Portal de Periódicos da CAPES

Bridging the Brain to the World: A Perspective on Neural Interface Systems

2008; Cell Press; Volume: 60; Issue: 3 Linguagem: Inglês

10.1016/j.neuron.2008.10.037

1097-4199

John P. Donoghue,

Neural dynamics and brain function

Neural interface (NI) systems hold the potential to return lost functions to persons with paralysis. Impressive progress has been made, including evaluation of neural control signals, sensor testing in humans, signal decoding advances, and proof-of-concept validation. Most importantly, the field has demonstrated that persons with paralysis can use prototype systems for spelling, "point and click," and robot control. Human and animal NI research is advancing knowledge about neural information processing and plasticity in healthy, diseased, and injured nervous systems. This emerging field promises a range of neurotechnologies able to return communication, independence, and control to people with movement limitations. Neural interface (NI) systems hold the potential to return lost functions to persons with paralysis. Impressive progress has been made, including evaluation of neural control signals, sensor testing in humans, signal decoding advances, and proof-of-concept validation. Most importantly, the field has demonstrated that persons with paralysis can use prototype systems for spelling, "point and click," and robot control. Human and animal NI research is advancing knowledge about neural information processing and plasticity in healthy, diseased, and injured nervous systems. This emerging field promises a range of neurotechnologies able to return communication, independence, and control to people with movement limitations. Rapid growth and development at the intersection of neuroscience, computer science, engineering, and medicine has allowed the creation of revolutionary neurotechnologies to evaluate and treat nervous system disorders and to restore lost neural functions. Available neurotechnologies can relieve symptoms of Parkinson's disease through electrical stimulation of deep brain structures and restore hearing by stimulating auditory nerve fibers. Neural interface (NI) systems that sense neural signals, also called brain computer interfaces (BCIs), are early-stage neurotechnologies designed to restore control, communication, and independence to persons with paralysis when the motor control structures are disconnected from muscle output. When motor pathways fail NIs offer a physical bridge for movement intention to reach the external world by detecting neural signals that reflect desired actions and transforming them into commands for action, bypassing muscles and damaged neural structures. The emerging neurotechnology field has moved quickly in recent years to demonstrate that people with paralysis can use an NI to perform potentially useful functions. Practical NI systems are not yet widely or commercially available, but many of the critical barriers to success are being tackled. Although the roots of NIs can be traced back well into the last century, scientific as well as public interest in the potential for NI technology was ignited by demonstrations of monkeys substituting neural signals from their motor cortex for hand motions (Serruya et al., 2002Serruya M.D. Hatsopoulos N.G. Paninski L. Fellows M.R. Donoghue J.P. Instant neural control of a movement signal.Nature. 2002; 416: 141-142Crossref PubMed Scopus (1053) Google Scholar, Taylor et al., 2002Taylor D.M. Tillery S.I. Schwartz A.B. Direct cortical control of 3D neuroprosthetic devices.Science. 2002; 296: 1829-1832Crossref PubMed Scopus (1336) Google Scholar). Using neural signals in place of arm motion, able-bodied monkeys moved computer cursors to accomplish goal-directed actions. This proof of concept was soon followed by the launch of a pilot clinical trial in which humans with longstanding tetraplegia demonstrated the ability to use motor cortex activity immediately to operate computer software and control a robotic arm (Hochberg et al., 2006Hochberg L.R. Serruya M.D. Friehs G.M. Mukand J.A. Saleh M. Caplan A.H. Branner A. Chen D. Penn R.D. Donoghue J.P. Neuronal ensemble control of prosthetic devices by a human with tetraplegia.Nature. 2006; 442: 164-171Crossref PubMed Scopus (2301) Google Scholar). Each of these studies, and many complementary studies, was based on a novel approach in which arm movement intentions were captured from the spiking patterns of a population of cortical neurons in motor cortex. During this same period, both the level of interest and accomplishments in a then nearly parallel effort using field potential (FP)-based NI technologies also accelerated. Humans with severe paralysis demonstrated the ability to use FP-NI systems, based on scalp-based electroencephalogram (EEG) sensors, for applications ranging from letter-by-letter spelling (Kubler et al., 2005Kubler A. Nijboer F. Mellinger J. Vaughan T.M. Pawelzik H. Schalk G. McFarland D.J. Birbaumer N. Wolpaw J.R. Patients with ALS can use sensorimotor rhythms to operate a brain-computer interface.Neurology. 