|Home | About | Journals | Submit | Contact Us | Français|
Neural prosthetic systems aim to help disabled patients by translating neural signals from the brain into control signals for guiding computer cursors, prosthetic arms, and other assistive devices. Intracortical electrode arrays measure action potentials and local field potentials from individual neurons, or small populations of neurons, in the motor cortices and can provide considerable information for controlling prostheses. Despite several compelling proof-of-concept laboratory animal experiments and an initial human clinical trial, at least three key challenges remain which, if left unaddressed, may hamper the translation of these systems into widespread clinical use. We review these challenges: achieving able-bodied levels of performance across tasks and across environments, achieving robustness across multiple decades, and restoring able-bodied quality proprioception and somatosensation. We also describe some emerging opportunities for meeting these challenges. If these challenges can be largely or fully met, intracortically based neural prostheses may achieve true clinical viability and help increasing numbers of disabled patients.
Neurological injury and disease often result in the permanent loss of motor function. In many cases, the disability is so severe that it is not possible to feed oneself or even communicate (e.g., ). Though surgical and pharmacological interventions have made it possible to repair nerves and promote recovery in the peripheral nervous system, most central nervous system impairments still do not have effective treatments. Electronic medical systems that interface with the nervous system, termed neural prostheses (also referred to as brain–machine interfaces or brain–computer interfaces), have started to fill some of these treatment gaps. Recent successes include cochlear implants for the profoundly deaf and deep brain stimulators to alleviate Parkinsonian tremor. In the relatively near term, epileptic-seizure disruption systems, artificial vision systems, and computer interface (communication) and prosthetic arm (motor) systems are believed to be possible (e.g., –).
In this perspective and review paper, we focus on communication and motor prostheses where a patient’s expectation is to be able to type, move a computer mouse, or move a prosthetic arm as swiftly and as accurately as when able bodied –. Though complete restoration of lost neurological function is ideal, prostheses can become clinically viable when the anticipated quality-of-life improvement (benefit) outweighs the potential risks (burden). As such, considerable attention is being paid to maximizing benefit while minimizing burden , . We briefly describe later the range of measurement techniques currently available, the need for quantitative assessment and comparison of benefits and burdens, and how the benefit-to-burden ratio is of central interest. We then turn our attention to the need for, and possibilities for, increasing performance, robustness, and sensory restoration in intracortically based neural prostheses.
Sensors which are external to the body, such as electroencephalography (EEG), magnetoencephalography, and functional near-infrared imaging, have been investigated extensively in recent decades (e.g., ). More recently, implantable electrode-array based techniques such as epi- and subdural electrocorticography (ECoG) and finely spaced ECoG electrodes (μECoG) have become a major research area. The closer proximity to the neural signal source should result in higher signal quality, which should in turn enable higher prosthetic performance (see ,  for recent reviews). Over roughly the past decade, neural prostheses based on intracortical electrode arrays have also been investigated fairly extensively due to their ability to record action potentials from individual neurons, as well as local field potentials from small clusters of neurons (e.g., , ). This is possible since each intracortical electrode penetrates one to several millimeters into the brain (typically cerebral cortex), bringing the electrode tip into close proximity of the neurons of interest. Higher performance should result since these individual neurons are widely believed to be the fundamental information encoding units in the nervous system.
Thus, it is anticipated, but remains to be fully realized and demonstrated, that intracortical electrode arrays which have access to the highest temporal (individual action potentials, millisecond timescale) and spatial (individual neurons, micron spatial scale) information possible should enable the highest performance and capability prosthetic systems (e.g., , ). If increased performance, capability, and robustness—as well as the new capability of high fidelity sensory “write in”—can be realized with intracortically based neural prostheses, this could offer an important option for patients and their treating physicians to consider.
Importantly, internal sensors such as ECoG and intracortical electrodes require a surgical implantation procedure, introducing risks to a patient’s well-being that would not be present with nonsurgical options. However, the use of any medical intervention is associated with potential negative influences on wellbeing that may not be as obvious. For example, many patients would opt for a single, risky, potentially debilitating surgical procedure if it meant that they would be free from taking medication several times a day. These interventions (medication and surgery) may thus be thought of as nearly equivalent in this example, or indeed, the “riskier” option might be preferable, based on the impact on the patient’s daily well-being. We propose the term “burden” to represent this aggregate negative impact on well-being. Surgical risk or invasiveness represents only one part of this burden.
