Review: Human intracortical recording and neural decoding for brain computer interfaces

Abstract

Brain Computer Interfaces (BCIs) use neural information recorded from the brain for voluntary control of external devices. The development of BCI systems has largely focused on improving functional independence for individuals with severe motor impairments, including providing tools for communication and mobility. In this review, we describe recent advances in intracortical BCI technology and provide potential directions for further research.

I. Introduction

BRAIN computer interfaces (BCIs) use neural information recorded from the brain for voluntary control of external devices [1]–[10]. While there are many types of BCIs, all have three components: a sensor to record neural activity or its proxy, a decoder that converts neural activity into a command signal, and an effector such as a computer cursor or robotic arm (Figure 1). The sensor detects and records information. Information sources range from the locally detected electrical discharges of individual neurons, up to aggregate signals reflecting the simultaneous behavior of millions of neurons. Sensor choice impacts whether surgical implantation is required, whether or not the technology is portable or requires a specialized environment, the extent of expert assistance required to run the BCI system, and overall cost.

Fig. 1.

Illustration of a closed-loop intracortical BCI setup. The user has an intracortical electrode (Utah array) inserted into the motor cortex acting as the sensor, which then sends real-time voltage signals to the external processing system, converting raw voltages into spikes. The decoder then maps the high-dimensional spike data to low-dimensional output that is fed to the effector as computer cursor control. The real-time cursor control is then provided as feedback to the user, who is then able to modulate further neural activity accordingly.

The decoder is a mathematical algorithm that converts the complex neural data into a signal used to drive the effector. Effective neural decoding is predicated in part on advances in fundamental neuroscience (i.e. understanding the signal and noise characteristics of the sensor output) and logical algorithm choices. Finally, the effector choice reflects the use-case of the BCI. Many BCI systems are being developed to improve functional independence for individuals with severe motor limitations. Early pilot clinical trials are currently investigating the role of intracortical BCIs (iBCIs) for the purpose of restoring communication, such as typing on a computer screen [11]–[17] or as an effector for physical responses, such as controlling a robotic or prosthetic arm [18]–[21] or functional electrical stimulation systems [22], [23].

Other articles in this special edition publication describe BCI sensor technologies such as EEG, fMRI, fNIRS, or ECoG. In this review we provide an introduction to iBCI devices, as pertaining to ongoing feasibility clinical trials. We organize our discussion along each of the BCI components – the sensor, decoder and effector – and conclude by outlining the future directions for the field.

II. Intracortical recording devices as BCI sensors

The BCI sensor choice determines the temporal and spatial resolution of the neural signal [2], [8], [24]. The temporal resolution refers to the timescale over which neural signals are sampled and can be detected. For instance, in its current instantiation, near-infrared spectroscopy requires up to 25 seconds to generate a reliable binary command signal [25] whereas modern intracranial recording devices are able to record high-resolution field potentials [26] and intracortical recording devices are able to record action potentials with sub-millisecond precision. The spatial resolution refers to the volume of neuron-derived information that is aggregated together in the sensor. For instance, EEG recordings reflect changes in huge numbers of changes in neural electrical fields oriented perpendicular to the skull [27], [28], whereas implanted intracortical (or subcortical) recordings are able to focus on individual neuronal action potentials from ensembles of simultaneously recorded neurons.

Since intracortical devices record action potentials, effective decoding is predicated on understanding the functional role of the cortical area being recorded. Accordingly, we begin this section with a discussion of properties of the primary motor cortex relevant for iBCI systems.

A. Neurophysiological foundations of motor-imagery based BCI systems

Cortical regions are demarcated at a centimeter scale and are identified by their location and function [29]. For the purposes of BCI decoding in controlling assistive technologies, many researchers have focused on decoding neural activity from the motor cortex in both non-human primates (NHPs) [30]–[42] and early pilot clinical trials [11]–[15], [18]–[20], [23], [43]–[45].

