Abstract
A brain-computer interface records neural activity, decodes it in real time, and translates it into commands for an external device, creating an output pathway that bypasses peripheral nerves and muscles. Such systems serve both to restore communication and movement to people with severe paralysis and to probe how intention is encoded in the brain. This article defines what distinguishes a brain-computer interface from other neurotechnology, contrasts the invasive and non-invasive methods of acquiring the signal, and surveys the control paradigms — from event-related potentials and sensorimotor rhythms to single-neuron firing rates — on which they are built. It traces the field from its 1970s origins to intracortical prostheses that decode attempted handwriting and speech, and introduces information transfer rate. Three interactive demonstrations explore a matrix speller, cursor control, and the arithmetic of information transfer rate.
Keywords: brain-computer interface, neuroprosthetics, electroencephalography, information transfer rate, neural decoding
A brain-computer interface (BCI) is a system that measures activity in the central nervous system and converts it, in real time, into an artificial output that replaces, restores, or supplements the natural output of the brain (Wolpaw et al., 2002). The defining feature is the pathway itself: a conventional action reaches the world through peripheral nerves and muscles, whereas a brain-computer interface reads the neural signal directly and routes it to a computer, a robotic limb, or a stimulator, bypassing the body’s usual motor channels entirely. This is why the technology matters most to people who have lost those channels — through amyotrophic lateral sclerosis, brainstem stroke, or spinal cord injury — while retaining the cortical activity from which intention can be read (Chaudhary et al., 2016).
The idea was named before it was feasible. Vidal set out the prospect of direct brain-computer communication in 1973, asking whether the electrical signals of the brain could be put to work as carriers of information in a man-computer dialogue (Vidal, 1973). Half a century of work on signal acquisition, decoding algorithms, and implant durability has since turned the proposal into working systems that let paralysed users spell, steer a cursor, and control a robotic arm. The sections below define the interface precisely, distinguish the ways its signal is acquired, describe the neural phenomena used as control signals, and explain how the throughput of a given system is measured.
Key Takeaways
- A brain-computer interface creates a direct output pathway from the brain to an external device, bypassing peripheral nerves and muscles (Wolpaw et al., 2002).
- Signals are acquired along a spectrum of invasiveness, trading spatial resolution against surgical risk: scalp electroencephalography, electrodes on the cortical surface, and microelectrode arrays penetrating the cortex (Lebedev & Nicolelis, 2017).
- Control can be built on evoked potentials such as the P300, on voluntarily modulated sensorimotor rhythms, or on the firing of individual neurons (Farwell & Donchin, 1988; Pfurtscheller & Neuper, 2001; Hochberg et al., 2006).
- Information transfer rate — combining the number of choices, the accuracy, and the time per selection — is the standard measure of how much a brain-computer interface can convey (Wolpaw et al., 2002).
- Intracortical systems now decode attempted handwriting and speech at conversational rates, and a fully implanted device has restored communication in a locked-in patient (Vansteensel et al., 2016; Willett et al., 2021; Willett et al., 2023).
What a Brain-Computer Interface Is
In the Medical Subject Headings vocabulary, a brain-computer interface is instrumentation — hardware together with software — that establishes a direct communication pathway between the brain and an external device, recording neural signals and converting them into commands. The essential requirement, and the one that separates a brain-computer interface from ordinary biofeedback or neural monitoring, is that the system must produce an output the user controls: the neural signal has to be turned into an action in the world, in real time, under the user’s intention, with feedback closing the loop so that performance can improve (Wolpaw et al., 2002).
Three properties follow from this definition. First, a brain-computer interface reads a signal generated by the central nervous system, not by muscles; a device driven by residual eye movements or faint muscle twitches, however useful, is not a brain-computer interface, because its input is not neural. Second, the translation is closed-loop: the user perceives the result of each command and adjusts, and the decoder is often retrained on the user’s activity, so the human and the machine adapt to each other (Chaudhary et al., 2016). Third, the purpose is to replace or restore an output that the nervous system can no longer express through its normal channels, which is why the clinical motivation — communication and control for people with severe motor impairment — has driven the field from the outset.