Burden might be apportioned in various ways: “up front” burden in the form of surgical implantation risk; daily burden in the case of donning, doffing, recharging, or operating a complex system; cosmetic burden; financial burden; and others that could possibly include the burden on caregivers and society at large. Some degree of burden is unavoidably associated with the use of any assistive technology. Balancing the burden of technology use versus its benefits is a vital part of the decision-making process between physician and patient.
Fig. 1(a) provides one possible way of visualizing risk-benefit for current and future neural prostheses, with “performance” as a surrogate for benefit and “burden” as a measure of overall risk. At present, this can only be a notional sketch as the field is currently striving to conduct experiments that will better allow various neural prosthesis technologies to be placed quantitatively within this design space. The current lack of standardization of tasks (e.g., selecting keys on a keyboard, moving around obstacles), environments (e.g., quiet room with no distractions, or the opposite), or metrics (e.g., time to target, bits per second, how to quantitate burden) used or reported in the literature makes quantitative comparison among various laboratories results virtually impossible. However, if standardization could be agreed upon (e.g., ISO 9241-9 for performance , clinical consensus on defining burden), this type of performance versus burden plot could become powerful in helping to guide both clinical decision making and further research directions. Clearly, the optimal system would incorporate the lowest burden with the highest performance. However, absolute performance and absolute burden are likely not very meaningful on their own. It is their ratio that is important. In the end, it is this ratio that will help guide clinicians and patients when discussing appropriate treatment options.
Finally, it is widely recognized that there is no “one size fits all” when it comes to benefit-risk ratio in medical treatments and decisions, and this is also the case for neural prostheses. Interpreting the benefit-risk ratio associated with each treatment option is a personal decision. Some patients will not elect to have surgery under any circumstance, preferring the daily burden of system maintenance to the up-front burden of surgical risk. Thus, completely nonimplantable technologies (e.g., EEG) are clearly extremely important and must be pursued. Other patients seek the restoration of able-bodied levels of control of, for example, prosthetic arms and hands which have tens of degrees of freedom and extreme performance requirements, and thus may be willing to accept surgically implanted technologies, which carry more upfront risk (e.g., intracortical arrays). At the present time, nearly all neural prosthesis technologies carry significant burden that prevents their widespread adoption. There are many opportunities to improve the performance (see Section III-A) and robustness (see Section III-B) of intracortically based systems to levels that could potentially allow their consideration as a clinically viable assistive technology.
We now consider three key challenges for next-generation intracortically based prostheses. While other challenges certainly exist, for example the need for transmitting information wirelessly from the neural sensors to a nearby prosthetic device –, such needs are not specific to intracortical array-based systems and, therefore, are not considered here.
Communication prostheses aim to provide the ability to select from among discrete targets, for example keys on a keyboard, in order to restore the ability to communicate. Performance for such systems is readily defined as the speed with which keys are selected times the accuracy with which keys are selected. More formally, this is known as bit rate (bits per second, b/s) and is often reported (e.g., ). The quest for higher performance led us to propose that “plan” or “preparatory” motor neural activity, which is present in posterior parietal, dorsal premotor (PMd), and primary motor (M1) cortex before movement (e.g., ), or even without any resulting movement, could be used in conjunction with a maximum likelihood (or other) discrete decoder to predict which one of the various possible targets should be selected . This approach was born out in experiments with rhesus monkeys , and further optimization analyses and experiments demonstrated what is, to the best of our knowledge, the highest level of performance reported to date (6.5 b/s, or approximately 15 words/min) . This is on par with able-bodied single-finger typing rates on an iPhone for example [Fig. 1(a), green] .
