last phase of preliminary training is to introduce the monkey to the restraint system that will be used during the experiment. In many paradigms, monkeys are seated in a primate restraint chair during the course of the experiment. Primate chairs typically restrain the monkey by fixing the monkey’s collar while the monkey stands or sits on a perch. In some cases, the monkey’s body will be entirely enclosed, which eases subsequent manipulations, but it may be desirable for the monkey’s arms to be unrestrained. Many paradigms require fixing the monkey’s head in place with a head-mounted post. Allowing one to two weeks for monkeys to become comfortable sitting in the chair with all of the intended restraints will generally facilitate subsequent training.

Training for the Experimental Paradigm

Given the tremendous variety of interesting behaviors that macaques are capable of producing, highly varied behavioral paradigms have been developed by different investigators. A standard paradigm suitable for the investigation of a wide range of neurobiological questions cannot, therefore, be prescribed. Nevertheless, there are some rules of thumb that can be used for the development of successful training schemes.

Most importantly, the use of positive reinforcement is strongly recommended. Monkeys are subject to a number of potential stressors in the laboratory. Using positive reinforcement to encourage voluntary cooperation from the monkeys in a study only helps to reduce the stress to which the animals are subject and may be more effective than alternative methods (Schapiro et al., 2003). Reducing the stress of the monkeys may not only benefit the animal’s welfare, duress may result in unforeseen physiological changes that increase the variability of experimentally obtained data (Reinhardt, 2003).

The first step in training a monkey to perform a complex behavioral task is to decompose the task into a series of simpler tasks that progressively approximate the final behavior. To the extent possible, it is most efficient to leverage the natural behavioral repertoire of the animal. For example, if a component of the final task requires the monkey to pull and hold a lever, training might begin by rewarding the monkey for simply touching the lever. This takes advantage of monkey’s natural tendency to explore the environment and associates reward with a component of the desired behavior. Once the monkey begins routinely grabbing the lever to obtain reward, one can introduce the requirement that the lever be held for a short time prior to obtaining reward.

In order to successfully apply positive reinforcement training, a few key rules should be followed. First, an appropriate biologically relevant reward should be chosen (typically food or drink). Access to the reward should be restricted to periods of successful behavior. If, for example, the monkey is to be rewarded with drops of orange juice, orange juice should not be provided in other circumstances (e.g., in the home cage). Finally, the difficulty of obtaining reward should be incrementally increased such that the monkey is regularly receiving rewards throughout the course of training. Frequent failure to obtain reward can extinguish the desired behavior.

MICROELECTRODES

Microwire Arrays

Small-gauge insulated wires (microwires) can be used to manufacture simple microelectrodes and microelectrode arrays for recording single unit activity.* Successful recording critically depends on the size and material properties of the wire selected. Determining the optimal size for microwire electrodes, however, involves balancing several factors. The larger the gauge of the wire, the more tissue damage will be done when the electrode is inserted. Furthermore, the reduced input impedance of larger-diameter wires makes isolating single units more difficult. Smaller-diameter wires will do less tissue damage and have higher input impedance. However, the lower tensile strength of smaller-diameter wires makes them more difficult to insert into the cortex without bending.

In addition to the size, strength, and conductivity of the wire, other aspects of the material must be considered. The saline environment of the brain is highly corrosive, and both the insulation and the wire itself must be constructed of material that can maintain their integrity in such an environment. In addition, the presence of foreign material evokes an elaborate immunological response from the surrounding tissue in the CNS (Polikov et al., 2005). Within 6 weeks of implantation, electrodes are often encased in a sheath of reactive astrocytes, typically called a glial scar. Materials used to construct microelectrodes would ideally minimize or eliminate this response.

The shape of the tip of the microelectrode also significantly impacts the quality and longevity of recording. The simplest, and perhaps most commonly used, method involves simply cutting off the end of the wire to make a blunt tip with the full diameter of the wire exposed. A variety of other tip shapes have been successfully employed, including sharp pointed tips, conical tips, and beveled tips. Sharp-tipped electrodes that have been insulated such that only the very tip of the electrode is exposed have the advantage of having very high input impedance and high SNR, ratio making the isolation of single units comparatively easy. In addition, a sharp tip makes puncturing the cortex easier and minimizes the depression of cortical tissue. However, the high input impedance of such electrodes limits the distance over which neuronal signals can be detected and necessitates placing the electrode tip very close to a cell to pick up the signal. Thus, these electrodes are most useful in acute recording experiments with restrained animals where a microdrive can easily be used to position the electrode in a desirable location and the chance that units will be lost too quickly is minimized by reducing motion in the system as much as possible.

