* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download Signal acquisition and analysis for cortical control of neuroprosthetics
Convolutional neural network wikipedia , lookup
Holonomic brain theory wikipedia , lookup
Functional magnetic resonance imaging wikipedia , lookup
Artificial general intelligence wikipedia , lookup
Embodied language processing wikipedia , lookup
Neuroethology wikipedia , lookup
Cognitive neuroscience of music wikipedia , lookup
Neurophilosophy wikipedia , lookup
Artificial neural network wikipedia , lookup
Single-unit recording wikipedia , lookup
Clinical neurochemistry wikipedia , lookup
Cortical cooling wikipedia , lookup
Binding problem wikipedia , lookup
Neural coding wikipedia , lookup
Neural modeling fields wikipedia , lookup
Multielectrode array wikipedia , lookup
Central pattern generator wikipedia , lookup
Feature detection (nervous system) wikipedia , lookup
Neuroeconomics wikipedia , lookup
Types of artificial neural networks wikipedia , lookup
Recurrent neural network wikipedia , lookup
Perceptual control theory wikipedia , lookup
Neural oscillation wikipedia , lookup
Synaptic gating wikipedia , lookup
Neuroplasticity wikipedia , lookup
Neuroanatomy wikipedia , lookup
Brain–computer interface wikipedia , lookup
Nervous system network models wikipedia , lookup
Neural correlates of consciousness wikipedia , lookup
Premovement neuronal activity wikipedia , lookup
Optogenetics wikipedia , lookup
Neural engineering wikipedia , lookup
Neuropsychopharmacology wikipedia , lookup
Neuroprosthetics wikipedia , lookup
Neural binding wikipedia , lookup
Development of the nervous system wikipedia , lookup
Signal acquisition and analysis for cortical control of neuroprosthetics Stephen I Helms Tillery1 and Dawn M Taylor2,3 Work in cortically controlled neuroprosthetic systems has concentrated on decoding natural behaviors from neural activity, with the idea that if the behavior could be fully decoded it could be duplicated using an artificial system. Initial estimates from this approach suggested that a high-fidelity signal comprised of many hundreds of neurons would be required to control a neuroprosthetic system successfully. However, recent studies are showing hints that these systems can be controlled effectively using only a few tens of neurons. Attempting to decode the pre-existing relationship between neural activity and natural behavior is not nearly as important as choosing a decoding scheme that can be more readily deployed and trained to generate the desired actions of the artificial system. These artificial systems need not resemble or behave similarly to any natural biological system. Effective matching of discrete and continuous neural command signals to appropriately configured device functions will enable effective control of both natural and abstract artificial systems using compatible thought processes. Addresses 1 The Biodesign Institute & Harrington Department of Bioengineering, ECG 334, Arizona State University, Tempe, Arizona, 85287-9709, USA e-mail: [email protected] 2 Department of Biomedical Engineering, Case Western Reserve University, 10900 Euclid Avenue, Cleveland, Ohio, 44106, USA 3 Louis Stokes Cleveland VA Medical Center, 10701 East Boulevard, Cleveland, OH 44106, USA e-mail: [email protected] Current Opinion in Neurobiology 2004, 14:758–762 This review comes from a themed issue on Neurobiology of behaviour Edited by Alexander Borst and Wolfram Schultz Available online 5th November 2004 0959-4388/$ – see front matter # 2004 Elsevier Ltd. All rights reserved. DOI 10.1016/j.conb.2004.10.013 Abbreviations FES functional electrical stimulation Introduction ‘We will be engaged in the development of principles and techniques by which information from the nervous system can be used to control external devices such as prosthetic devices, communications equipment, teleoperators [. . .] and ultimately perhaps even computers’ [1]. Current Opinion in Neurobiology 2004, 14:758–762 Karl Frank, Founder of the Laboratory of Neural Control at the NINDS, 1968. In the above quote from 1968, Frank described the mission of the newly created Laboratory of Neural Control at the National Institutes of Health. This mission is now coming to fruition, ironically enough, with the first major milestones being in the realm of computer control and, more recently, control of robotic arms. Recent advances in recording and computer technology have made it possible to utilize signals from large numbers of neurons to drive external systems in real