Decades of basic science research and clinical findings have motivated the use of the motor cortex as a source of neural information for BCIs. It is clear that the primary motor cortex plays a key role in directing voluntary, dexterous movement [2], [7], [46]–[51], although successful decoding can also be achieved from other cortical regions [21], [52], [53]. First, neurons in a small area of primary motor cortex yield a rich variety of signals related to voluntary limb movement. Experiments have demonstrated that neurons modulate activity based on arm position [54], [55], velocity [56] and acceleration [57]; kinetic parameters [58] such as force [59] muscle activation [49], [60]; preparatory activity [44], [48], [61], and even task-related parameters such as instruction cues [62] and movement observation [63]. Importantly, neurons recorded on adjacent electrodes separated by micrometers [64] may display highly variable behavior [50], [65], meaning that recordings from small implantable electrodes may have the potential to sample this rich variety of behavior. Thus, there is tremendous potential to acquire a variety of task-relevant signals for subsequent decoding.

Second, in humans, the hand area of motor cortex is an accessible surgical target. It is easy to identify it pre-operatively using known anatomical landmarks [66], [67], and as part of pre-operative functional MRI protocols [68]–[70].

Third, neural modulation is present in motor cortex despite longstanding downstream injury of pyramidal neurons. This point is non-trivial: destruction of a neuron’s axonal projection triggers degeneration of the axon in a process called Wallerian degeneration, which can lead to cell death [71]. Despite controversies in the non-human primate literature as to the extent of corticospinal destruction resulting from downstream axonal damage [72], [73], recent successes in motor cortex decoding in people with paralysis suggest that sufficient neural information for BCI use [11]–[15], [18]–[20], [43]. Thus, it appears as though motor cortex retains relevant physiological properties despite longstanding damage to the motor system. This includes people with amyotrophic lateral sclerosis (see below), from whom motor cortical recording has provided control over external devices despite disease-related injury or death of pyramidal cells [74].

B. Intracortical recording devices

While there are multiple intracortical recording devices available [75], the intracranial device capable of single unit recordings currently used as part of ongoing clinical trials [2], [19], [21] and for recordings in individuals with epilepsy [76]–[78] is the Utah microarray (formally, Blackrock/NeuroPort array) [64]. Each array is a 4.2 × 4.2mm wafer with 100 shanks, each of which is 1.0 – 1.5mm in length spaced apart by 400μm. At the tip of each shank is a conducting electrode [79]. Extracellular voltage recordings are transmitted from the electrodes to a skull-mounted titanium pedestal via gold/platinum alloy cables. The pedestal has connector pins to which a cable can be attached to connect the device to external amplifiers and then computers for analysis [80].

The process of inserting multiple electrodes simultaneously into the brain is not trivial. While the brain is enveloped in the dura mater, the cortical surface is enveloped by both arachnoid and pial meninges. The pia mater in particular, is intimately attached to the cortical surface. In animal models, unassisted insertion of multiple electrodes on an array can cause cortical deformation, and may even result in cortical injury – thereby damaging the very neurons one is interested in recording [81], [82]. To circumvent this problem, Utah arrays are typically inserted into cortex using a pneumatic insertion device [83]. The pneumatic device attempts to mitigate insertion trauma and ensure the electrodes are inserted at the desired cortical depth, in hopes of optimizing signal quality.

While the standard design of the Utah array allows for a single electrode per shank (with customizable shank lengths), other recording devices have been designed to allow multiple electrode per shank, thereby permitting recordings at various cortical depths simultaneously [84]. Another approach, the Neutrotrophic electrode, contains an insulated gold wire fixed within a glass cone, containing a medium for attracting neurite growth [85].

The interface between brain parenchyma and electrode recording device is complex, and the specific material considerations may dramatically impact recording characteristics [75]. For instance, factors such as material pliability, shank/electrode alloy selection, electrode coating techniques, array insertion techniques, and other factors may impact recording quality [86]–[90].

III. BCI Decoding

We described BCIs as devices that use neural information recorded from the brain for voluntary control of external devices. The process of actually converting the neural information (often high-dimensional) into the output signal (usually low-dimensional) is referred to as neural decoding. We begin by describing how neural features are acquired from raw voltage signals. We then outline how modern neural decoding algorithms transform neural features into the command signals used to control external devices.