The lower impedance of blunt-tipped electrodes means that they can pick up neuronal signals over a larger distance and thus record the activity of multiple units simultaneously. Although this can make the isolation of single-unit activity more difficult, blunt-tipped electrodes offer several advantages critical for long-term chronic recordings. The wider range over which neuronal signals can be detected means that the positioning of the electrode is less critical. This raises the possibility that signals will be detected, and makes the recordings more robust to motion. In practice, it is often possible to identify single units with blunt-tipped electrodes, and frequently, the activity of more than one unit can be discriminated with a single electrode.

Tungsten or stainless steel wires of 12 to 50 μm insulated with Teflon or S-isonel, are now used routinely for chronic cortical and subcortical recordings. Arrays of such microwires containing as many as 128 electrodes have been constructed by using printed circuit boards (PCB) and high-density miniaturized connectors. By minimizing the size of the implant, it has become feasible to record from large numbers of neurons simultaneously in multiple cortical and subcortical regions. In addition, these recordings can be maintained for many months.

Tetrodes

To overcome the difficulties of single-unit isolation with individual blunt-tipped microwires, probes comprising multiple wires can be constructed. Tetrodes of four microwires have gained considerable popularity in recent years (Buzsaki, 2004; Jog et al., 2002). The microwire components of a tetrode are of differing lengths, making the position of the recording tips distinct in three dimensions. Therefore, a spike from a neuron in the vicinity of the tetrode end will have slightly different magnitude on each of the four component electrodes due to the small differences in the distance between the neuron and the recording ends of the electrodes. Such differences in magnitude can be exploited to localize the neuronal source of the spike in space and improve the discrimination of activity generated by different neurons. At the same time, tetrodes inherit many of the advantages of blunt-tipped electrodes. The low impedance of the component electrodes allows them to pick up activity from many neurons and makes them more robust to mechanical perturbations. Thus far, however, tetrodes have primarily been used to record in rodents.

Thin-Film Electrodes

Photofabrication of thin-film circuits has long been recognized to offer the potential of creating electrodes with many recording sites in a very small volume. The first thin-film electrodes for neurophysiological use were constructed in the mid-1960s, and a wide variety of thin-film electrodes based on different materials (e.g., glass and ceramic) have been developed over the last 30 years (Pearce and Williams, 2007; Wise et al., 2004). Although many successful acute recordings in cell culture and in vivo have been conducted with thin-film electrodes, their use for chronic recordings has been hindered by several technological issues, including the strength of the substrate and identification of dielectric material appropriate for long-term implantation in nervous tissue (Polikov et al., 2005).

SURGICAL IMPLANTATION

While implanting several cortical or subcortical regions permits scientific questions to be addressed that could not be addressed in any other way, an extensive and difficult surgery lasting many hours must be performed. Experience suggests that the quality of the surgical procedure used to implant the electrodes dramatically affects the success of chronic recordings. Particular care must be taken to ensure the physiological well-being of an animal maintained under deep anesthesia for such extended periods of time. Here we will briefly outline the procedure that we have used successfully to implant more than 700 electrodes in the cortex of a monkey (Nicolelis et al., 2003).*

Arrays of microelectrodes are surgically implanted using aseptic technique. After the surface of the skull is exposed, craniotomies are drilled above each of the regions to be implanted. We have found that recording stability and longevity is enhanced by keeping the craniotomies as small as possible—no larger than strictly necessary to accommodate the array. After resection of the exposed dura mater, an electrode array is mounted on a micropositioner and positioned such that the electrodes of the array are approximately normal to the cortical surface. High-density electrode arrays can compress the cortex, sometimes considerably, prior to penetrating the pia mater. Compression of the cortical surface only a few millimeters can produce several undesirable effects, including ischemia and neuronal death. The best method of minimizing these kinds of tissue damage has been a matter of controversy. Some have found that a very rapid, ballistic insertion procedure produces the best results (e.g., Rennaker et al., 2005). In contrast, we have found that our arrays can be inserted with minimal cortical dimpling by advancing them very slowly (approximately 100 μm/min). In our experience, this procedure produces the highest yield of electrodes with good SNRs ratios and recordings that can endure in the macaque as long as 18 months (Nicolelis et al., 2003). To ensure the correct placement of the electrodes, single- and multiunit activity is monitored, and neuronal response properties are qualitatively assessed. For example, the position of an array in somatosensory cortex can be determined by observing the activity elicited by tactile stimulation of the appropriate part of the body. Once the array is known to be in the desired location, the implant is fixed in position using dental cement. Following the experiment, the positions of the electrodes can be confirmed histologically.