time. These advances have captured the public imagination with images of high-tech devices being controlled by one’s thoughts. In this high-profile research environment, it seems appropriate to step back and put this work into perspective. Here, we discuss the progress that has been made in the broad field of brain interfacing for device control, and contrast the philosophies and successes experienced by a range of approaches to the design of control algorithms. Why cortical control As a result of advances in medical technology, more people are now surviving severely paralyzing injuries or diseases [2]. This has spawned an increase in the development of assistive technology to improve the quality of life of these individuals. For people who are almost completely paralyzed or ‘locked-in’, owing to amyotrophic lateral sclerosis or brainstem stroke, specialized typing software can provide a way to communicate with their caregivers and commercially available wheelchairmounted robotic arms can enable them to interact physically with the world. For people with high-level spinal cord injuries, restoration of arm and hand function is now possible through the use of implanted functional electrical stimulation (FES) systems that activate paralyzed muscles through controlled stimulation of the peripheral nerves [3,4]. By decoding movement intent directly from the brain and then creating the desired movement through FES, paralyzed individuals could, once again, move their arms and hands just by thinking of doing so [5,6]. Although restoring movement by thought has much conceptual appeal, the use of cortical signals has to be weighed against the other command options available to a person. For completely paralyzed individuals, brain signals are their only option. For people with high-level spinal cord injuries (C1–C4), command signals can still be generated from the neck upwards. However, many www.sciencedirect.com Signal acquisition and analysis for cortical control of neuroprosthetics Tillery and Taylor 759 options, such as mouth-operated joysticks, tongue-touch keypads and voice commands, interfere with talking, eating and normal social interaction. Making useful arm movements (FES-activated or robotic) requires multiple proportional command signals that are not always easily generated for a given task from the face and neck. Accessing desired movement directly from the brain could replace, or at least augment, these less convenient command options and improve the quality of life of severely paralyzed individuals. Volitional control of cortical neurons Some have questioned whether or not we have volitional control over the activity of our cortical neurons. The entire field of motor cortical physiology is devoted to studying those areas of the cortex which are ‘turned on’ when we want to move. Could there be a more direct example of ‘volitional control’ over neuronal firing rates? However, we are still learning the extent to which one has direct access to the firing patterns of individual neurons. Certainly, when we send out motor signals we do not think about inducing a particular firing pattern; we think of the desired result (e.g. pick up the glass) and, through practice and experience, we produce the firing patterns necessary to make the desired action take place. It was not until the 1960s that explicit studies of an animal’s ability to control the firing rates of individual neurons began in earnest [7–14]. In these operant conditioning studies, animals were provided with feedback of the firing rates of cortical neurons and/or rewarded for changing these signals on command. Animals showed the ability to raise or lower neural firing rates at will, alter the normal relationship between firing rate and muscle activity, alter the firing relationship between two neighboring neurons and put firing rates of individual neurons into a sequence of specific ranges on command. Recent experiments in humans have shown a similar ability to regulate the firing patterns of recorded cortical neurons [15,16] or invasively recorded field potentials [17,18] at will. A long history of electroencephalographic studies have already shown that people can learn to modulate scalp surface potentials as well [19]. Conversion algorithms The aim of a neuroprosthetic system is to ‘read’ a person’s intent from the recorded neuronal signals and use that information to command another device. The form of the command signal (i.e. discrete or continuous) depends on the device one is aiming to control. To generate