A. Distinguishing between open- and closed-loop systems

To understand BCI decoding, it is important to distinguish between open- or closed- loop decoding. An open-loop system is one where no feedback is provided back to the user. For instance, by recording from several dozen neurons in three motor-related cortical areas in NHPs, the complex movement of 25 joint angles can be faithfully reconstructed for complex reaching and grasping motions [91]. This movement reconstruction maps action potential events from dozens of neurons (high-dimensional) to kinematic reconstructions (low-dimensional). However, the result of the low-dimensional decoder output was not provided as feedback to the NHP at the time of experiment, thus making this an open-loop experiment. By contrast, a closed-loop system is one where the decoded output is provided as feedback to the BCI user, allowing the user to adjust volitional control in real-time. A now-classic example of closed-loop control was performed by Fetz [92]. An electrode was placed into the motor cortex of a NHP and neurons were found that responded to wrist torque; the firing rate of a neuron was used to drive an auditory signal. With practice, the NHP learned to modulate the firing rate of the neuron in real-time to earn a reward. Since the output of the effector (the auditory signal) was provided as feedback in real-time and used to drive further neural behavior, this was a demonstration of closed-loop single-neuron BCI control used to control an external device.

B. From voltages to spikes

Modern microelectrode arrays sample voltages at several thousand times per second and are designed to record single neuron action potentials through each channel. Since action potentials are voltage waves with canonical shapes and wavelengths [71], raw signals are band-pass filtered (typically, 0.5–7500Hz) and action potentials are identified as large sudden changes in the voltage (typically, 3–5 RMS Volts) lasting around 1ms. Due to the characteristic appearance of band-pass filtered signals, action potentials are also known as spikes.

As an alternative to spikes, local field potentials (LFPs) are acquired when raw voltage signals are band-pass filtered in lower frequency ranges (usually in the 0.1 – 200Hz range). LFPs reflect spatially averaged electrical fields generated from groups of neurons simultaneously [26], [93], and are important for coupling different regions of cortex simultaneously, and impact spike timing [94]–[96]. Specific frequency bands are relevant for motor control [97], [98]. LFPs have also been used for neural decoding with intracortical electrodes in NHPs [99]–[103] and in humans [104], [105].

C. Linear neural decoding strategies are predicated on cosine tuning

In what are now classic experiments, NHPs were taught to move a planar manipulandum to one of eight different directions. The spike rate of individual neurons and the direction of limb movement could be described parsimoniously using a sinusoidal curve with a period of 360 degrees [54]. For each neuron, the vector corresponding to the maximum firing rate (the phase offset) is referred to as the neuron’s “preferred direction”. To perform neural decoding using the population vector algorithm (PVA), the firing rate of each neuron is used to scale its preferred direction vector, and the sum of all vectors is used as the low-dimensional output. PVA is a special case of solving a linear regression problem: if the preferred directions are uniformly distributed, then the surface mapping the binned firing rates to the kinematic variables can be solved by solving for the least-squares estimator [106]. Motivated by the sinusoidal tuning properties of motor cortex neurons, the PVA/linear regression algorithm was shown to be successful in NHPs [30], [31], [33], [37], [38], and was modified [107] to allow an individual with tetraplegia to control a robotic limb [18], [19].

Firing rates recorded from a single neuron may be highly variable despite very similar (or even identical) experimental repetitions[108]–[110]. One approach to addressing noise may be smoothing inputs [11], [111] or by using a Weiner filter. Some experimenters have taken the approach of modeling the firing rates as being the “read-out” of a time-evolving hidden variable. This is tantamount to modeling the system as a hidden Markov model, where the observed variables are the firing rates and the hidden variable is modeled as the kinematics of the end-effector. By taking advantage of the cosine-tuning properties of neurons, the Kalman filter provides a principled approach for real-time hidden variable estimation [112], [113]. The Kalman filter has been implemented in early clinical research settings during to provide users with cursor [12]–[16], [43] and robotic limb control [20].

D. Non-linear decoding approaches move away from cosine tuning

The aforementioned decoding approaches -- the PVA, linear decoder, and Kalman -- all rely on binned spike data as input, assuming a rate code. An alternate way of approaching neural decoding is to model precise spike times as important. Rather than modeling neural behavior as a time-evolving rate-function, spike times may be seen as a realization of a point processes with an inhomogeneous (i.e. time-varying) conditional intensity function [35], [114]–[117]. Approaching spike trains in this way allows the experimenter to model how the intensity function varies as a function of: (1) the neuron’s own intrinsic spike history, (2) the behavior of other recorded neurons, and/or (3) external variables of interest [117]. Recent decoding approaches have used point process methods in both NHP [35], [116] and human subjects [117].

Future decoding approaches will undoubtedly take advantage of ongoing developments in the machine learning literature. For instance, recurrent neural networks were used as a decoder in NHPs using high dimensional neural input along with an element of temporal history [118], [119]. Neural decoding has also been done with kernel embedding of neural data [40], [120]–[123].