SIMULTANEOUS MULTICHANNEL RECORDING

Microwire recordings from multiple single-units have been conducted since the 1960s; however, it is only in recent years that technologies for recording and analyzing the activity of very large numbers of individual neurons have become commonly available. Several commercial systems are now available that can be purchased off the shelf, which permit the sampling, amplification, and analysis of tens and even hundreds of channels simultaneously. One such system, which we have used extensively, is the Multichannel Acquisition Processor (MAP).* The standard MAP is a modular system that provides programmable amplification, filtering, and spike sorting on up to 128 electrodes. The current system permits the discrimination of as many as four units on each channel, permitting the investigator to record the activity of 512 single cells. With minor hardware customization, we were able to daisy chain four MAP systems together to simultaneously record from 512 channels in a single macaque monkey in a recent experiment (Nicolelis et al., 2003).

The MAP system begins its signal conditioning in preamplifiers connected by short leads to head-stages plugged into the connectors of the microelectrode arrays on the monkey’s head. The signal is amplified (by a factor of 16), band-pass-filtered (100 Hz – 16 KHz), and transmitted to the main component of the MAP system via ribbon cables. The signal is amplified again (by a factor of 1, 10, or 20), band-pass-filtered a second time (400 Hz – 8 KHz), and passed to a programmable amplifier (ranging from 1 to 30 times). The resultant signal is digitized with a resolution of 12 bits at 40 KHz. Finally, programmable digital signal processors (DSPs) capture segments of the signal that cross a user-definable threshold and permit real-time spike detection.

One of the great advantages of the MAP system is the ability to control the amplification, filtering, and spike sorting settings of the DSPs from a single off-the-shelf microcomputer. Manual spike sorting can be conducted in real time with the support of several available statistical tools. Generally, a combination of principal components analysis to identify discrete clusters and time-voltage boxes defined by the investigator are used to isolate waveforms that belong to a single unit. These parameters are downloaded to the DSPs for automatic spike detection.

DATA ANALYSIS

Simultaneously recording the activity of large numbers of neurons is technically challenging but, as can be seen throughout this volume, this challenge is being overcome for many animal species and intraoperative recordings in human subjects. Neurophysiologists working with animals models are motivated to overcome such challenges because they believe that important functional properties of the nervous system cannot be recovered from serially sampling individual neurons. One of the greatest challenges facing investigators, however, is the development of analytical techniques capable of quantifying ensemble level properties. Here we will briefly review several methods we have used successfully in our laboratory.*

We have employed several standard multivariate analysis methods to characterize the spatiotemporal pattern of neuronal ensemble activity associated with sensory processing or motor control.

Principal Component Analysis

Principal component analysis (PCA) is a common method for identifying patterns in high-dimensional data sets. It can be employed to reduce the number of variables needed to describe a data set by identifying those that explain the majority of the variance in the data. The components extracted with PCA are orthogonal by definition, and therefore associated with independent sources of variance. Functional interactions between neurons in the ensemble recorded can be extracted and their relationship to particular behaviors examined.

In our application of PCA to neurophysiological data, we begin by binning the data by counting the number of spikes each neuron fired within a certain period of time (typically 10–20 ms). This results in a set of vectors containing the spike counts of each neuron in the sample. A correlation matrix is constructed from these vectors, and the eigenvectors are calculated. The components are formed by the weighted linear sum of the firing of individual neurons, and the contribution of any particular neuron is reflected in the magnitude of its weight. Interesting patterns can be identified by examining the way in which the weights of individual neurons cluster in a space defined by the components, which account for the majority of the variance in the data set.

Independent Component Analysis

Independent component analysis (ICA) is a statistical technique based on a model that assumes that an observed signal is a linear mixture of signals originating from an unknown set of sources. A number of different ICA algorithms have been developed. In general, ICA attempts to isolate factors based on higher-order statistics (e.g., kurtosis) of the data. One recent study used simulated spike train data constructed to reflect common sources of input for different subsets of the artificial units to examine the ability of several different ICA algorithms to segregate sources in neuronal data (Laubach et al., 1999). It was found that ICA can accurately reconstruct several underlying sources of synchronous firing with a population of neurons. Because different sources of input to a given population of neurons are represented in a PCA as a linear combination of the principal components, independent sources cannot be identified.

Discriminant Analyis

Discriminant analysis (DA) is a technique for classifying observations into known classes. By carrying out a series of calculations similar to multivariate analysis of variance (MANOVA), a combination of variables is identified that can discriminate between classes. Subsequently, the model is refined by systematically calculating the effect of including or excluding a variable on the F-value to determine which variables should be included in the final model. In our studies, we have employed DA to single-trial-related differences on neuronal ensemble firing patterns.

Artificial Neural Networks

Artificial neural networks (ANNs) can be used to detect patterns in high-dimensional data sets without making any assumptions about the structure of the data. A large number of ANN architectures have been developed, and many software packages are available for ANN implementation. We have used both back propagation and competitive ANNs to successfully distinguish patterns of neuronal ensemble activity on a single-trial basis.