command signals for continuous systems, such as robotic arms or computer cursors, the dominant approach has been to record an ensemble signal during natural movements (usually arm movements) and then develop a decoding scheme that enables reconstruction of those movements from the neuronal signals [20–24,25,26,27]. The core assumption is that, if one can understand exactly what an www.sciencedirect.com animal is trying to do, then it becomes a matter of engineering to accomplish the same task with another device. The analysis of ensemble neural activity during natural arm movements has increased our fundamental understanding of how the nervous system controls movements but it might not be the best approach for identifying decoding algorithms to convert neural activity into device commands. First, there is no reason to believe that the neural activity associated with thoughts of moving an external device, such as a computer cursor or a robot, will be the same as that generated when thinking about moving one’s own arm. Two studies that compared neural activity during target-directed arm movements with equivalent movements made using a brain-controlled virtual cursor showed substantial differences in the directional tuning of the neurons between the two tasks [24,26]. The different dynamics of an external system might also require the neural activity to change when switching from arm control to device control. The same two laboratories have recently reported training monkeys to use intracortical signals for closed-loop control of a robotic arm in three-dimensional space [25] or in two spatial dimensions plus gripper closure [26]. Both studies report that the animals easily adjusted to the novel dynamics of the robotic systems but, again, the neural activity changed when the animals switched from normal arm movements to the brain control tasks. Another limitation of optimizing decoding algorithms on neural activity recorded during natural movements is that it does not take advantage of the inevitable learning that takes place when a person or an animal has real-time feedback of their brain-controlled movements. The operant conditioning work of the 1960s and 1970s clearly demonstrated that animals could learn to alter their neural output with practice when given real-time feedback of their neural activity [7–14]. More recent studies in closedloop two- or three-dimensional continuous braincontrolled movement tasks have shown that monkeys improve performance in these tasks with training [24,26]. Recent human studies that used brain surface signals to control a cursor demonstrated that performance improvements can occur within minutes [18]. In the previously described monkey study, in which animals controlled a three-dimensional cursor or robotic arm, the strength of directional tuning of the individual recorded units significantly improved with regular practice. Within just 12 training sessions, each with a duration of about one hour, the average strength of directional tuning of the recorded units during the brain control task significantly exceeded the strength of directional tuning exhibited during normal arm movements [24]. In further Current Opinion in Neurobiology 2004, 14:758–762 760 Neurobiology of behaviour offline studies, maximum likelihood analysis of neuronal activity recorded during normal target-directed arm movements showed that the intended target could only be predicted with about 65% accuracy [28]. The small ensemble of recorded neurons did not naturally convey enough information about the intended target for accurate prediction. However, when the same animal made similar target-directed cursor movements using its brain signals directly, the visual feedback of the braincontrolled cursors enabled the animal to learn to modulate its recorded signals more effectively. A similar maximum likelihood estimation of target location based on neural activity recorded during the closed-loop braincontrolled movements improved prediction accuracy to 80%. With continued practice, the animal learned to make long continuous sequences of three-dimensional brain-controlled movements to 50–70 targets at a time, without missing [24]. Neuron selection Determining which neuronal ensembles will form the best mapping between cortical activity and a specific device command is still an unresolved issue. First, different brain regions might contain units that are more appropriate for one neuroprosthetic task than another. For example, recent work suggests that, although the motor cortex might