More recently, we have attempted to further increase performance by investigating multiple algorithmic avenues. First, we believe that it is possible to readily detect when a target (key) is being planned, so as to enable freely paced selections , . Second, we found that it is possible to use the neural activity on each (specific) electrode array to determine where on the computer screen the targets should be placed, so as to optimally engage the information available in the neural population . Third, we discovered that estimating eye position from the neural activity, in addition to the desired arm movement, effectively reduced noise and increased performance . Fourth, we designed a factor-analysis-based decoder that is able to segregate out both fast and slow time-varying activity fluctuations, which are nominally present in all neural recordings, resulting again in increased performance . Finally, we investigated a variety of more optimal decoding and firing rate estimation algorithms, some of which can increase performance  and some of which revealed that present methods appear adequate . Taken together, it would appear likely that free-paced discrete decoding systems for communication prostheses could be demonstrated to push well beyond 6.5 b/s.
Motor prostheses enable moment-by-moment control of a computer cursor or an end effector such as a prosthetic limb. In the case of a computer cursor, motor prostheses have the potential to provide computer mouse like functionality. These systems also typically target motor cortices for electrode implantation, although additional areas have been investigated . In experiments with rhesus monkeys, intracortical motor prostheses have demonstrated control of computer cursors in both 2-D – and 3-D  as well as control of prosthetic limbs (3-D and hand open/close) . Results from recent clinical trials of these systems with tetraplegic patients provide reasonable confidence that these proofs-of-concept and subsequent advances with rhesus monkeys will continue to translate readily to humans , .
These initial demonstrations of intracortical motor prostheses are compelling; however, their performance is far from that of the native limb. Thus, over the past decade research groups have focused on methods for improving performance. Roughly, this study falls into three categories: adding more sensory information, learning studies, and neural decoder design, which are complimentary because all three approaches could likely be used together to achieve even higher performance. Adding more sensory information, beyond vision alone, is a relatively new approach which reasons that by providing the full compliment of sensory signals present during natural arm movements the performance and capability of prosthetic arm movements should increase (e.g., ; see Section III-C). Learning studies aim to measure the neural system’s ability to adapt to a neural prosthesis over short (hours) ,  and long (days) ,  timescales. Importantly, these studies suggest that relative performance, with respect to the first hour or day of the experiment, can indeed improve considerably over time as the user and neurons adapt to operating with the motor prosthesis.
In contrast, neural decoder design attempts to maximize performance of motor prostheses from their very first use onward. Work in this area falls into two subcategories: off-line reconstruction and online control. In off-line work, advances are designed and tested with prerecorded neural and behavioral data and the goal is to reconstruct the (natural arm) behavioral data from the neural observations –. Improved off-line reconstruction may not always translate into better online control, as system properties may change once the user is in closed loop with the neural decoder and the computer cursor (i.e., the new prosthetic plant has different dynamics than the natural arm) . Thus, to fully vet potential algorithmic advances they must be tested online, as part of closed-loop prosthetic experiments –, , .
However, online control experiments are more costly than offline reconstruction studies, both in terms of physical resources and time. Two intermediate approaches have been proposed recently that attempt to bridge the gap between off-line and online testing. One method is to design off-line simulations that attempt to better model the properties of online control . Another method is to construct a human-in-the-loop “online prosthetic simulator” (OPS) wherein the subject’s arm movements are used to create synthetic neural data that then course through a decoder to move a cursor on a screen . By varying the number and noise characteristics of the synthetic neural signals, as well as the type of decoder and its parameter settings, it may be possible to discover decoders and parameter settings particularly well suited for online control across a range of neural signal properties. This approach appears promising as an initial OPS study employing a Kalman filter suggested that extremely brief neural integration times, in contrast to somewhat longer integration times indicated by traditional off-line reconstruction studies, should result in better online control and this has been born out in recent online control experiments with monkeys –.
In our recent online closed-loop neural decoder study with rhesus monkeys , we focused on issues of closed-loop design. By viewing motor prostheses as a feedback control system, we made design decisions that alter existing decode algorithms and improve prosthetic performance on a task requiring the user to acquire and hold targets [e.g., achieving 70–80% of normal reaching speed, approaching the green region in Fig. 1(a)] . These performance enhancements generalize to more complex tasks, including an obstacle avoidance task, and also increase robustness . In order to measure performance gains, we held task parameters constant as we made algorithmic changes, allowing us to compare performance metrics across algorithms.