Preprocessing of the data is often necessary to obtain the best classification performance from an ANN. Typically, some type of normalization or feature extraction is performed on the data prior to feeding it to an ANN. Both the spontaneous and evoked firing rates of different neurons can vary considerably; however, neurons with a high average firing rate are not necessarily the most informative. Therefore, to prevent neurons with higher firing rates from obscuring the contribution of other units, we typically normalize the spike counts from each unit to have mean zero and unit standard deviation. Classification of large high-dimensional data sets can often be improved by reducing the dimensionality of the input. Reducing the number of input variables also reduces the amount of data necessary to train an ANN. A simple-minded feature extraction method is to identify parts of the data that are most likely to contain information related to the desired classification. For example, neurons that are not modulated by any of the stimuli used in an experiment could be excluded from subsequent analysis. Unsupervised methods of data reduction, such as PCA and ICA, can dramatically reduce the number of input variables while retaining most of the information in a given data set.

EXAMPLE: BRAIN–MACHINE INTERFACES

Estimates indicate that over 200,000 people suffer paralysis in the United States as a consequence of spinal cord injury, with 11,000 new patients appearing every year (Nobunaga et al., 1999). Although progress has been made in attempts to remediate the loss of function associated with spinal cord injury through encouraging neuronal growth to reconstruct the loss connectivity, it is unclear at this time how successful this line of investigation will be.

An alternative approach to restoring the function to paralyzed patients, originally proposed by Schmidt (1980), involves creating an artificial interface to functional neural tissue to bypass the spinal cord and permit control of an external device. In order to replace the function of a lost limb, such as the arm, an actuator of sufficient dexterity must be designed. Most importantly, a level of control that will permit the kind of articulate movements characteristics of the upper limb is required. A large-scale, multisite multielectrode capable of monitoring the activity of ensembles of neurons will be a critical component of any such system. Indeed, recent studies in rodents (Chapin et al., 1999; Talwar et al., 2002), nonhuman primates (Serruya et al., 2002; Taylor et al., 2002; Wessberg et al., 2000), and human patients (Birbaumer et al., 1999) have demonstrated that contemporary multielectrode recording methods can be used to generate control signals for artificial actuators. Here we describe a recent study from our laboratory demonstrating the use of a brain–machine interface (BMI) to permit macaque monkeys to control a robotic arm in the absence of actual limb movements (Carmena et al., 2003).*

Figure 9.1A illustrates the elements of the paradigm. Monkeys were seated in a primate restraint chair facing a video monitor on which all stimuli were displayed and provided with a joy-stick-like manipulandum with a force-sensing handle. Signals recorded by a commercial data acquisition system were collected by custom software that produced predictions of motor parameters in real time. The resulting predictions were used to control a robotic arm, and feedback about the motion of the arm was reflected in the behavior a cursor presented on the video monitor.

FIGURE 9.1. (See color insert following page 140.

FIGURE 9.1

(See color insert following page 140.) Experimental setup, behavioral tasks, changes in performance with training, EMG records during pole and brain control, and stability of model predictions. (A) Behavioral setup and control loops, consisting of data (more...)

Two female Rhesus macaques were trained to perform three behavioral tasks (Figure 9.1B) to obtain fluid rewards. In the original task (task 1), the monkey attempted to move a small red disk (the cursor) onto a large green disk (the target) that appeared, at the beginning of each trial, in a random location. In task 2, the monkey was presented with a yellow disk (the cursor) whose diameter reflected the gripping force exerted by the monkey. On each trial the monkey was presented with a pair of red circles, concentric with the cursor, to indicate the gripping force the monkey was required to make in order to obtain a reward. The final task (task 3), required the monkey to combine the elements of task 1 and task 2—to move the cursor to the intended location and generate a particular gripping force. After a period of training, arrays containing between 16 and 64 microwires were surgically implanted in the frontal and parietal regions of the cortex of each animal. Implanted areas included dorsal premotor cortex (PMd), supplementary motor area (SMA), and primary motor cortex (M1) in both hemispheres. In addition, monkey 1 was implanted in primary somatosensory cortex (S1), and monkey 2 was implanted in the medial intraparietal area (MIP). In total, monkey 1 received implants containing 96 microwires, and monkey 2 received 320 microwires. After the monkeys recovered from the surgery, training was resumed while the activity of cortical cells was recorded.

