| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
SYMPOSIUM REPORT |
1 Wadsworth Center, Laboratory of Nervous System Disorders, New York State Department of Health and State University of New York, Albany, NY 12201, USA
 |
Abstract |
|---|
 Top Abstract References |
|---|
(Received 1 December 2006;
accepted after revision 10 January 2007;
first published online 25 January 2007)
Corresponding author J. R. Wolpaw: Wadsworth Center, New York State Dept. Health, PO Box 509, Empire State Plaza, Albany, NY 12201-0509, USA. Email: wolpaw{at}wadsworth.org
Introduction
Brain–computer interfaces (BCIs) are a fundamentally new approach to restoring communication and control to people with severe motor disorders such as amyotrophic lateral sclerosis (ALS), brainstem stroke, spinal cord injury, muscular dystrophies, and cerebral palsy (Wolpaw et al. 2002; Wolpaw & Birbaumer, 2006 for review). All other assistive technology methods depend on the brain's natural output pathways of peripheral nerves and muscles (or peripheral nerves and glands; Wilhelm et al. 2006), and take outputs that the person still retains (e.g. vertical eye movement in a person with a brainstem stroke) and use these to replace missing functions (e.g. using gaze direction to select letters on a computer screen). In contrast, BCIs give the brain entirely new output pathways. They take electrophysiological or other measures of brain activity and from these measures determine the person's wishes. Intent, which is normally achieved by speaking or by another motor action, is instead achieved by producing brain signals that encode the intent so that a computer can translate it into control of a device such as a computer cursor or a neuroprosthesis.
BCI research, which was confined to only three groups 20 years ago and only six to eight groups as recently as 10 years ago, is now a burgeoning enterprise, with over 100 groups throughout the world engaged in a broad spectrum of research and development efforts, and more entering the field every month. Up to now, this work has demonstrated that a variety of different brain signals, recorded in a variety of different ways and analysed with a variety of different algorithms, can support some degree of real-time communication and control (Vaughan & Wolpaw, 2006). As a result of this collective effort, two facts are becoming increasingly clear. One is encouraging, the other is sobering.
First, BCIs do offer a potentially valuable new option for restoring communication and control to people with severe disabilities; and practical dissemination of BCI technology has in fact begun. Second, however, the development of BCIs that are at once practical, reliable and capable of high-speed complex communication and control is an enormously difficult problem, and one that is far from solution. Furthermore, the origin of the difficulty is not clear – it is not simply a need for better recording methods or improved analysis algorithms – and thus the best route to its solution is also not clear. The origin of this difficulty and how it might be at least circumvented and at best overcome are the topic of this article.
Limitations of current BCIs
BCI studies usually take place in highly controlled laboratory environments or in similarly constrained clinical situations. The BCI user, whether human or animal, typically assumes a specific posture in a simple stereotyped setting free of distractions and operates the BCI for brief periods under close supervision. In spite of these controlled conditions, one of the hallmarks of the results achieved is their variability. Users do much better on some days than others, and performance can vary widely even within a single session and from trial to trial. This high variability is perhaps best illustrated by BCI-based movement control. For example, Fig. 1 compares cursor movement times when the cursor is controlled by a joystick to cursor movement times when the cursor is controlled by a set of single neurons in motor cortex (Hochberg et al. 2006). BCI control is slower than joystick control and is also far more variable. Such variability appears to be a characteristic feature of all BCI approaches, whether non-invasive (e.g. EEG) or invasive (e.g. electrocorticographic (ECoG) or intracortical). In spite of prolonged practice and frequent recalibration of the algorithms that translate brain signals into output commands, variability in performance remains substantial. In contrast to the actions carried out by the brain's normal neuromuscular pathways, which are very consistent from trial to trial (e.g. Fig. 1), the actions carried out through BCIs display a disconcerting, and, up to now at least, ineradicable variability. This variability is likely to be even greater when BCIs are taken out of the protected settings in which they are now typically used and are applied to the day-to-day needs of people with severe disabilities.
|
|
Normal brain functions are widely distributed and undergo continual adaptation
Until less than 200 years ago, the function of the central nervous system (CNS) was not clear: on the one hand, the CNS was thought to provide an interface between a person's immortal soul and the material world, while at the same time it was thought to manage low-level reflex interactions between the organism and its environment (Wolpaw, 2002). In the early 19th century, philosophical developments and experimental discoveries led to the formulation and widespread acceptance of the sensorimotor hypothesis, the hypothesis that the entire function of the CNS is to convert sensory inputs into appropriate motor outputs. This hypothesis sets the agenda of neuroscience: to understand the physiological, anatomical, genetic, developmental, metabolic, hormonal and environmental factors that shape and control this conversion, as well as the pathological processes that can damage or disrupt it.