E. Adapting the user to the decoder vs. the decoder to the user

Two approaches for permitting BCI control have been described [124]. In the wrist torque experiments described above, a NHP learned to modulate firing rates based on sound feedback [92]. Here, the decoder was fixed (mapping firing rate to pitch) and the NHP had to adapt its behavior in order to use the system. Thus, cortical activity adapted to improve decoder use [6]. In fact, NHPs can learn to use sub-optimally parameterized decoders, achieving effective closed-loop control over several days [34], [125]. Moreover, when NHPs are taught to use a particular decoder and the parameters are changed, they are able to adapt to the new parameterizations[38].

Having users adapt their cortical activity to control external devices has been used in a variety of BCI environments [17], [40], [92], [125]–[130]. However, the process of developing reliable control requires days, weeks, or even months to learn to use. By contrast, the second approach -- which has been explored in the iBCI literature in NHPs and humans -- has been to adapt the algorithm to the user. Here, the decoder parameters are computed in a calibration process. One approach to calibration is to record from the cortex while the NHP is performing stereotyped movements [30]–[33], [131]. However, this approach cannot be applied in humans with paralysis.

The state-of-the-art approach to calibration of motor-imagery BCI devices in humans is a three-step process [11]. Users are first asked to attempt, or to imagine, performing an action -- without feedback -- in an open-loop step. Next, a trained technician uses the recorded data to compute decoder parameters. Finally, the user is provided closed-loop control. Decoders built solely from open-loop data yield sub-optimal control in closed-loop decoding, due to context shifts of the neural tuning properties [43], [132], [133]. Accordingly, modern state-of-the-art approaches to decoder calibration in humans require technician supervised closed-loop batch-based calibration sequences [11]–[14], [16], [20], [43], [134]. An alternative to requiring batch-based technician supervision is to perform closed-loop decoder adaptation. Here, the parameters for the decoder are updated as the user is controlling the end-effector in real-time [6], [35], [36], [124], [135]–[137]. Using this approach, NHPs can rapidly gain control of a decoder using a randomly initialized set of decoder parameter settings within minutes [135]. However, not all parameterizations will allow NHP to successfully achieve BCI control, and evidence suggests that low-dimensional features of neural ensembles must be maintained in order for NHPs to be able to learn to use decoders [38], [40].

IV. The BCI effector and its uses in restorative technology

Following the discussion of sensors and decoders, we now turn our attention to discussing the effectors. We focus our discussion for two different disease states in humans the challenges of which stand to be reduced through intracranial recordings.

A. Communication using single neurons by people with locked-in syndrome

The motor system is required for strong, coordinated and fluid voluntary movement; damage occurring at any point in the system may affect motor control. Limb weakness may result in paralysis of the legs (paraplegia) or four limbs (tetraplegia). An even more extensive form of paralysis is referred to as locked-in-syndrome (LIS), wherein individuals lack the ability for most voluntary movements (which usually includes limb, facial movements and speaking), though the individual may remain cognitively normal. One of the most common causes of LIS is amyotrophic lateral sclerosis (ALS). There are as many as 30,000 individuals with ALS in the United States, with as many as 5500 new cases diagnosed each year [138]. Individuals with LIS are dependent on caretakers for all aspects of their wellbeing. Despite the dramatic limitations in motor control, studies indicate that quality of life markers for individuals with ALS consistently demonstrate positive life attitudes [139], [140].

Research in LIS has focused on communication – an identified priority for those living with ALS. A recent Japanese mail-back questionnaire survey was conducted with 37 individuals living with ALS, 29 of whom had undergone tracheostomy for assisted respiratory support [141]. Nearly all respondents described feeling stressed when communicating, and most of the respondents who were already using assisted communication devices complained of difficulties using their present devices. The respondents who were interested in using BCI as a communication tool prioritized conversations with family or caregivers, as well as electronic communication.

BCI technology is a viable approach to providing tools for communication for individuals with ALS [13]–[16]. Kennedy and colleagues were the first to implant electrodes into an individual with ALS [128] using a specialized electrode with two recording wires and a neurotrophic factor [85]. While the first report demonstrated closed-loop control of neurons adjusting auditory signals, users subsequently demonstrated the ability for neurons to drive a computer cursor in one direction[142], and to use a wireless interface driving a one-dimensional signal [143].