Figures 9.1C–9.1E illustrate procedural motor learning as animals interacted with the BMIc in each of the three tasks. Improvement in behavioral performance with the BMIc was indicated by a significant increase in the percentage of trials completed successfully (Figures 9.1C–E, top graphs), and by a reduction in movement time (Figures 9.1C–E, bottom graphs). For task 1, both monkeys had some training in pole control of task 1 (data not shown) several weeks before the series of successive daily sessions illustrated in Figure 9.1C. For both tasks 1 and 2, after a relatively small number of daily training sessions, the monkeys’ performance in brain control reached levels similar to those during pole control (Figures 9.1C, 9.1D). For tasks 2 and 3, all behavioral data are plotted given that in both cases pole and brain control was used since the first day of training. Behavioral improvement was also observed in task 3, which combined elements of tasks 1 and 2 (Figure 9.1E). In all three tasks, the levels of performance attained during brain control mode by far exceeded those predicted by a random walk model (dashed and dotted lines in Figures 9.1C–9.1E). Moreover, both animals could operate the BMIc without any overt arm movement and muscle activity, as demonstrated by the lack of EMG activity in several arm muscles (Figure 9.1G). The ratios of the standard deviation of the muscle activity during pole versus brain control for these muscles were 14.67 (wrist flexors), 9.87 (wrist extensors), and 2.77 (biceps).

A key novel feature of this study was the introduction of the robot equipped with a gripper into the control loop of the BMIc after the animals had learned the task. Figure 9.1C shows that, because the intrinsic dynamics of the robot produced a lag between the pole movement and the cursor movement, the monkeys’ performance initially declined. With time, however, the performance rapidly returned to the same levels as seen in previous training sessions (Figure 9.1C). It is critical to note that the high accuracy in the control of the robot was achieved by using velocity control in the BMIc, which produced smooth predicted trajectories, and by the fine tuning of robot controller parameters. These parameters were fixed across sessions in both monkeys. The controller sent velocity commands to the robot every 60–90 ms. Each of these commands compensated for potential position errors of the robot hand from previous commands.

In all experiments, the animals continuously received visual feedback of their performance. Unlike previous results in owl monkeys where an open-loop BMI was implemented (Wessberg and Nicolelis, 2004), after the model parameters were fixed, its predictions did not drift substantially from initial best performance even during 1-h recordings. As shown in the examples of Figure 9.1F, prediction of grasping force (mean = SE, r = 0.84 = 5 × 10−3) in monkey 1, and hand position (r = 0.63 = 3 × 10 −3) and velocity (r = 0.73 = 5 × 10−3) in monkey 2 remained very stable despite some transient fluctuations (slopes for black, magenta, and cyan lines are −2.16 × 10−4, −5.15 × 10−4, and −1.1 × 10−3). One possibility is that the presence of continuous visual feedback helped to stabilize model performance.

This paradigm allowed us to address a number of important questions about the development of motor control signals in the context of a BMI. Prior studies had shown that neuronal ensemble activity could be used to predict the endpoint of a monkey’s reach. Here we examined whether the activity of such neuronal ensembles could also be used to predict other aspects of motor behavior, such as the velocity and acceleration of the movement. Figure 9.2A shows records of the monkey’s hand position, hand velocity, and gripping force for a representative epoch. Overlaid are the predictions generated in real time by independent linear models running in parallel. As can be seen by examining the figures, accurate predictions of each of these parameters were obtained. After training, the linear models could account for up to 85% of the variance in hand position, 80% of the variance in hand velocity, 95% of the variance in gripping force, and 61% of the variance in simultaneously recorded EMG activity (Figure 9.2B and C).

FIGURE 9.2. (See color insert following page 140.

FIGURE 9.2

(See color insert following page 140.) 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. (A) Motor parameters (blue) and (more...)

Investigators developing experimental BMIs have focused their efforts on different components of the network of cortical areas involved in motor planning and motor control. Due to its proximity to motor output, two laboratories have focused their efforts on M1 (Serruya et al., 2002; Taylor et al., 2002). Others have argued that the more abstract cognitive representations believed to be present in posterior parietal cortex (PPC) are likely to be the best source of BMI control signals (Andersen et al., 2004; Musallam et al., 2004; Pesaran et al., 2002; Shenoy et al., 2003). Our own work has indicated that the best way of taking advantage of the highly distributed character of cortical motor planning is to sample multiple components of the frontal parietal network (Nicolelis, 2001; Nicolelis et al., 2003; Wessberg et al., 2000). Recording different cortical areas involved in motor control permitted us to quantify the contribution of neurons in different regions to the predicted motor parameters.