Because it seeks to establish new output pathways for the brain, new ways of acting on the world, BCI research is a departure from, or an addition to, this agenda. Nevertheless, it depends on the same brain structures and processes that have evolved to control the brain's standard output pathways, and thus it is likely to be governed by the same principles that apply to standard outputs. The research of the past 150 years, and especially of the past several decades, has revealed two basic principles concerning how the brain converts sensory inputs into motor outputs.
First, the task of creating motor outputs is distributed throughout the CNS from the cerebral cortex to the spinal cord. No single area is wholly responsible for an action. As summarized on the left side of Fig. 2, the selection, formulation, and reliable execution of actions such as walking, speaking, or playing the piano are accomplished by collaboration among cortical areas, basal ganglia, thalamic nuclei, cerebellum, brainstem nuclei, and spinal cord interneurons and motoneurons. For example, the high-speed real–time interactions needed to ensure precise and coordinated movements are handled in considerable part by spinal cord reflex pathways. The product of this widely distributed brain activity is appropriate excitation of the spinal cord motoneurons that activate muscles and thereby produce actions. While activity in a variety of brain areas correlates with motor action, the activity in any one area may vary substantially even in highly constrained settings. This variability contrasts with the trial-to-trial consistency of the motor action itself.
|
The challenge of BCI development
Unlike all normal motor actions, which are produced by spinal motoneurons, BCI outputs are produced by brain signals that reflect activity in one or more brain areas. In normal life, the brain activity responsible for these signals simply contributes to motoneuron control. In contrast, when these signals operate a BCI, the brain activity responsible for them becomes the output of the CNS. Figure 2 illustrates this fundamental change. The neurons that produce the brain signals assume the role normally performed by spinal motoneurons; their activity becomes the final product, the output, of the entire CNS. How well they can perform in this new role depends on how well the many brain areas that normally collaborate to control spinal motoneurons can adapt to control instead the neurons that are producing the crucial brain signals. Can they adapt to optimize the brain signals produced by cortical neurons instead of the muscle contractions produced by spinal motoneurons? For example, can the cerebellum, which normally ensures that spinal motoneurons activate muscles so that movement is smooth, rapid, and accurate, change its role to ensure that cortical neurons produce brain signals that move a cursor (or a neuroprosthesis) smoothly, rapidly, and accurately? On the answers to these and related questions depend the ultimate capacities and practical usefulness of BCIs.
The studies to date indicate that the adaptation necessary to control cortical neurons rather than spinal motoneurons is possible but as yet imperfect. As Fig. 1 and the videos referenced above illustrate, trial-to-trial variability is high, and the cursor movements produced are nowhere near as smooth, rapid and accurate as normal limb movements. Furthermore, these imperfections seem to be similarly prominent whether the brain signals used for control are the activity of individual cortical neurons or the amplitudes of EEG rhythms. Thus, the control deficits cannot be readily ascribed to the recording method.
While future refinements in recording and analysis methods and in training algorithms are likely to improve control to some degree, the extent and nature of the control deficits so prominent in the work to date suggest that substantial progress requires a more realistic strategy that recognizes the unique challenge of BCI usage for the CNS. A realistic strategy should minimize the difficulty of the challenge, and should make the challenge as similar as possible to the demands of normal muscle-based control.