Ideally, performance could be compared across the published literature; however, task designs and parameters vary considerably [see Section II and Fig. 1(a)]. By application of Fitts’ law, which is used to evaluate the behavior of standard computing interfaces such as mice or track pads, it may be possible to compare performance when target sizes and distances are altered (e.g., ). However, these comparisons can be confounded by differences in required target hold times. Tasks with small or nonexistent hold times (and thus no “click/select signal”) are fundamentally different than tasks with significant hold times (or another click/select signal). With no hold time, the neural cursor simply needs to run through the target; hold times require the cursor to slow down or stop when on target. Both tasks require an optimization between speed and accuracy, but hold time tasks add complexity by requiring stopping ability. Another potentially subtle difference is that for some tasks the neural decoder is reinitialized on every trial, while in others the decoder is free-running across trials. This choice can have a large impact on performance, as reinitialization provides more consistency across trials for the users, but consequently does not generalize to more complex tasks.
Incongruities in task design not only make it difficult to assess algorithmic advances, but also obscure comparisons between different signal sources [see Fig. 1(a)]. One possible solution is to standardize task design in an effort to provide consistent quantification . Another possibility is to begin directly assessing these technologies as computer interface devices, as defined by ISO 9241-9 . By mapping neural prostheses to these standards, the field will formally define new computer mouse and keyboard equivalents, providing quantifications of performance that can be directly compared to standard computer mouse and keyboard systems. It is important to note that existing computer interface standards are only a starting point, as neurally controlled systems are likely to be more sensitive to additional factors, such as environmental distractors, than traditional interfaces.
The monkey studies and methods described earlier are restricted to neural control while head position and body posture are relatively stationary, and the cognitive load of the environment is low. While these animal model studies appear to translate readily to immobile tetraplegic patients , , it is not clear that our understanding of motor/prosthetic cortical control will generalize when the user is capable of walking and moving freely around a real-world environment, as would be the case if such technology was applied to amputees. Fortunately, recent advances in electrophysiology techniques are enabling the study of motor control during free behavior in naturalistic environments . These studies have the potential to expand our understanding of the benefits of neural prostheses for a larger patient population across tasks and environments.
Another key challenge for intracortical based neural prostheses is robustness, which in turn depends on recording stability. By robustness we mean the ability to perform well, as discussed earlier and across multiple tasks and environments: 1) across several hours as we do daily when using a computer mouse, which is possible if the system works well enough that it is not “frustrating” to use; and 2) across several years, which is possible if the neural signals from the electrode array do not “disappear” (see Fig. 1(b), green line).
Previous intracortical work has focused primarily on decoding information from the activity of single neurons (e.g., , , – ). This has two major difficulties. First, as schematically illustrated in Fig. 1(b) (solid red line), waveform amplitudes can change dramatically from day to day, and even within a day and even while seated quietly , . Fig. 2(a) shows an example of an electrode at various time points recorded over two weeks using a wireless system . While the cause of these fast changes are not known for certain, they could be due to electrode movement termed micromotion , , or changes in the extracellular impedance (e.g., ). If these changes were limited to moderate amounts of signal change, due to either slow diurnal cycles or even occasional fast acceleration-induced shifts , it ought to be possible to use adaptive algorithms to track these changes . This could then be combined with low-power spike sorting circuitry ,  with the eventual goal of full implantation, and thereby retain the highest level of information extraction possible. However, an even more serious problem is that the amplitudes of single units can decay dramatically over the course of months or several years, also schematically illustrated in Fig. 1(b) (dashed red lines). Fig. 2(b) shows an array for which this happened particularly fast. The peak-to-peak voltage of the neuron with the largest action potential, on each electrode initially showing 200 µV or greater single unit activity, declined by 39% from 17 to 50 days after implantation (see Fig. 2 legend for details). This decline may be associated with the formation of scar tissue, engineering or material failures, or micromotion.
Despite these problems, it is important to recognize that neural prosthetic system performance may be less dependent than previously thought on single-neuron action potential amplitude and waveform isolation quality. This is likely an underappreciated point. Several studies have found only a small difference in system performance when decoding from sorted single units rather than unsorted multiunits garnered by setting a threshold level and counting threshold-crossing events , , . Fig. 2(c) shows off-line decoder performance based on threshold-crossing event counts from the same array as in Fig. 2(b), during this period of rapid single unit decline. The ability to predict which one of four targets (directions) the monkey was reaching to showed no significant change over this time period . Similarly, using a continuous linear decoder off-line, there was no significant change in two other performance metrics (correlation coefficient with hand position and mean distance to target; data not shown). This array continues to be used for online high-performance neural prosthetics experiments using threshold crossings over a year after implantation –. A second animal participating in these ongoing experiments was implanted approximately three years ago, and comparable online performance ,  as well as off-line performance (average correlation coefficient with hand position of 0.85; linear decode) can be achieved.