It is worth distinguishing the interface from the neurostimulation technologies with which it is often grouped. A brain-computer interface is primarily a recording and decoding system, reading activity outward; a device such as a cochlear implant or a deep-brain stimulator writes activity inward. Many modern systems do both — sending sensory feedback back to cortex while reading motor intention out — but the interface is defined by the outward, brain-to-device pathway that makes volitional control possible.
Signal Acquisition
Every brain-computer interface begins with a choice about where to place the sensor, and that choice sets the ceiling on everything downstream. The options form a spectrum of invasiveness in which spatial and temporal resolution are bought at the price of surgical risk (Lebedev & Nicolelis, 2017).
At the non-invasive end sits electroencephalography (EEG), which records the summed electrical activity of large populations of neurons from electrodes on the scalp. It requires no surgery and is inexpensive and portable, which has made it the workhorse of brain-computer interface research, but the skull blurs and attenuates the signal, so EEG offers coarse spatial resolution and a low signal-to-noise ratio. Most non-invasive systems therefore rely on comparatively slow, robust signals — evoked potentials and rhythm changes — rather than fine-grained motor detail (Wolpaw et al., 2002).
Electrocorticography (ECoG) places an electrode array directly on the surface of the cortex, beneath the skull, capturing higher-frequency activity and finer spatial detail than the scalp allows while not penetrating the brain tissue itself. A fully implanted electrocorticographic system has been used to give a person with late-stage amyotrophic lateral sclerosis reliable, home-based control of a spelling program (Vansteensel et al., 2016).At the invasive end, intracortical microelectrode arrays are inserted into the cortex to record the firing of individual neurons and small ensembles. This yields by far the richest signal — enough to reconstruct intended movements in several dimensions — at the cost of surgery and of the long-term challenge of keeping electrodes stable and functional in living tissue. The clinical intracortical work most associated with this approach records from arrays in motor cortex, and it is the basis of the high-performance systems described later (Hochberg et al., 2006; Hochberg et al., 2012).
The Signal-Acquisition Spectrum
| Method | Sensor placement | Signal resolution | Surgical risk | Typical control signals |
|---|---|---|---|---|
| Electroencephalography | On the scalp; non-invasive. | Coarse; skull blurs and attenuates the signal. | None. | Evoked potentials (P300), sensorimotor rhythms, slow cortical potentials. |
| Electrocorticography | On the cortical surface, beneath the skull. | Intermediate; higher-frequency, finer detail without penetrating tissue. | Moderate; requires craniotomy. | High-gamma activity, attempted speech and movement. |
| Intracortical array | Penetrating the cortex, in motor areas. | Finest; individual-neuron and small-ensemble firing. | High; implantation plus long-term stability problems. | Population firing rates for cursor, robotic-arm, handwriting, and speech decoding. |
Control Signals and Paradigms
Having acquired a signal, the system needs a neural phenomenon that the user can reliably produce and the decoder can reliably detect. Three broad strategies dominate.
The first uses event-related potentials, most famously the P300, a positive deflection arising roughly 300 milliseconds after a rare, attended stimulus. In the classic matrix speller, the rows and columns of a grid of letters flash in turn; the row and column containing the letter the user is attending to elicit a P300, and their intersection identifies the target (Farwell & Donchin, 1988). Because the P300 appears automatically in response to an attended oddball, this paradigm needs little training, though each selection requires many flashes to average out noise.