2005; 64: 1775-1777Crossref PubMed Scopus (395) Google Scholar, Wolpaw et al., 2002Wolpaw J.R. Birbaumer N. McFarland D.J. Pfurtscheller G. Vaughan T.M. Brain-computer interfaces for communication and control.Clin. Neurophysiol. 2002; 113: 767-791Abstract Full Text Full Text PDF PubMed Scopus (5469) Google Scholar) to 2D cursor control (McFarland et al., 2008McFarland D.J. Krusienski D.J. Sarnacki W.A. Wolpaw J.R. Emulation of computer mouse control with a noninvasive brain-computer interface.J. Neural Eng. 2008; 5: 101-110Crossref PubMed Scopus (141) Google Scholar). By reaching these major milestones, NI systems have come to a threshold of being able to substantially alter the functional capabilities of persons who have any of a wide range of movement limitations. However, NIs in any form must be sufficiently reliable, beneficial, and easy to use for them to become widely adopted and commercially attractive. Both engineering and fundamental scientific issues remain, but considerable progress made so far has helped codify the major obstacles remaining to create a practical human NI system. These initial advances have focused debate and motivated considerable research necessary to realize this entirely new way to help those with movement limitations. The field has also stimulated inquiry into the nature of neural signals and neural coding, investigation of neural implant safety, innovative engineering of "smart" microscale implantable systems, and cross-field discussion of issues and needs of those with movement limitations. The fervor of activity has attracted a number of basic laboratories to engage in multidisciplinary research; shaped scientific meetings and journals (see, e.g., IEEE Trans. Neural Syst. Rehabil. Eng. volume 14:2); evoked vigorous dialog, especially over the use of invasive and noninvasive technologies; and garnered much public and media attention. The move from preclinical to pilot clinical trials has provided a solid example of translational success in neuroscience. Finally, the emergence of NI systems has promoted useful dialog regarding the ethics of communicating directly with the brain, working with a potentially vulnerable user population, and managing conflict of interest when attempting to move scientific and engineering discovery into commercial distribution. All of these can be seen as healthy signs of an emerging area that presents formidable challenges. Many review articles on the various designs and types of NI systems are now available (e.g., J. Physiology volume 579:3). Instead of re-reviewing these reports, the goal of this perspective is to provide a current point of view from one immersed in the field—first, to attempt to identify and clarify some major current key issues, and second, to provide personal impressions of the future of NI systems. Topics are presented as contemporary questions frequently raised in the current NI community. Nearly all in the field will agree that one major goal of NI research is to create a bridge from the brain to the outside world—a kind of replacement part, or prosthesis, for the motor system. A system that senses brain signals may have other roles in evaluating disease states, for example to predict seizure onset in epilepsy, or to guide therapy; these important potential uses of an NI are outside the scope of this perspective. Opinions vary on the target population for NI devices. Concepts for NIs range from an externally driven, reliable, brain-activated switch for a person who is totally unable to move or speak, to an implanted system that provides direct brain-actuated dexterous limb movement for someone with limb paralysis. The design and implementation of these visions share common features but also present independent problems that lead down divergent paths. However simple or elaborate, a functional NI system is potentially of enormous value for individuals with movement limitations. Many disorders leave the cerebral mechanisms for volitional movement intact, but disconnect motor signals from the muscles, preventing normal movement and, in the worst cases of complete "locked in" paralysis, blocking all forms of communication as well. Paralysis originates in diverse ways that include: injury to descending motor pathways in the spinal cord, brainstem, or cerebrum through stroke or trauma (spinal cord injury [SCI], cerebral