These results suggest that threshold-crossing data (i.e., multiunit activity per electrode) are a rich source of data capable of supporting high prosthetic performance, and are less sensitive to action potential amplitude fluctuations and the multi-month/year decline in action potential amplitudes than appreciated historically. While this is encouraging news in the short term for animal and human translational studies alike , , next-generation systems will require stable recordings for multiple decades, which is a much larger challenge. This will likely require a comprehensive understanding of the (presumably) multiple mechanisms of electrode-tissue recording failure, including but not limited to electrode encapsulation, engineering and materials failures, and micromotion. Developing an accurate compartmental model of these effects could inform design choices for new arrays, as well as the instrumentation used to record from arrays. Studying these effects over clinically relevant timescales of multiple years may require the capability to record waveform and impedance data using multichannel low-power wireless devices . In this way, array lifetime data can be gathered continuously between conventional experiments. Fully implantable wireless devices ,  may also be crucial for removing infection risks and transcutaneous connector tethering forces that may be a major cause of array failure after multiple years. Finally, there may be opportunities to standardize good surgical techniques that have been developed somewhat independently by various groups, but have not been systematically compared.
The performance and capability of prostheses will ultimately be limited if vision alone provides feedback . Natural motor control relies on the integration and use of multiple senses, including proprioception and somatosensation, to coordinate multijoint movements, respond rapidly to unexpected perturbations, and operate in the absence of vision. An action as simple as picking up a styrofoam cup of water relies heavily on somatosensation since vision is often obscured, and even if the hand is clearly visible there is only a minute visual difference between gripping the cup too tightly and crushing it and not gripping it tightly enough and dropping it (e.g., –). It is possible to place joint angle sensors in prosthetic arms, and pressure sensors on prosthetic hands, affording the opportunity to then “write in” this artificial sensory information directly to the brain. The key challenge is how to format and deliver the artificial sensor signals into the appropriate region(s) of the brain in order to produce low-latency, naturalistic sensation that is readily usable as part of (increased) prosthetic performance and capability (e.g., ).
Some progress has been made with electrical stimulation already [see Fig. 1(c)]. Electrical intracortical microstimulation (eICMS) in the proprioceptive area of the primary somatosensory cortex (area 3a) has been used as a potential means to deliver an artificial sense of proprioception to a monkey . This study demonstrated that it is possible for monkeys to detect brief stimulus trains and to discriminate between trains of different frequencies. Similarly in the tactile domain, Romo and colleagues have demonstrated that vibrotactile stimuli delivered mechanically to the finger tip of monkeys can be substituted by eICMS (e.g., , ). The monkeys could readily equate sensations evoked by vibrotactile stimulation to those evoked by microstimulation in primary somatosensory cortex (S1). This is a strong argument in favor of perceptions evoked by eICMS being similar to those evoked by skin vibrotactile stimulation. However, in a similar study monkeys needed two weeks to transfer from vibrotactile to eICMS indicating a potentially different perceptual sensation from the two different stimulus types . Irrespective of this potential caveat, this study also showed that such electrical write in to somatosensory cortex can be combined with a motor cortex read out; it was possible to record electrically during eICMS with only a small (1 ms) blackout period due to eICMS artifacts and, in this way, conduct conjoint neural prosthetic “read out” and eICMS “write in” experiments. In a related study, owl monkeys learned to discriminate spatiotemporal microstimulation patterns . Information was conveyed to the brain through the interplay of microstimulation patterns delivered to multiple electrodes and the temporal order in which these electrodes were stimulated.