be useful for encoding continuous motion, dorsal premotor cortex [29,30,31] or posterior parietal cortex [32,33] might be able to provide a higher-level estimate of the goal of a movement. However, the pattern classification analysis employed in these recent studies is best suited for selection among a limited set of discrete goals as opposed to determining a reach goal from a continuous range of possible goals throughout the workspace. The discrete pattern classification techniques demonstrated thus far might be most effectively applied in computer-based systems designed to enable severely paralyzed individuals to choose efficiently between a fixed set of different letters, words or icons distributed around a computer screen. Many complex assistive devices have been designed to convert discrete goal requests into detailed plans for action (e.g. NavChair [34,35]). Continuous movement command signals can also be effectively used in discrete choice selection tasks just by moving the cursor to one of a fixed number of targets. To speed up the goal-selection process, trajectories do not need actually to reach the goal target. They only need to proceed far enough toward the desired target for the controller to determine which target is the intended goal. Post hoc analysis of continuous brain-controlled cursor movements in an eight-target three-dimensional centerout task has shown that discrete target selection could have been accurately predicted just from the initial part of the brain-controlled trajectories [25,28]. This is not surprising, considering that studies of motor cortical areas Current Opinion in Neurobiology 2004, 14:758–762 have shown that target direction is strongly encoded early on in center-out reaching movements [36]. Training a person to use their brain signals for continuous cursor control might provide them with more flexibility, giving them both rapid discrete choice selection and continuous cursor control using the same learned brain-control skills. Ensemble size Another issue to address is how many neurons are really needed for good long-term reliable control of various neuroprosthetic systems. The number of neurons obtainable with the current hardware must meet or exceed the number needed for adequate command before the benefits will outweigh the risks of such a device. Several research laboratories have now shown good real-time twoand three-dimensional continuous or discrete control using ensembles in the tens of neurons [22–24,25, 30,31,32,33]. However, there are suggestions that good reliable long-term control will require as many as hundreds of neurons [26,29,37]. A couple of issues stand in the way of fully putting this debate to rest. One issue is the stability of the recorded units over time. Recording many units will provide a more robust command source that can remain stable as neurons drop out or enter the mix over time. It is likely that recording stability will increase as better implantable electrodes are designed [38–41]. Another unresolved issue is the extent to which neurons will learn to produce the desired command signals with continuous daily practice. The monkey studies to date have only utilized the recorded neurons in brain control tasks for a limited time each day. The animals then go back to using those same neurons for normal motor control activities for the rest of the time. It is likely that the recorded neurons will become very proficient at command of a given device over time, once their neural activity is consistently used exclusively for that task. This was seen to some extent in the three-dimensional brain control experiment described above, in which the tuning quality during brain control improved with regular daily practice [24]. This neural plasticity might also be able to compensate for shifts in the recorded populations over time but only if brain interfacing and decoding systems are appropriately designed to facilitate the transition and retraining of changing neural populations. Although being able to record a lot of neurons sounds desirable, more is not always better. In one recent experiment, limiting the neurons used for decoding to only the best units improved movement accuracy. Movement quality deteriorated when additional lower-quality units were included (R Wahnoun, J He, SI Helms Tillery, unpublished). Selective use of the recorded neurons is also important because the number of recorded signals and the complexity of the conversion algorithm will affect www.sciencedirect.com Signal acquisition and analysis