The first example of reliable multi-dimensional cursor control in humans using a population of neurons in motor cortex was reported in 2006 [11]. Two individuals with paralysis due to cervical spinal cord injury demonstrated the ability to move a 2-dimensional neural cursor, play “pong” (similar to the classic video game but controlled entirely though neural spike trains) and one controlled a rudimentary robotic device. The feasibility of long-term intracortical recording as a communication device was demonstrated when a participant demonstrated successful decoding in an implanted array over 1000 days[13] and subsequently more than five years after implantation [20]. Recently, customized keyboards have been developed to take advantage of the directional cursor control, enabling typing speeds of over 10 words per minute using Internet chat applications[16]. Through this and other advances, it is hoped that intracortical recording will provide the basis for individuals with LIS and ALS to communicate readily [144].

B. Restoration of arm control is a priority for individuals with tetraplegia

There are millions of people worldwide[145] and roughly a quarter-million people in the United States United States[146] who live with a traumatic spinal cord injury (SCI). Younger individuals tend to develop traumatic injuries from high-impact events (e.g. motor vehicle collisions, diving accidents, etc.) whereas older patients develop injuries from low-energy falls, with concomitant degenerative spine disease. Of those who suffer a traumatic SCI, approximately half will develop some degree of tetraplegia[146]. A recent survey of individuals with SCI in the United States identified their specific functional needs[147]. For individuals with tetraplegia, the ability to regain arm and hand function was the most common outcome selected from seven different physiological outcomes. A more recent systematic literature review including 3187 patients also supports this finding[148].

BCI technology is well suited to realize the needs identified in the aforementioned surveys. Several years after the first demonstration of cursor control for communication, implanted electrodes were used to allow an individual who was incompletely locked-in after a brainstem stroke the ability to control a robotic limb using decoded spikes. For the first time in nearly 15 years, the individual was able to use an assistive technology to reach out and grasp a bottle, bring it towards her mouth and drink from it[20] (Figure 3). This research was then independently replicated and expanded when a 52 year-old woman diagnosed with spinocerebellar degeneration was able to control a robotic limb with seven independent degrees of freedom to feed herself [19]; and, when a 32 year-old man with a SCI was able to control a limb using recordings from the posterior parietal cortex[21]. More recently, a brain-spinal interface was demonstrated in NHPs [149], highlighting the potential for intracortical recording to drive restorative approaches for people with paraplegia. Finally, there have been two recent examples of intracortical recording driving a surface FES device in non-human primates [150]–[152] and humans with [22] and an implanted, percutaneous FES system [23], the latter successfully providing reanimation of elbow flexion/extension and hand grasp plus robot-assisted arm elevation in a 53 year-old man with high cervical spinal cord injury.

While the aforementioned demonstrations of robotic and native limb control are at a relatively primitive stage, they are proof of principal that neural information can be used for multi-dimensional limb control. These results demonstrate both the feasibility and potential utility from intracortical recordings in humans.

V. Future directions

Most consumer devices undergo iterative improvements, and this is especially true for successful medical devices [153]. The original deep brain stimulators and cardiac pacemakers began as experimental devices with protruding wires, bulky electronic systems and limited efficacy. Iterative engineering solutions and partnerships between academia and industry allowed these devices to become miniaturized and fully implantable; both devices are now part of the standard of care for their respective disease states and have been implanted in patients around the world.

The development of reliable closed-loop BCIs for people with paralysis benefits from multidisciplinary collaborations across a variety of specialties, including neuroscientists, clinicians, engineers, computer scientists, and applied mathematicians. Keeping in mind that the guiding principles for device development ought to be based on the needs of those who will be using them [141], [147], [148], [154], [155], we highlight several directions for innovation in the near future.

First, much like deep-brain stimulators and cardiac pacemakers, the next generation of devices that will provide robust and reliable decoding of neural information should be fully implantable (i.e. without anything protruding through skin). Recent prototypes have been developed to wireless transmit real-time spike data through a hermetically sealed and fully implanted titanium system[156], [157]. This device has been successfully used in NHPs and is being transitioned toward human medical device manufacturing.

Second, the current devices used for closed-loop control of cursors and robotic limbs all require a dedicated technician to oversee their use [18]–[21]. A major engineering achievement will be the development of a robust and reliable closed-loop algorithms that works 24-hours a day with minimal outside intervention from trained technicians[15].