palsy); degenerative disorders that lead to the loss of motor neurons (such as amyotrophic lateral sclerosis [ALS]) or motor pathways (e.g., multiple sclerosis); degenerative disorders of the muscle (muscular dystrophy); or limb loss. This range of conditions limiting useful movement affects hundreds of thousands in the US alone. An NI offers a physical means to reconnect action intentions to the world, as illustrated in Figure 1. A note on nomenclature is valuable because multiple sets of terminology to name NIs exist. BCI can mean either brain computer interface or brain-controlled interface. The former reflects the idea that neural output, normally meant to control muscles, is now directed at controlling a computer; the latter reflects the fact that the interface is being run directly from the brain without the usual somatic intermediaries. Additionally, neural signals may not go to a computer, but to a machine like a robot; hence the term brain machine interface (BMI) is used. By contrast, NIs may control a range of assistive technologies that are not computers or "machines," such as a push-buttons, so terms that reflect function as a replacement part, such as neural prosthesis or neuromotor prosthesis (NMP), have also been used. Here, I have adopted the term NI system as a general name to encompass the range of these neurotechnologies. There is widespread agreement that any NI requires three major components (Figure 1): (1) a sensor to detect neural signals, (2) a signal processor that converts neural activity into a command related to a desired action, and (3) a device to effect action, often called an assistive technology (AT) in the clinical realm. Ongoing concerns relate mainly to the first two areas, including the optimal types of signals and sensors and the ability to obtain them, decoding approaches, and the necessary capabilities of the control signal. At this point, there has been less attention paid by the NI research community toward explicit AT needs, but this likely reflects the relatively early stage of the field and the substantial challenges of going from neural signals to a command signal. A stable and reliable control signal can be readily applied to many useful ATs. Thus, despite still having various designations, the overall concept of providing a link from neural signals as a means to compensate for loss of control is seen as the central principle that unifies this field. There is not general agreement on how to categorize various emerging types of NI systems, and this reflects the diversity of technologies being developed. Systems may be grouped by the nature of the control signal, sensor location, or output form. One constructive classification method stems from the cerebral processes that the particular NI system engages to provide control: (1) indirect NIs—those systems that co-opt neural events not intrinsically or originally related to intended movement in order to achieve action and (2) direct NIs—those that attempt to control action by using those neural events that underlie the intended movements. Thus, an indirect NI provides a surrogate (replacement) output, because the source of control comes from a signal that substitutes for the missing motor command. Learning to raise or lower the amplitude of an EEG signal over the auditory cortex to activate a switch (Wilson et al., 2006Wilson J.A. Felton E.A. Garell P.C. Schalk G. Williams J.C. ECoG factors underlying multimodal control of a brain-computer interface.IEEE Trans. Neural Syst. Rehabil. Eng. 2006; 14: 246-250Crossref PubMed Scopus (120) Google Scholar) would be one clear example of an indirect NI system, because this signal is not ordinarily directly coupled to movement. One subtype of indirect NI involves learning to associate the power or amplitude of a brain rhythm with a desired action. For example, learned suppression of a cortical rhythm reflecting attention could substitute for action of the hand on a computer mouse. Through a decoder and simple hardware one could couple the amount of attentional suppression to upward movement of computer cursor on a monitor so that cursor motion is achieved without the user engaging hand movement circuitry. Instrumental conditioning or biofeedback-like training is used to form an arbitrary association between whatever modulates the brain signal and the desired action for this learned form of indirect control. A second subtype of indirect system is based on capturing event-related potentials (ERPs) that respond to a time-locked event, which signals the user's intent. The clearest and most successful example of this type of indirect NI is the P300 evoked potential system, in which control is derived from amplitude differences in this response to attended and nonattended computer-flashed stimuli (Birbaumer, 2006Birbaumer N. Breaking the silence: brain-computer interfaces (BCI) for communication and motor control.Psychophysiology. 