for cortical control of neuroprosthetics Tillery and Taylor 761 the amount of hardware needed to utilize the signals for device control. For a mobile wheelchair user, the racks of external equipment needed to apply complex sorting and conversion algorithms to high-density data might not be worthwhile if most of the command information can still be obtained using a smaller subset of neurons and simple decoding algorithms implemented on small portable hardware. To date, cortical activity has only been used to control, at most, three simultaneous degrees of freedom. At first glance, complex systems, such as paralyzed arms and hands or multijoint robotic devices, appear to require complex command signals, suggesting that large numbers of recorded neurons will be needed for adequate control. However, careful system design will minimize the number of command signals and, therefore, the number of neurons required for any given device. For example, a paralyzed arm, along with most multiaxis robotic systems, can be directed using simple three-dimensional endpoint commands. The device controller can transparently take care of the complex details of determining muscle activations or joint torques needed to achieve the desired endpoint path. Hand function, which at first glance appears to be extremely complex with more than 20 degrees of freedom, can be adequately described using just a few aggregate variables [42–44], and restoration of practical hand function has been demonstrated with current FES systems that utilize very simple command signals [3–6]. Conclusions: moving forward The combination of our current knowledge of motor systems with the development of relatively cheap, high-speed computing and multichannel recording systems has produced rapid advances in the field of cortically controlled neuroprosthetics. Both human and animal studies have now demonstrated that it is possible to utilize neural signals in real time to control the motion of a robotic arm or a computer cursor and to make discrete choice selections. This technology holds great promise for improving the quality of life of severely paralyzed individuals. However, the field is in something of an arms race, with a push to record more neurons and implement more complex algorithms to extract details about arm movements from healthy behaving animals. This approach has provided rapid advances in both neuroprosthetics and neuroscience, but learning the strengths and limitations of neural plasticity is essential for rapidly moving this field forward. Future research must be directed toward practical implementation in people with severe movement impairments in order to understand fully and utilize the potential of the brain in the paralyzed population. Update Since this article was written, Cyberkinetics Neurotechnology Systems Inc. released their initial findings from www.sciencedirect.com the first subject enrolled in a pilot study to test their product, the BrainGateTM neural interfacing system (http://phx.corporate-ir.net/phoenix.zhtml?c=182802&p= irol-newsArticle&ID=628347&highlight=). Although human testing of intracortical signals for computer control has been underway for several years (Kennedy et al. [15,16], now with Neural Signals Inc.), the early studies were limited to subjects who were almost completely paralyzed. The new study by Cyberkinetics Neurotechnology Systems Inc marks the first use of chronic multichannel intracortical microelectrodes in a subject with a high-level spinal cord injury. This first subject demonstrated that single and multiunit action potentials can be recorded from the motor cortex three years after a spinal cord injury and that the subject could volitionally modulate his recorded neural signals to move a computer cursor in a controlled manner. Notably, the subject is able to carry on conversations with ease while manipulating the brain-controlled cursor. Although the accuracy of cursor movements leaves room for improvement, the quality of the cursor movements appears to exceed those obtained by this group in previous monkey studies that utilized similar hardware and software [23]. Over time, results from this study will provide valuable information on the effects of long-term training in humans. The influx of venture capitol funding for this project played a substantial part in reaching this milestone, and reflects the promise of continued progress in this field. References and recommended reading Papers of particular interest, published within the annual period of review, have been highlighted as: of special interest of outstanding interest 1. Frank K: Some approaches to the technical problem of chronic excitation of peripheral nerve. Ann Otol Rhinol Laryngol 1968, 77:761-771. 