The results of this comparison are illustrated in Figures 9.2D–9.2F. The first thing to note is that neurons in all recorded cortical areas contributed to predictions of each of the parameters we investigated. The contributions of neurons in different areas, however, were distinct. M1 neurons alone were the best predictors of all motor variables, accounting for 73% of the variance in hand position, 66% for velocity, and 83% of gripping force. The activity of neurons residing in SMA was also a good predictor of variance in hand position (51%) and velocity (51%), but a poor predictor of gripping force (19%). Similarly, S1 ensemble activity predicted hand position and velocity well (48% and 35%, respectively), but gripping force poorly. The quality of predictions obtained from the activity of PMd neuronal ensembles paralleled the predictions from S1 neuronal activity—position and velocity were predicted well (48% and 46% respectively), while predictions of gripping force were worse (29%). In contrast, neuronal ensemble activity in PP produced accurate predictions of gripping force (73%), somewhat weaker predictions of velocity (52%), and a poor prediction of hand position (25%). Though information related to motor planning is clearly widely distributed in the cortex, these results indicate that subtle differences exist in the relationship between neuronal populations in separate cortical areas and different movement parameters. This kind of comparative knowledge is essential for choosing the best target regions for the therapeutic use of BMIs.

Another outstanding question is how many neurons must be sampled to achieve the desired level of control. Although some researchers have argued that small samples (i.e., 10–30 neurons) are sufficient for a BMI-based prosthetic (Serruya et al., 2002; Taylor et al., 2002), others believe that much larger numbers of neurons (i.e., 100s to 1000s) will be necessary to replicate the dexterous movements human beings are capable of making (Nicolelis, 2001; Wessberg et al., 2000). To examine this issue, we randomly excluded neurons from our analysis and produced predictions based on the remaining population. As can be seen in Figures 9.2G–9.2I, as the population of neurons becomes smaller, the quality of the predictions degrades. It is important to note that this degradation is smooth, indicating that the quality of the predictions depends more on the quantity of neurons in the sample than the inclusion of a few critical neurons.

A related question, and a subject of some debate, is what kind of signal is optimal for the prediction of motor behavior. To access this, we compared predictions based on multiunit recordings (i.e., the single from several neurons simultaneously recorded on a single electrode) to predictions based on single-unit activity. The results of this analysis are presented in Figures 9.2G–9.2I. These results clearly demonstrate the superiority of the predictions obtained from single units. For each motor parameter, the single unit predictions were more accurate than multiunit predictions. On the other hand, this data also demonstrates that accurate predictions can be obtained from multiunit recordings. This is significant because it is more difficult to obtain single-unit recordings than multiunit recordings, and more difficult to retain them as well. In this context, it is important to note that, although single-unit recordings produced better predictions of motor parameters, our results indicate that increasing the number of multiunit recordings can compensate for this discrepancy.

Reliable, long-term operation of a BMIc was achieved by extracting multiple motor parameters (i.e., hand position, hand velocity, and gripping force) from the simultaneously recorded activity of frontopariental neural ensembles. Macaque monkeys learned to use the BMIc to reach and grasp virtual objects with a robot even in the absence of overt arm movements. Accurate performance was possible because large populations of neurons from multiple cortical areas were sampled. Thus, the present study shows that large ensembles are preferable for efficient operation of a BMI. This conclusion is consistent with the notion that motor programming and execution are represented in a highly distributed fashion across frontal and parietal areas, and that each of these areas contains neurons that represent multiple motor parameters. We suggest that, in principle, any of these areas could be used to operate a BMI, provided that a large-enough neuronal sample was obtained. Although analysis of neuron dropping curves (Figures 9.2D–9.2F) indicates that a significant sample of M1 neurons consistently provides the best predictions of all motor parameters analyzed, neurons in areas such as SMA, S1, PMd, and PP contribute to BMI performance, as well.

SUMMARY

In this chapter, we have discussed methods for simultaneously recording large populations of neurons in the brains of monkeys engaged in purposeful behavior. These techniques have already yielded significant results and will no doubt be fruitfully used to study a wide range of cognitive and sensory processing questions in the future. Particularly exciting developments are the appearance of wireless recording systems (Lei et al., 2004; Wise et al., 2004). Such systems will very likely prove critical to the development of practical BMIs. In addition, they promise the development of experimental paradigms in freely moving moneys. Also of significant promise is the ongoing development of thin-film electrodes with large numbers of contacts in small volumes. In the near future, such technology may permit the simultaneous recording of thousands of neurons in the primate brain.