The second uses sensorimotor rhythms — oscillations over motor cortex, in the mu (roughly 8–12 Hz) and beta (roughly 13–30 Hz) bands, whose power falls when a person moves or imagines moving a limb. Through motor imagery — imagining movement of, say, the left versus the right hand without executing it — a trained user can voluntarily modulate these rhythms and so drive a cursor or a switch; the phenomenon is a learned skill that improves with feedback (Pfurtscheller & Neuper, 2001). A related, slower signal is the slow cortical potential, a shift in cortical voltage over seconds that patients can be trained to control; an early device of this kind, the Thought Translation Device, let paralysed users select letters and was among the first to give people who were locked in a means of written communication (Birbaumer et al., 1999).
The third strategy reads the firing of individual neurons directly, using intracortical arrays. Because populations of motor-cortex neurons encode the direction and speed of intended movement, their activity can be decoded into continuous control of a cursor or a robotic limb, and even into the reach and grasp needed to pick up an object (Hochberg et al., 2012; Collinger et al., 2013). This single-neuron approach delivers the highest performance and underlies the recent communication systems that decode attempted handwriting and speech.
Figure 1
The Closed Loop of a Brain-Computer Interface
The first demonstration builds a matrix speller of the P300 kind: choosing a target letter and running the row-and-column flashes shows how the intersection of the two attended flashes identifies the letter, and how averaging over more flash cycles sharpens the detection.
A P300 matrix speller
Click a letter to attend to it, then raise the number of flash repetitions. Each attended row and column carries a faint P300 hidden in noise; averaging over more repetitions sharpens the estimate until the detected intersection (navy) lands on your target (gold outline).
Averaging N repetitions cuts the noise standard deviation by a factor of √N, so a signal invisible in a single flash becomes reliable once enough repetitions are combined. This is why a P300 speller trades speed for accuracy. The noise is fixed-seeded, so the result depends only on the target and the repetition count.
Sensorimotor-rhythm control is continuous rather than discrete. The second demonstration maps the imagined-movement signal — the difference in rhythm power between two hemispheres — onto the velocity of a cursor, so that steering toward a target becomes a matter of modulating that difference, exactly the skill a trained user acquires.
Cursor control from a sensorimotor rhythm
Imagined movement of one hand versus the other shifts the balance of mu-rhythm power between the hemispheres. Set that control signal and pick a target side; the signal biases a noisy cursor left or right, and reaching the intended edge in time is a successful selection.
Control is continuous, not a single choice: the user holds a bias in the rhythm long enough to carry the cursor to a target against the noise. Weak or wrong-signed bias lets noise decide. The step noise is fixed-seeded, so the trajectory depends only on the control you set.
Measuring Performance
Because brain-computer interfaces differ so widely — a binary switch, a 36-character speller, a continuous cursor — comparing them requires a common currency, and the field’s standard is information transfer rate (ITR), expressed in bits per minute. It combines three quantities: the number of possible selections, the accuracy of each selection, and the time each selection takes. A system with many choices conveys more information per correct selection, but errors are costly and slowness drags the rate down, so the measure captures the real trade-off a user faces (Wolpaw et al., 2002).
The bits conveyed per selection, for a system with N equally likely targets selected with accuracy p, is given by the expression Wolpaw and colleagues adapted from information theory. Multiplying by the number of selections per minute converts it to a rate. The third demonstration computes this directly: adjusting the number of targets, the accuracy, and the time per selection shows how each pushes the bit rate up or down, and why very high accuracy matters more than it first appears.
Information transfer rate
The throughput of a brain-computer interface combines how many choices it offers, how accurately each is made, and how long each takes. Adjust the three and watch bits per minute respond — and see how sharply errors below perfect accuracy cut the rate.
B = log₂N + p log₂p + (1 − p) log₂[(1 − p) / (N − 1)], rate = B × 60 / T. The default N = 6, p = 0.90, T = 5 s gives 1.88 bits per selection and 22.6 bits per minute, matching the worked example. Notice how the gap between the two bars widens fast as accuracy drops.
Worked Example
Consider a matrix speller offering N = 6 choices, at which a user makes one selection every T = 5 seconds with an accuracy of p = 0.90. How much information does it convey?