2. DeVivo MJ, Krause JS, Lammertse DP: Recent trends in mortality and causes of death among persons with spinal cord injury. Arch Phys Med Rehabil 1999, 80:1411-1419. 3. Peckham PH, Keith MW, Kilgore KL, Grill JH, Woulle KS, Thrope GB, Gorman P, Hobby J, Mulcahey MJ, Carroll S et al.: Efficacy of an implanted neuroprosthesis for restoring hand grasp in tetraplegia: a multicenter study. Arch Phys Med Rehabil 2001, 82:1380-1388. 4. Peckham PH, Kilgore KL, Keith MW, Bryden AM, Bhadra N, Montague FW: An advanced neuroprosthesis for restoration of hand and upper arm control using an implantable controller. J Hand Surg [Am] 2002, 27:265-276. 5. Lauer RT, Peckham PH, Kilgore KL, Heetderks WJ: Applications of cortical signals to neuroprosthetic control: a critical review. IEEE Trans Rehabil Eng 2000, 8:205-208. 6. Pfurtscheller G, Muller GR, Pfurtscheller J, Gerner HJ, Rupp R: ‘Thought’ — control of functional electrical stimulation to restore hand grasp in a patient with tetraplegia. Neurosci Lett 2003, 351:33-36. 7. Olds J: Operant conditioning of single unit responses. In 23rd International Congress of Physiological Sciences 1965; Tokyo, Japan: Excerpta Medica Foundation: 1965:372-380. 8. Olds J, Olds ME: Interference and learning in palaeocortical systems. In Brain Mechanisms and Learning. Edited by Delafresnaye JF. Oxford, UK: Charles C. Thomas; 1961:153-187. Current Opinion in Neurobiology 2004, 14:758–762 762 Neurobiology of behaviour 9. Fetz EE, Baker MA: Operantly conditioned patterns on precentral unit activity and correlated responses in adjacent cells and contralateral muscles. J Neurophysiol 1973, 36:179-204. 10. Fetz EE, Finocchio DV: Operant conditioning of isolated activity in specific muscles and precentral cells. Brain Res 1972, 40:19-23. 11. Fetz EE, Finocchio DV: Operant conditioning of specific patterns of neural and muscular activity. Science 1971, 174:431-435. 12. Fetz EE: Operant conditioning of cortical unit activity. Science 1969, 163:955-958. 13. Schmidt EM, McIntosh JS, Durelli L, Bak MJ: Fine control of operantly conditioned firing patterns of cortical neurons. Exp Neurol 1978, 61:349-369. 14. Schmidt EM, Bak MJ, McIntosh JS, Thomas JS: Operant conditioning of firing patterns in monkey cortical neurons. Exp Neurol 1977, 54:467-477. 15. Kennedy PR, Bakay RA, Moore MM, Adams K, Goldwaithe J: Direct control of a computer from the human central nervous system. IEEE Trans Rehabil Eng 2000, 8:198-202. 16. Kennedy PR, Bakay RA: Restoration of neural output from a paralyzed patient by a direct brain connection. Neuroreport 1998, 9:1707-1711. 17. Kennedy PR, Andreasen D, Ehirim P, King B, Kirby T, Mao H, Moore MM: Using human extra-cortical local field potentials to control a switch. J Neural Eng 2004, 1:72-77. This study showed that signals recorded extradurally over the sensorimotor cortex using permanently implanted skull screws can be used to turn a switch on and off. 18. Leuthardt EC, Schalk G, Wolpaw JR, Ojemann JG, Moran DW: A brain–computer interface using electrocorticographic signals in humans. J Neural Eng 2004, 1:63-71. This study also used signals recorded with low-impedance electrodes, this time placed subdurally, to control the motion of a cursor in a video game. The results demonstrated rapid learning and useful signal modulation in the gamma band. 19. Wolpaw JR, Birbaumer N, McFarland DJ, Pfurtscheller G, Vaughan TM: Brain–computer interfaces for communication and control. Clin Neurophysiol 2002, 113:767-791. 20. Humphrey DR, Schmidt EM, Thompson WD: Predicting measures of motor performance from multiple cortical spike trains. Science 1970, 170:758-762. 21. Schwartz AB, Taylor DM, Tillery SI: Extraction algorithms for cortical control of arm prosthetics. Curr Opin Neurobiol 2001, 11:701-707. 22. Serruya MD, Hatsopoulos NG, Paninski L, Fellows MR, Donoghue JP: Brain–machine interface: instant neural control of a movement signal. Nature 2002, 416:141-142. 23. Serruya M, Hatsopoulos N, Fellows M, Paninski L, Donoghue J: Robustness of neuroprosthetic decoding algorithms. Biol Cybern 2003, 88:219-228. 24. Taylor DM, Tillery SI, Schwartz AB: Direct cortical control of 3D neuroprosthetic devices. Science 2002, 296:1829-1832. 