REFERENCES

  1. Andersen RA, Musallam S, Pesaran B. Selecting the signals for a brain-machine interface. Curr Opin Neurobiol. 2004;14:720–726. [PubMed: 15582374]
  2. Birbaumer N, Ghanayim N, Hinterberger T, Iversen I, Kotchoubey B, Kubler A, Perel-mouter J, Taub E, Flor H. A spelling device for the paralysed. Nature. 1999;398:297–298. [PubMed: 10192330]
  3. Buzsaki G. Large-scale recording of neuronal ensembles. Nat Neurosci. 2004;7:446–451. [PubMed: 15114356]
  4. Carmena JM, Lebedev MA, Crist RE, O’Doherty JE, Santucci DM, Dimitrov DF, Patil PG, Henriquez CS, Nicolelis MAL. Learning to control a brain-machine interface for reaching and grasping by primates. Plos Biol. 2003;1:193–208. [PMC free article: PMC261882] [PubMed: 14624244]
  5. Chapin JK, Moxon KA, Markowitz RS, Nicolelis MAL. Real-time control of a robot arm using simultaneously recorded neurons in the motor cortex. Nat Neurosci. 1999;2:664–670. [PubMed: 10404201]
  6. Crist RE, Li W, Gilbert CD. Learning to see: experience and attention in primary visual cortex. Nat Neurosci. 2001;4:519–525. [PubMed: 11319561]
  7. Evarts EV. A technique for recording the activity of subcortical neurons in moving animals. Electroencephalogr Clin Neurophysiol. 1966;24:83–86. [PubMed: 4169750]
  8. Evarts EV. Methods for recording activity of individual neurons in moving animals. In: Rushmer RF, editor. Methods in Medical Research. Chicago, IL: Year Book Medical Publishers; 1968. pp. 241–250.
  9. Jasper HH, Ricci G, Doane B. Microelectrode analysis of cortical cell discharge during avoidance conditioning in the monkey. Electroencephalogr Clin Neurophysiol. 1960;(Supplement 131):137–156. [PubMed: 14400427]
  10. Jog MS, Connolly CI, Kubota Y, Iyengar DR, Garrido L, Harlan R, Graybiel AM. Tetrode technology: advances in implantable hardware, neuroimaging, and data analysis techniques. J Neurosci Methods. 2002;117:141–152. [PubMed: 12100979]
  11. Laubach M, Shuler M, Nicolelis MA. Independent component analyses for quantifying neuronal ensemble interactions. J Neurosci Methods. 1999;94:141–154. [PubMed: 10638821]
  12. Lei Y, Sun N, Wilson FA, Wang X, Chen N, Yang J, Peng Y, Wang J, Tian S, Wang M, et al. Telemetric recordings of single neuron activity and visual scenes in monkeys walking in an open field. J Neurosci Methods. 2004;135:35–41. [PubMed: 15020087]
  13. Loop MS, Crossman DK. High color-vision sensitivity in macaque and humans. Vis Neurosci. 2000;17:119–125. [PubMed: 10750833]
  14. Musallam S, Corneil BD, Greger B, Scherberger H, Andersen RA. Cognitive control signals for neural prosthetics. Science. 2004;305:258–262. [PubMed: 15247483]
  15. Nakatsukasa M, Hirasaki E, Ogihara N. Energy expenditure of bipedal walking is higher than that of quadrupedal walking in Japanese macaques. Am J Phys Anthropol. 2006;131:33–37. [PubMed: 16485295]
  16. Nicolelis MA. Actions from thoughts. Nature. 2001;409:403–407. [PubMed: 11201755]
  17. Nicolelis MA, Dimitrov D, Carmena JM, Crist R, Lehew G, Kralik JD, Wise SP. Chronic, multisite, multielectrode recordings in macaque monkeys. Proc Natl Acad Sci US A. 2003;100:11041–11046. [PMC free article: PMC196923] [PubMed: 12960378]
  18. Nicolelis MA, Fanselow EE, Ghazanfar AA. Hebb’s dream: the resurgence of cell assemblies. Neuron. 1997;19:219–221. [PubMed: 9292712]
  19. Nobunaga AI, Go BK, Karunas RB. Recent demographic and injury trends in people served by the model spinal cord injury care systems. Arch Phys Med Rehabil. 1999;80:1372–1382. [PubMed: 10569430]
  20. Ogihara N, Usui H, Hirasaki E, Hamada Y, Nakatsukasa M. Kinematic analysis of bipedal locomotion of a Japanese macaque that lost its forearms due to congenital malformation. Primates. 2005;46:11–19. [PubMed: 15688121]
  21. Orban GA, Van Essen D, Vanduffel W. Comparative mapping of higher visual areas in monkeys and humans. Trends Cogn Sci. 2004;8:315–324. [PubMed: 15242691]
  22. Paxinos G, Huang X-F, Toga AW. The rhesus monkey brain in stereotaxic coordinates. New York: Academic Press; 1999.
  23. Pearce TM, Williams JC. Microtechnology: Meet neurobiology. Lab on a Chip. 2007;7:30–40. [PubMed: 17180203]
  24. Pesaran B, Pezaris JS, Sahani M, Mitra PP, Andersen RA. Temporal structure in neuronal activity during working memory in macaque parietal cortex. Nat Neurosci. 2002;5:805–811. [PubMed: 12134152]
  25. Polikov VS, Tresco PA, Reichert WM. Response of brain tissue to chronically implanted neural electrodes. J Neurosci Methods. 2005;148:1–18. [PubMed: 16198003]
  26. Reinhardt V. Working with rather than against macaques during blood collection. J Appl Anim Welfare Sci. 2003;6:189–197. [PubMed: 14612267]
  27. Rennaker RL, Street S, Ruyle AM, Sloan AM. A comparison of chronic multi-channel cortical implantation techniques: manual versus mechanical insertion. J Neurosci Methods. 2005;142:169–176. [PubMed: 15698656]
  28. Schapiro SJ, Bloomsmith MA, Laule GE. Positive reinforcement training as a technique to alter nonhuman primate behavior: quantitative assessment of effectiveness. J Appl Anim Welfare Sci. 2003;6:175–189. [PubMed: 14612266]
  29. Schiller PH. On the specificity of neurons and visual areas. Behav Brain Res. 1996;76:21–35. [PubMed: 8734041]
  30. Schmidt EM. Single neuron recording from motor cortex as a possible source of signals for control of external devices. Ann Biomed Eng. 1980;8:339–349. [PubMed: 6794389]
  31. Serruya MD, Hatsopoulos NG, Paninski L, Fellows MR, Donoghue JP. Instant neural control of a movement signal. Nature. 2002;416:141–42. [PubMed: 11894084]
  32. Shenoy KV, Meeker D, Cao S, Kureshi SA, Pesaran B, Buneo CA, Batista AP, Mitra PP, Burdick JW, Andersen RA. Neural prosthetic control signals from plan activity. Neuroreport. 2003;14:591–596. [PubMed: 12657892]
  33. Sperling HG, Harwerth RS. Red-green cone interactions in the increment-threshold spectral sensitivity of primates. Science. 1971;172:180–184. [PubMed: 4993975]
  34. Talwar SK, Xu SH, Hawley ES, Weiss SA, Moxon KA, Chapin JK. Behavioural neuroscience: Rat navigation guided by remote control—Free animals can be “virtually” trained by microstimulating key areas of their brains. Nature. 2002;417:37–38. [PubMed: 11986657]
  35. Taylor DM, Tillery SIH, Schwartz AB. Direct cortical control of 3D neuroprosthetic devices. Science. 2002;296:1829–1832. [PubMed: 12052948]
  36. Van Essen DC, Lewis JW, Drury HA, Hadjikhani N, Tootell RB, Bakircioglu M, Miller MI. Mapping visual cortex in monkeys and humans using surface-based atlases. Vision Res. 2001;41:1359–1378. [PubMed: 11322980]
  37. Wessberg J, Nicolelis MA. Optimizing a linear algorithm for real-time robotic control using chronic cortical ensemble recordings in monkeys. J Cogn Neurosci. 2004;16:1022–1035. [PubMed: 15298789]
  38. Wessberg J, Stambaugh CR, Kralik JD, Beck PD, Laubach M, Chapin JK, Kim J, Biggs J, Srinivasan MA, Nicolelis MAL. Real-time prediction of hand trajectory by ensembles of cortical neurons in primates. Nature. 2000;408:361–365. [PubMed: 11099043]
  39. Wise KD, Anderson DJ, Hetke JF, Kipke DR, Najafi K. Wireless implantable microsystems: High-density electronic interfaces to the nervous system. Proceedings of the IEEE. 2004;92:76–97.