2006; 43: 517-532Crossref PubMed Scopus (452) Google Scholar). The P300 response can be used, without learning, to select one attended character within a larger matrix of characters to create a spelling device (Figure 4). Research on indirect systems is driving inquiry into the various types of brain rhythms and the ability of humans to learn to control them, as well as the nature of ERPs related to cognitive and other events. By contrast, a direct NI system attempts to reconnect the neural spiking patterns related to movement, say for the arm, directly back to a device that carries out arm-like functions (Donoghue et al., 2007Donoghue J.P. Nurmikko A. Black M. Hochberg L.R. Assistive technology and robotic control using motor cortex ensemble-based neural interface systems in humans with tetraplegia.J. Physiol. 2007; 579: 603-611Crossref PubMed Scopus (141) Google Scholar). Thus, arm movement control signals used to guide hand movement for mouse control of a computer cursor are instead used to guide the cursor directly from the brain. Consequently, a direct NI control signal does not require any initial learning because it maps neural activity related to the intended motor feature directly to the desired action. There has been a major emphasis on arm function for direct NI systems because neural control of the arm in nonhuman primates at the single and neuron population is comparatively well-understood and because restoration of arm-like functions, such as point and click actions of a computer mouse, is both enabling and highly desired by those with tetraplegia (Anderson, 2004Anderson K.D. Targeting recovery: priorities of the spinal cord-injured population.J. Neurotrauma. 2004; 21: 1371-1383Crossref PubMed Scopus (1159) Google Scholar). Direct systems, which necessarily intercept movement commands from only one part of a distributed motor control system, rely on a very limited sample of ongoing processes captured at an intermediate stage. Learning, either by the human or decoders, is therefore likely to play a critical part in optimizing direct NI function and in compensating for missing or disconnected parts of the motor system. An implicit assumption for a direct NI is that control would be more natural and intuitive because it begins with the signal ordinarily used to perform a particular missing action (i.e., hand motor commands to achieve hand-like actions). Success in testing this idea will be discussed below. One contentious but key issue is selecting the optimal neural signal to provide control. In its ideal form, the neural control signal would achieve the quality of the communication link between the brain and the able body. Greater information content, speed of transfer (information rate), reliability, and signal accessibility are features that influence optimal neural signal selection. Lack of fundamental knowledge concerning the nature of neural signals and information coding in the brain, as well as insufficient human NI experience, limits our ability to judge how much control potential there is in various neural signals, although small samples are unlikely to provide elaborate control without support from physical systems. Divergent opinions on the best signal sources have emerged based on the main classes of neural signals. Two broad types of electrical potentials form important information carrying modes or signs of information processing in the nervous system (Bullock, 1997Bullock T.H. Signals and signs in the nervous system: the dynamic anatomy of electrical activity is probably information-rich.Proc. Natl. Acad. Sci. USA. 1997; 94: 1-6Crossref PubMed Scopus (162) Google Scholar; Figure 2): action potentials (spikes) and FPs. Both classes are currently used as NI control sources, and the field has divided, to some degree, along the lines of those using FP, largely in humans, and those using spike-based systems in animal models and, more recently, humans. It is widely held that action potentials, or spikes, are the major neural information-carrying mode of the nervous system, and would thus seem to be the richest source of movement information. Most agree that information is largely carried by spike rate (number of spikes in a specific interval or a related function). The vast majority of systems neurophysiologists investigate information codes at the level of spikes from single cells and, to a lesser extent, evaluate additional information conveyed by populations of spiking neurons. It is not surprising then that researchers from this background form the base of those working on direct NI systems, which are considered direct because they use spikes. The amount of movement information available from spiking activity is impressive. Hand velocity, position, forces, and goals, among other variables, can all be gleaned from single cells in motor cortex (Scott, 2008Scott S.H. Inconvenient truths about neural processing in primary motor cortex.J. Physiol. 