While eICMS of the somatosensory cortex has delivered promising results, it remains an open question whether eICMS was perceived as the natural feeling of touch. Binary decisions, such as left versus right reported movements, do not afford this insight. A task with continuously varying stimulation parameters (e.g., frequency, intensity, position, and timing) is needed to mimic naturalistic tactile stimulations. However, neural recordings during such continuous and varied eICMS would likely present a problem. Even with advanced eICMS electrical artifact filtering, the blackout periods would be problematic since the continuous stimulation would cause continuous artifacts. This also hampers bidirectional neural prostheses since they require continuous reading out of motor commands and continuous writing in of proprioceptive and tactile information. A second caveat of using eICMS to imprint meaningful activity patterns into intrinsically highly interconnected areas like S1 is the activation of fibers of passage leading to a spatiotemporal “blur” of activity . There is substantial evidence from experimental work as well as from computer simulations that electrically evoked activity in gray matter is attributed mainly to the excitability of axonal elements and only to a small degree to somata and dendrites. More specifically, 2-photon calcium imaging was used to show recently that microstimulation-evoked activity is surprisingly sparse, and that the pattern of activation is greatly altered by small shifts in electrode position, suggesting that activation occurs primarily via neuronal processes in a small volume surrounding the electrode tip .
What is needed is a method for continuously stimulating and recording, as well as high spatial (individual neuron, specific neuron types, specific projection pattern) and temporal (millisecond timescale) write-in resolution. While electrical stimulation is challenged by these requirements, optogenetic stimulation methods have been designed to address exactly these issues. Optogenetics is based on light-sensitive proteins derived from bacteria and algae. Those proteins are integrated into the membranes of neurons using viral vectors and/or transgenic approaches thus rendering the cells light-responsive. Short light pulses of specific wavelengths in the millisecond range can be used to activate (blue light on Channelrhodopsin 2, ChR2) or inhibit (yellow light on Natronomonas Halorhodopsin, NpHR) neurons expressing the light-sensitive proteins , . By using cell-type specific promoters, these opsins can also be targeted to subpopulations of neurons –. Further, the light stimulation does not cause any electrical artifact thus allowing continuous electrical recordings during stimulation. Finally, the activation or inhibition is extremely fast (millisecond range). This temporal precision is paired with a high spatial resolution. Since the opsins are only expressed by directly infected neurons, no fibers of passage are activated by optical stimulation, preventing spatial “blur.” By combining opsin targeting, light beam focusing, and opsins with different activation spectra neurons can be activated in a spatially highly precise manner. Further, optical stimulation can be finely titrated by varying the light power. An additional means to restrict the activation to a small cortical field is a combination of excitatory and inhibitory opsins. In such a scenario, the patterned illumination with light of different wavelengths (e.g., a blue light beam surrounded by a yellow circle in an area injected with ChR2 and NpHR) would lead to submillimeter precision of activation.
Optogenetic tools were first used in cell cultures and mice, which are amenable to genetic manipulation. To apply them to neural prosthetics research, these tools must now be adapted to function in organisms which are better suited to study complex motor tasks such as rats and rhesus monkeys . In the first study of optogenetics in primates, Lenti viruses were successfully used to target the excitatory channel ChR2 to excitatory neurons . Adeno-associated viruses (AAV) can also be used to deliver ChR2 and other opsins, and these have traditionally been used for gene therapy in humans due to their lower biosafety risks (see  for additional discussion). We have recently used AAV vectors to introduce excitatory and inhibitory opsins into rhesus monkey cortex . Subsequent illumination of the ChR2 and NpHR injected regions with blue or yellow light (the preferred activation wavelength of the inserted proteins) caused strong excitation [see Fig. 3(a)] or strong inhibition [see Fig. 3(b)] of single neuron activity. The expression of the opsins was targeted to neurons excluding astrocytes by using neuron specific promoters. New targeting tools capitalizing on specific viral tropisms and traveling properties now also allow expressing opsins specifically in projecting neurons. This approach does not require any specific promoter fragment or genetic definition of target cells which is a clear advantage for use in less genetically tractable species such as rats and monkeys .