Footnotes

*

Though other species of monkeys (e.g., New World monkeys) are used in neurophysiological research, the macaque monkey has become a standard model, and we will focus most of our discussion on the use of multielectrode recording techniques with macaques.

*
*

Descriptions and assessments of methods for training nonhuman primates in the laboratory have been difficult to find in the literature, and as a result, monkey training techniques have typically been developed in an ad hoc fashion within individual laboratories. For a systematic treatment of these issues, we recommend that interested readers consult a special issue of the Journal of Animal Welfare Science (Vol. 6 Num. 3, 2003), devoted to nonhuman primates training.

*
*

An extensive description of surgical procedures employed for chronic implantation of microwire arrays in Rhesus monkeys can be found in chapter 2.

*

Plexon Systems, Inc., Dallas, Texas.

*

For a more detailed discussion, see chapter 4.

*

For additional details, see the manuscript of the original paper.

Copyright © 2008, Taylor & Francis Group, LLC.
Bookshelf ID: NBK3891PMID: 21204441
Contents

Views

  • PubReader
  • Print View
  • Cite this Page

In this Page

Other titles in this collection

Related information

  • PMC
    PubMed Central citations
  • PubMed
    Links to PubMed

Similar articles in PubMed

See reviews...See all...

Recent Activity

ClearTurn OffTurn On

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...