Process control versus goal selection
BCIs provide new output pathways for the brain. A BCI output pathway can function in two different ways: it can control a process or it can select a goal. These two options are shown in Fig. 3. A BCI output pathway can, like spinal motoneurons, control all the details of the process that accomplishes the user's intent. For example, it can specify each of the sequence of individual movements that bring the output device, whether a cursor or a neuroprosthesis, to its target. To do this effectively, it must manage intricate high-speed interactions with the device as the movement proceeds. Alternatively, the new output pathway provided by a BCI can simply communicate the goal (e.g. the target to which the cursor should move) to software that then manages the high-speed interactive process that moves the cursor to the target. Up to the present, many noninvasive and almost all invasive BCI studies have adopted the process-control strategy (e.g. Table 1), while non-invasive BCI studies using the P300 evoked potential (e.g. Farwell & Donchin, 1988) and a few invasive studies using cortical neuron activity (e.g. Musallam et al. 2004) have adopted the goal-selection strategy.
|
Furthermore, the goal-selection strategy has the potential to provide more natural control, that is control closer to normal neuromuscular control. As discussed above and illustrated in Fig. 2, normal neuromuscular control is a product of the combined activity of multiple areas from the cortex to the spinal cord. The cortex alone does not control the motoneurons, and, for some important aspects of movement, has relatively little influence. Much of the process of movement control is delegated to subcortical and spinal areas. In a much cruder but qualitatively similar approach, BCI goal selection obtains intent alone from the cortex and delegates actual control of the process to downstream hardware and software. Thus, with further development, including refinement and elaboration of feedback from the downstream apparatus to the CNS, the goal-selection strategy has the potential to emulate with increasing fidelity the brain's normal output pathways. In contrast, process-control BCI methods that vest control entirely in the cortex are likely to remain an artificial and fundamentally unnatural approach.
Conclusion
A BCI changes the final product of CNS activity from spinal motoneuron control to control of the brain area responsible for the signals that the BCI uses to determine the user's intent. Thus, BCI usage presents a unique challenge. It requires that the many CNS areas normally involved in producing motor actions adapt so as to optimize cortical neuron control rather than spinal motoneuron control. The variability characteristic of current BCI performance may stem largely from the difficulty presented by the need for this new and unnatural adaptation. The difficulty might be decreased by switching from a process-control strategy in which the BCI handles all the complex high-speed interactions involved in movement, to a less demanding goal-selection strategy in which the BCI simply communicates the user's goal to software that itself handles the high-speed interactions that achieve the goal. In addition, by assigning lower-level aspects of motor control to an artificial structure (rather than requiring that cortex do everything), the goal-selection strategy imitates the distributed operation typical of normal motor control and may thus provide BCI function that users find more similar to normal motor function.
|
Footnotes |
|---|
|
References |
|---|
|
TopAbstractReferences |
|---|
Carey MR & Lisberger SG (2002). Embarrassed, but not depressed: Eye opening lessons for cerebellar learning. Neuron 35, 223–226.[CrossRef][Medline]
Carmena JM, Lebedev M, Crist RE, O'Doherty JE, Santucci DM, Dimitrov DF, Patil PG, Henriquez CS & Nicolelis MAL (2003). Learning to control a brain–machine interface for reaching and grasping by primates. Plos Biol 1, 1–16.
Cohen TE, Kaplan SW, Kandel ER & Hawkins RD (1997). A simplified preparation for relating cellular events to behavior: mechanisms contributing to habituation, dishabituation, and sensitization of the Aplysia gill-withdrawal reflex. J Neurosci 17, 2886–2899.
Farwell LA & Donchin E (1988). Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalogr Clin Neurophysiol 70, 510–523.[CrossRef][Medline]
Garcia KS, Steele PM & Mauk MD (1999). Cerebellar cortex lesions prevent acquisition of conditioned eyelid responses. J Neurosci 19, 10940–10947.
Hansel C, Linden DJ & D'Angelo E (2001). Beyond parallel fiber LTD: the diversity of synaptic and non-synaptic plasticity in the cerebellum. Nat Neurosci 4, 467–475.[Medline]
Hochberg LR, Serruya MD, Friehs GM, Mukand JA, Saleh M, Caplan AH, Branner A, Chen D, Penn RD & Donoghue JP (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 442, 164–171.
King DAT, Krupa DJ, Foy MR & Thompson RF (2001). Mechanisms of neuronal conditioning. Int Rev Neurobiol 45, 313–337.[Medline]
Lieb JR & Frost WN (1997). Realistic simulation of the Aplysia siphon-withdrawal reflex circuit: roles of circuit elements in producing motor output. J Neurophysiol 77, 1249–1268.