New optoelectronic engineering challenges and opportunities also exist. New“coaxial” optrodes with a small single tip including fiber and electrode have been developed to reduce cortical damage . These next-generation optrodes, and other variants, e.g., micro LED arrays , can be introduced in combination with multielectrode recording arrays in order to combine large-scale electrical recordings with artifact-free optical stimulation (e.g., , ). Such an optoelectronic array would allow spatiotemporal write in of sensory information directly to selected cortical sensory neurons while simultaneously reading out the neural population’s response (e.g., to “verify” write in, or study neural-circuit dynamics). An additional multielectrode array in an output area (e.g., PMd/M1) would allow the simultaneous recording and decoding of the desired prosthetic arm movement. Combining these high-throughput write-in and read-out arrays with dexterous behavioral tasks in nonhuman primates has the potential to accelerate neural prosthesis design and development, ultimately leading to neurally controlled prosthetic arms that enable the users to feel what they are grasping.
The design of neural prostheses has progressed rapidly over the past decade, yielding communication and motor prosthetic systems based on EEG, ECoG, and intracortical electrode array measurements. This perspective and review paper focused on intracortical-based neural prostheses, where the need for further advances in neural “read out” as well as “write in” was highlighted and organized around three challenges. First, new mathematical decode algorithms are needed to more optimally convert the neural activity from intracortical arrays into higher performance systems that are on par with able-bodied performance. Second, new electrode array designs and algorithms are needed to increase system robustness from weeks/months, at present, out to several years/decades. Third, while eICMS can provide at least some semblance of artificial sensation, it is anticipated that the ability to continuously write information into specific neurons, of different types and projection patterns, will be needed to restore natural levels of sensation. Optogenetic methods and oICMS are a potential opportunity for meeting this challenge. Overall, a diverse set of approaches will certainly be needed to fully meet these three key challenge areas. Together, increased performance, increased robustness, and the increased ability to deliver artificial sensation should further increase the benefit (and benefit-risk ratio) intracortical-based neural prostheses can offer patients.
The work of V. Gilja was supported in part by the NSF and NDSEG Graduate Research Fellowships. The work of C. A. Chestek was supported in part by the NSF Graduate Research Fellowship and in part by the William R. Hewlett Stanford Graduate Fellowship. The work of I. Diester was supported in part by the Human Frontier Science Program Fellowship and the German Academic Exchange Service Award. The work of J.M. Henderson was supported in part by the NIH-NINDS under Grant R01-NS066311 supplement, the Stanford Institute for Neuro-Innovation and Translational Neuroscience (SINTN), and the Stanford Bio-X/NeuroVentures Program. The work of K. Deisseroth was supported in part by the William M. Keck Foundation, the Snyder Foundation, the Albert Yu and Mary Bechmann Foundation, the Wallace Coulter Foundation, CIRM, the McKnight Foundation, the Esther A. and Joseph Klingenstein Fund, NSF, NIMH, NIDA, the NIH Director’s Pioneer Award, and DARPA REPAIR under Grant N66001-10-C-2010. The work of K. V. Shenoy was supported in part by the Burroughs Wellcome Fund Career Award in the Biomedical Sciences, Christopher and Dana Reeve Foundation, DARPA REPAIR under Grant N66001-10-C-2010, McKnight Foundation, NIH-NINDS BRP under Grant R01-NS064318, NIH-NINDS under Grant R01-NS066311 supplement, the NIH Director’s Pioneer Award under Grant DP1-OD006409, SINTN, and the Stanford Bio-X/Neuro Ventures Program.
Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org.
Vikash Gilja, Department of Computer Science and SINTN, Stanford University, Stanford, CA 94305 USA (Email: ude.drofnats@ajlig)
Cindy A. Chestek, Department of Electrical Engineering and SINTN, Stanford University, Stanford, CA 94305 USA (Email: ude.drofnats@cydnic)
Ilka Diester, Department of Bioengineering, Stanford University, Stanford, CA 94305 USA (Email: ude.drofnats@retseid).
Jaimie M. Henderson, Departments of Neurosurgery and Neurology and Neurological Sciences, Stanford University, Stanford, CA 94305 USA (Email: ude.drofnats@jredneh).
Karl Deisseroth, Departments of Bioengineering and Psychiatry and Behavioral Sciences and the Neurosciences Program, Stanford University, Stanford, CA 94305 USA, and also with the Howard Hughes Medical Institute, Chevy Chase, MD 20815 USA (Email: ude.drofnats@oressied)
Krishna V. Shenoy, Departments of Electrical Engineering and Bioengineering and the Neurosciences Program, Stanford University, Stanford, CA, 94305 USA (Email: ude.drofnats@yonehs)