25. Taylor DM, Helms Tillery SI, Schwartz AB: Information conveyed through brain-control: cursor versus robot. IEEE Trans Neural Syst Rehabil Eng 2003, 11:195-199. This study extended the previous results from this group [24], providing the first demonstration of closed-loop three-dimensional control of a robotic arm using intracortical signals in monkeys. 26. Carmena JM, Lebedev MA, Crist RE, O’Doherty JE, Santucci DM, Dimitrov D, Patil PG, Henriquez CS, Nicolelis MA: Learning to control a brain–machine interface for reaching and grasping by primates. PLoS Biol 2003, 1:E42. 27. Kemere C, Shenoy KV, Meng TH: Model-based neural decoding of reaching movements: a maximum likelihood approach. IEEE Trans Biomed Eng 2004, 51:925-932. This study showed that a neural decoding scheme that begins with a careful understanding of the controlled system (the arm in this case) can Current Opinion in Neurobiology 2004, 14:758–762 produce control with half as many neurons as one that makes no assumptions about movement at all. 28. Helms Tillery SI, Taylor DM, Schwartz AB: Training in cortical control of neuroprosthetic devices improves signal extraction from small neuronal ensembles. Rev Neurosci 2003, 14:107-119. 29. Hatsopoulos N, Joshi J, O’Leary JG: Decoding continuous and discrete motor behaviors using motor and premotor cortical ensembles. J Neurophysiol 2004, 92:1165-1174. These authors present a double-dissociation in which an ensemble of M1 neurons provides better rendering of movement trajectory than a similar ensemble of dorsal premotor cortex neurons, whereas the dorsal premotor cortex neuron ensemble provides a better prediction of the target of an upcoming movement than the M1 ensemble. 30. Ryu SI, Santhanam G, Yu BM, Shenoy KV: High speed prosthetic icon positioning. In Proceedings of the Society for Neuroscience 34th Annual Meeting; San Diego, CA: 2004. 31. Santham G, Ryu SI, Yu BM, Shenoy KV: High information transfer rates in a neural prosthetic. In Proceedings of the Society for Neuroscience 34th Annual Meeting; San Diego, CA: 2004. 32. Shenoy KV, Meeker D, Cao S, Kureshi SA, Pesaran B, Buneo CA, Batista AP, Mitra PP, Burdick JW, Andersen RA: Neural prosthetic control signals from plan activity. Neuroreport 2003, 14:591-596. This study showed that as few as 40 neurons from the posterior parietal cortex can provide a good prediction of the target of an upcoming movement. It also addressed the practical issue of identifying a ‘go’ signal as well as an intended target. 33. Musallam S, Corneil BD, Greger B, Scherberger H, Andersen RA: Cognitive control signals for neural prosthetics. Science 2004, 305:258-262. Signals recorded from the posterior parietal cortex and premotor cortex were shown in this study to contain information about several elements related to task performance, including movement, target and even motivational states. 34. Simpson RC, Levine SP: Automatic adaptation in the NavChair Assistive Wheelchair Navigation System. IEEE Trans Rehabil Eng 1999, 7:452-463. 35. Levine SP, Bell DA, Jaros LA, Simpson RC, Koren Y, Borenstein J: The NavChair Assistive Wheelchair Navigation System. IEEE Trans Rehabil Eng 1999, 7:443-451. 36. Fu Q-C, Flament D, Coltz JD, Ebner TJ: Temporal encoding of movement kinematics in the discharge of primate primary motor and premotor neurons. J Neurophysiol 1995, 73:836-854. 37. Wessberg J, Stambaugh CR, Kralik JD, Beck PD, Laubach M, Chapin JK, Kim J, Biggs SJ, Srinivasan MA, Nicolelis MA: Real-time prediction of hand trajectory by ensembles of cortical neurons in primates. Nature 2000, 408:361-365. 38. Edell DJ, Toi VV, McNeil VM, Clark LD: Factors influencing the biocompatibility of insertable silicon microshafts in cerebral cortex. IEEE Trans Biomed Eng 1992, 39:635-643. 39. Maynard EM, Fernandez E, Normann RA: A technique to prevent dural adhesions to chronically implanted microelectrode arrays. J Neurosci Methods 2000, 97:93-101. 40. Fofonoff TA, Martel SM, Hatsopoulos NG, Donoghue JP, Hunter IW: Microelectrode array fabrication by electrical discharge machining and chemical etching. IEEE Trans Biomed Eng 2004, 51:890-895. 41. Rousche PJ, Kipke DR: Next generation of cortical devices. Hackensack, NJ: World Scientific Publishing Co; 2004. 42. Santello M, Flanders M, Soechting JF: Postural hand synergies for tool use. J Neurosci 1998, 18:10105-10115. 43. Mason CR, Gomez JE, Ebner TJ: Hand synergies during reach-to-grasp. J Neurophysiol 2001, 86:2896-2910. 44. Mason CR, Theverapperuma LS, Hendrix CM, Ebner TJ: Monkey hand postural synergies during reach-to-grasp in the absence of vision of the hand and object. J Neurophysiol 2004, 91:2826-2837. www.sciencedirect.com