Third, while cursor control for communication has now been achieved in multiple individuals[15], [16], the ability to control complex robotic arm movements is still limited and has not yet reached the point of being a viable replacement for naturalistic control [18]–[21]. This is, in part, due to algorithmic deficiencies. For instance, algorithmic decoder advancements that incorporate novel neurophysiological insights are more likely to capitalize on the useable signals available through single cell recordings and optimize control [58], [124]. Decoding algorithms may develop strategies to continuously update parameters to deal with system noise [15], [35], [124], [135], [158]. Algorithms may be developed to determine what parameters the brain will be able to adapt to use [38], [40], and what features necessitate recalibration of the decoding parameters[6]. These innovations will also help to further the restoration of native limb movement, by iBCI control FES systems, with the goal of restoring arm/hand function and quality of life to individuals with cervical spinal cord injury or brainstem stroke [22], [23].

Fourth, current cursors and robotic limbs do not provide somatosensory information back to the user. In normal human reaching and grasping, continuous feedback is provided via proprioceptive, visual and tactile information. Closing the motor-sensory loop beyond the currently available visual feedback would likely enable movements closer to that achieved by people without physical disability [58], [159]. In the case of ALS, the sensory system is generally left intact, and devices could be developed in order to mimic proprioceptive information, such as by providing tendon stimulation [71]. By contrast, in complete SCI, sensory information from limb movement is impaired. One approach would be to directly stimulate the cortex in order to mimic proprioceptive information. Early success with intracortical stimulation in humans [160] suggests this may be a viable approach in the future.

Finally, developing clinically viable tools is predicated on demonstrating efficacy. Without being able to demonstrate reproducible, clinically viable outcomes, it will be difficult to bring these devices to those who need them most. With the field of clinical BCI for communication and motor rehabilitation still in its infancy, consensus is required as to how to measure outcomes appropriately [161].

VI. Conclusion

Intracortical brain computer interfaces have the capacity to improve the quality of life for individuals with severe motor deficits. Recent successes in early pilot clinical trials have demonstrated the ability for people with paralysis to control computer cursors and robotic limbs using only the intention to move one’s limb. Through multidisciplinary collaborative efforts, these devices will continue to evolve and we expect we move steadily toward meeting critical and ever-more complex needs for people with paralysis.

Fig. 2.

BCI control of a prosthetic limb. Enrolled in the BrainGate2 pilot clinical trial (https://www.clinicaltrials.gov/ct2/show/NCT00912041), Participant S3 reached out using a prosthetic arm and drank coffee for the first time in over a decade.

Biographies

David M. Brandman studied biophysics at the University of Brisith Columbia, Vancouver, BC, Canada, and received his medical doctorate from the University of Calgary, AB, Canada. He is a Ph.D. Candidate in Neuroscience at Brown University, Providence, RI, USA. He is currently a senior neurosurgical resident at Dalhousie University, Halifax, NS, Canada.

Sydney S. Cash received the B.S. degree from Yale University, New Haven, CT, USA, and the M.D. and Ph.D. degrees from Columbia University, New York, NY, USA. Afterwards he did an internship in medicine at Massachusetts General Hospital (MGH), Boston, MA, USA, and a Neurology Residency in the combined Massachusetts General Hospital Brigham and Women’s Hospital Partners program, Boston, MA, USA. This was followed by fellowship training in Epilepsy at MGH. He is now Associate Professor in the Department of Neurology and Assistant Director for Research in the Partners Neurology Training Program as well as co-Director of the NeuroTechnology Trials Unit in the department and Clinical Trials Director of the New England Pediatric Device Consortium, Lebanon, NH, USA.

Leigh R. Hochberg received the M.D. and Ph.D. degrees in neuroscience from Emory University, Atlanta, GA, USA. He is Associate Professor of Engineering, Brown University; Associate Director, VA Center of Excellence for Neurorestoration and Neurotechnology, Providence VA Medical Center; and Senior Lecturer in Neurology at Harvard Medical School. In addition, he maintains clinical duties on the Stroke and Neurocritical Care Services at the MGH and Brigham and Women’s Hospital, and is on the consultation neurology staff at Spaulding Rehabilitation Hospital. He directs the Laboratory for Restorative Neurotechnology at Brown and MGH and directs the BrainGate pilot clinical trials. His research focuses on the development and testing of implanted neural interfaces to help people with paralysis and other neurologic disorders.