Lisberger SG (1998). Physiologic basis for motor learning in the vestibulo-ocular reflex. Otolaryng Head Neck Surg 119, 43–48.[CrossRef][Medline]
Medina JF, Nores WL, Ohyama T & Mauk MD (2000). Mechanisms of cerebellar learning suggested by eyelid conditioning. Curr Opin Neurobiol 10, 717–724.[CrossRef][Medline]
Medina JF, Repa JC, Mauk MD & LeDoux JE (2002). Parallels between cerebellum- and amygdala-dependent conditioning. Nat Rev Neurosci 3, 122–131.[CrossRef][Medline]
Mendell LM (1984). Modifiability of spinal synapses. Physiol Rev 64, 260–324.
Musallam S, Corneil BD, Greger B, Scherberger H & Andersen RA (2004). Cognitive control signals for neural prosthetics. Science 305, 258–262.
Serruya MD, Hatsopoulos NG, Paminski L, Fellows MR & Donoghue JP (2002). Instant neural control of a movement signal. Nature 416, 141–142.
Taylor DA, Helms Tillery SI & Schwartz AB (2002). Direct cortical control of 3D neuroprosthetic devices. Science 296, 1829–1832.
Thompson RF, Bao S, Chen L, Cipriano BD, Grethe JS, Kim JJ, Thompson JK, Tracy JA, Weninger MS & Krupa DJ (1997). Associative learning. Int Rev Neurobiol 41, 151–189.[Medline]
van Alphen AM & De Zeeuw CI (2002). Cerebellar LTD facilitates but is not essential for long-term adaptation of the vestibulo-ocular reflex. Eur J Neurosci 16, 486–490.[CrossRef][Medline]
Vaughan TM & Wolpaw JR (ed.) (2006). The Third International Meeting on brain–computer interface technology: Making a difference. IEEE Trans Neural Syst Rehabil Eng, vol. 14.
Whelan P & Pearson KG (1997). Plasticity in reflex pathways controlling stepping in the cat. J Neurophysiol 78, 1643–1650.
Wilhelm B, Jordan M & Birbaumer N (2006). Communication in locked-in syndrome: effects of imagery on salivary pH. Neurology 67, 534–535.
Wolpaw JR (2002). Memory in neuroscience: rhetoric versus reality. Behav Cog Neurosci Rev 1, 130–163.
Wolpaw JR & Birbaumer N (2006). Brain–computer interfaces for communication and control. In Textbook of Neural Repair and Rehabilitation; Neural Repair and Plasticity, ed. Selzer ME, Clarke S, Cohen LG, Duncan P & Gage FH, pp. 602–614. Cambridge University Press, Cambridge.
Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G & Vaughan TM (2002). Brain–computer interfaces for communication and control. Clin Neurophysiol 113, 767–791.[CrossRef][Medline]
Wolpaw JR & Lee CL (1989). Memory traces in primate spinal cord produced by operant conditioning of H-reflex. J Neurophysiol 61, 563–572.
Wolpaw JR & McFarland DJ (2004). Control of a two-dimensional movement signal by a noninvasive brain–computer interface in humans. Proc Natl Acad Sci U S A 101, 17849–17854.
Wolpaw JR & Tennissen AM (2001). Activity-dependent spinal cord plasticity in health and disease. Annu Rev Neurosci 24, 807–843.[CrossRef][Medline]
|
Acknowledgements |
|---|
This article has been cited by other articles:
|
|
 |
|
  E. Buch, C. Weber, L. G. Cohen, C. Braun, M. A. Dimyan, T. Ard, J. Mellinger, A. Caria, S. Soekadar, A. Fourkas, et al. Think to Move: a Neuromagnetic Brain-Computer Interface (BCI) System for Chronic Stroke Stroke, March 1, 2008; 39(3): 910 - 917. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 L. G. Cohen and N. Birbaumer The physiology of brain-computer interfaces J. Physiol., March 15, 2007; 579(3): 570 - 570. [Full Text] [PDF]
|
|
|
|
|
|
J. P. Donoghue, A. Nurmikko, M. Black, and L. R. Hochberg Assistive technology and robotic control using motor cortex ensemble-based neural interface systems in humans with tetraplegia J. Physiol., March 15, 2007; 579(3): 603 - 611. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