2008; 586: 1217-1224Crossref PubMed Scopus (106) Google Scholar). Higher information levels, such as upcoming plans for hand motion, can be decoded from parietal and frontal neuron spiking (Achtman et al., 2007Achtman N. Afshar A. Santhanam G. Yu B.M. Ryu S.I. Shenoy K.V. Free-paced high-performance brain-computer interfaces.J. Neural Eng. 2007; 4: 336-347Crossref PubMed Scopus (55) Google Scholar, Pesaran et al., 2006Pesaran B. Nelson M.J. Andersen R.A. Dorsal premotor neurons encode the relative position of the hand, eye, and goal during reach planning.Neuron. 2006; 51: 125-134Abstract Full Text Full Text PDF PubMed Scopus (259) Google Scholar, Scherberger and Andersen, 2007Scherberger H. Andersen R.A. Target selection signals for arm reaching in the posterior parietal cortex.J. Neurosci. 2007; 27: 2001-2012Crossref PubMed Scopus (104) Google Scholar). Recording many cells at once adds information distributed across heterogeneous populations and reduces noise (information variability) by averaging across neurons (e.g., Maynard et al., 1999Maynard E.M. Hatsopoulos N.G. Ojakangas C.L. Acuna B.D. Sanes J.N. Normann R.A. Donoghue J.P. Neuronal interactions improve cortical population coding of movement direction.J. Neurosci. 1999; 19: 8083-8093PubMed Google Scholar). The number of neurons required for a reasonably reliable reconstruction of hand motion is remarkably small. For example, about 50 cells in the MI arm area can provide a very good estimate of hand motion in 2D or 3D space (Serruya et al., 2002Serruya M.D. Hatsopoulos N.G. Paninski L. Fellows M.R. Donoghue J.P. Instant neural control of a movement signal.Nature. 2002; 416: 141-142Crossref PubMed Scopus (1053) Google Scholar, Taylor et al., 2002Taylor D.M. Tillery S.I. Schwartz A.B. Direct cortical control of 3D neuroprosthetic devices.Science. 2002; 296: 1829-1832Crossref PubMed Scopus (1336) Google Scholar, Carmena et al., 2005Carmena J.M. Lebedev M.A. Henriquez C.S. Nicolelis M.A. Stable ensemble performance with single-neuron variability during reaching movements in primates.J. Neurosci. 2005; 25: 10712-10716Crossref PubMed Scopus (109) Google Scholar). Consequently, there has been great interest in using signals from populations of a few dozen neurons, now technically feasible to gather, as control signals for direct NI systems. The mixed signal, when many spikes are intermingled together from a single site, is called multiunit activity (MUA), which is thought to represent averaged spiking of a local population. MUA is also a spiking signal of interest for NI applications because it reduces the technical demands of isolating single neurons on each electrode, although MUA is just beginning to be studied in this domain (Stark et al., 2008Stark E. Drori R. Abeles M. Motor cortical activity related to movement kinematics exhibits local spatial organization.Cortex. 2008; (in press. Published online July 10, 2008)https://doi.org/10.1016/j.cortex.2008.03.011Abstract Full Text Full Text PDF PubMed Scopus (22) Google Scholar). FPs are the other type of neural electrical potential that is used to obtain control signals. FPs are more complex than spikes, in that they reflect the flow of transmembrane currents, usually of synaptic origin, summed across groups of neurons of varying size, frequency, and spatial distribution. While the recognition of FPs as an information-carrying signal is long standing, the recent surge in NI interest has renewed and accentuated study into the nature, origin, and significance of FPs and their relationship to spiking. FPs are both signals and signs (Bullock, 1997Bullock T.H. Signals and signs in the nervous system: the dynamic anatomy of electrical activity is probably information-rich.Proc. Natl. Acad. Sci. USA. 1997; 94: 1-6Crossref PubMed Scopus (162) Google Scholar) of underlying neural processes that generally reflect brain states, such as stages of sleep or alertness or higher cognitive processes. A comprehensive discussion of the many subtypes and sources of these signals is not possible here; only a brief explanation of the type of control signals possible from FP and recent advances in their use in NI will be presented (for additional perspectives see Clinical Neurophys. volume 117:3; Birbaumer, 2006Birbaumer N. Breaking the silence: brain-computer interfaces (BCI) for communication and motor control.Psychophysiology. 2006; 43: 517-532Crossref PubMed Scopus (452) Google Scholar, Vaughan et al., 2006Vaughan T.M. McFarland D.J. Schalk G. Sarnacki W.A. Krusienski D.J. Sellers E.W. Wolpaw J.R. The Wadsworth BCI Research and Development Program: at home with BCI.IEEE Trans. Neural Syst. Rehabil. Eng. 2006; 14: 229-233Crossref PubMed Scopus (259) Google Scholar). A schema to organize common types and subbands of signals relevant to NI control is shown in Figure 2. In this context, FPs include two major subgroups: (1) rhythmic signals that can be grouped as slow, medium, or fast, and (2) ERPs, which are responses triggered by a time-locked event. All three rhythms have been used as NI control signals. Slower cortical potentials ( 30 Hz to ∼100 Hz) are only recently being carefully evaluated with intracranial recordings because they are heavily filtered in more common scalp recordings. Middle and high bands appear to carry distinctly different information (Belitski et al., 2008Belitski A. Gretton A. Magri C. Murayama Y. Montemurro M.A. Logothetis N.K. Panzeri S. Low-frequency local field potentials and spikes in primary visual cortex convey independent visual information.J. Neurosci. 2008; 28: 5696-5709Crossref PubMed Scopus (286) Google Scholar). Beta oscillations appear in primary motor cortex (MI) in able-bodied monkeys (Baker et al., 2003Baker S.N. Pinches E.M. Lemon R.N. Synchronization in monkey motor cortex during a precision grip task. II. effect of oscillatory activity on corticospinal output.J. Neurophysiol. 2003; 89: 1941-1953Crossref PubMed Scopus (147) Google Scholar, Donoghue et al., 1998Donoghue J.P. Sanes J.N. Hatsopoulos N.G. Gaal G. Neural discharge and local field potential oscillations in primate motor cortex during voluntary movements.J. Neurophysiol. 1998; 79: 159-173PubMed Google Scholar, Murthy and Fetz, 1996Murthy V.N. Fetz E.E. Synchronization of neurons during local field potential oscillations in sensorimotor cortex of awake monkeys.J. Neurophysiol. 1996; 76: 3968-3982PubMed Google Scholar) and in humans with paralysis (Hochberg et al., 2006Hochberg L.R. Serruya M.D. Friehs G.M. Mukand J.A. Saleh M. Caplan A.H. Branner A. Chen D. Penn R.D. Donoghue J.P. Neuronal ensemble control of prosthetic devices by a human with tetraplegia.Nature. 2006; 442: 164-171Crossref PubMed Scopus (2301) Google Scholar), where they, rather than spiking patterns, mark the transition from preparation to intended action. By contrast, gamma rhythms are correlated with aspects of spiking as shown in visual (Belitski et al., 2008Belitski A. Gretton A. Magri C. Murayama Y. Montemurro M.A. Logothetis N.K. Panzeri S. Low-frequency local field potentials and spikes in primary visual cortex convey independent visual information.J. Neurosci. 2008; 28: 5696-5709Crossref PubMed Scopus (286) Google Scholar) and parietal cortex (Andersen et al., 2004Andersen R.A. Musallam S. Pesaran B. Selecting the signals for a brain-machine interface.Curr. Opin. Neurobiol. 2004; 14: 720-726Crossref PubMed Scopus (240) Google Scholar), suggesting that they might provide specific forms of information (Womelsdorf and Fries, 2006Womelsdorf T. Fries P. Neuronal coherence during selective attentional processing and sensory-motor integration.J. Physiol. (Paris). 2006; 100: 182-193Crossref PubMed Scopus (108) Google Scholar) useful for NI movement applications. Beyond rhythms, FPs include ERPs. ERPs signify large-scale potential shifts in neuronal populations that can be elicited and modulated by various external or internal events. As noted earlier, the P300 has been intensively investigated for NI applications. FP can radiate considerable distances, especially in the lower frequencies, and can therefore be recorded electrically outside as well as inside the head, unlike spikes. FP recorded by scalp electrodes is called the EEG; FP recorded inside the skull, close to the cortical surface (above or below the dura) is the electrocorticogram (ECoG); and the FP recorded intraparenchymally is the local field potential (LFP). The ease of recording FP at the scalp has made the EEG attractive as a signal source to create, test, and develop NI systems using humans in a number of laboratories, and has allowed them to be usefully adopted by persons with paralysis (Vaughan et al., 2006Vaughan T.M. McFarland D.J. Schalk G. Sarnacki W.A. Krusienski D.J. Sellers E.W. Wolpaw J.R. The Wadsworth BCI Research and Development Program: at home with BCI.IEEE Trans. Neural Syst. Rehabil. Eng. 2006; 14: 229-233Crossref PubMed Scopus (259) Google Scholar). EEG systems have drawbacks as potential sensors and FP signal sources for NIs that include