The bits per selection are given by B = log₂N + p log₂p + (1 − p) log₂[(1 − p) / (N − 1)]. Taking each term in turn: log₂6 = 2.585 bits, the information of a perfect 6-way choice. The second term is 0.90 × log₂0.90 = 0.90 × (−0.152) = −0.137. The third term is 0.10 × log₂(0.10 / 5) = 0.10 × log₂0.02 = 0.10 × (−5.644) = −0.564. Summing, B = 2.585 − 0.137 − 0.564 = 1.88 bits per selection.
The 10% error rate has cost nearly a third of the 2.585 bits a perfect selection would carry — a vivid illustration of why accuracy dominates the measure. Converting to a rate, the user makes 60 / 5 = 12 selections per minute, so the information transfer rate is 1.88 × 12 = 22.6 bits per minute. Raising accuracy to 100% while holding the speed would lift the rate to 2.585 × 12 = 31.0 bits per minute, showing how much throughput those errors consume.
Discussion
The trajectory of the field is a steady exchange of surgical risk for performance. Non-invasive systems ask nothing of the user surgically and have brought communication to many, but the blurred scalp signal caps their throughput, and even a good P300 speller conveys only tens of bits per minute (Wolpaw et al., 2002). Intracortical systems, by reading individual neurons, achieve far higher rates and richer control — multi-dimensional reach and grasp, and now decoded speech — but they require neurosurgery and confront the unsolved problem of keeping electrodes recording stably for years (Lebedev & Nicolelis, 2017). Much current engineering is aimed squarely at that middle ground: fully implanted, chronically stable systems that a person can use unsupervised at home (Vansteensel et al., 2016).
Two framings of the technology coexist. Clinically, a brain-computer interface is an assistive and restorative device, returning communication and movement to people who have lost them (Chaudhary et al., 2016). Scientifically, it is a research instrument: building a decoder that must work in real time forces an explicit, testable account of what motor and language cortex actually encodes, so each advance in decoding is also a claim about neural representation (Pandarinath et al., 2017). The clinical and the basic-science goals reinforce each other, which is why progress in one so often follows from the other.
Current Directions
The most striking recent progress has come from reframing communication as the decoding of attempted movement rather than the slow selection of symbols. Rather than spelling by cursor, a user attempts to write letters by hand, and a recurrent-network decoder reads the intended pen strokes from motor cortex; this brain-to-text-by-handwriting system reached typing speeds far beyond earlier point-and-select spellers (Willett et al., 2021). The same logic applied to speech has produced neuroprostheses that decode attempted speaking — both as text and as synthesised voice and facial animation — approaching conversational rates in people who can no longer speak (Willett et al., 2023; Metzger et al., 2023).
In parallel, the hardware is scaling. Efforts to increase the number of recording channels by orders of magnitude, and to make the implant compact and wireless, aim to raise both the bandwidth and the practicality of chronic systems (Musk & Neuralink, 2019). The open problems are less about the ceiling of what can be decoded in the laboratory than about durability, autonomy, and reach: keeping a system working for years without expert recalibration, letting users operate it independently, and extending access beyond the small number of research participants implanted so far (Vansteensel et al., 2016).
Common Misconceptions
- “A brain-computer interface can read a person’s private thoughts.”
- Current systems decode specific, trained signals — an attended flash, an imagined movement, an attempted word — not free-ranging private thought. They detect patterns in motor or attentional activity that the user deliberately produces, and they work only for the narrow repertoire the decoder was trained on (Wolpaw et al., 2002).
- “Non-invasive interfaces are simply worse than implanted ones.”
- They occupy different points on a risk-performance trade-off. Scalp electroencephalography sacrifices resolution but needs no surgery and reaches far more people; intracortical arrays offer much higher performance at the cost of an operation and long-term stability problems. Which is appropriate depends on the user and the task (Lebedev & Nicolelis, 2017).
- “A device driven by eye or muscle signals is a brain-computer interface.”
- By definition it is not. A brain-computer interface reads activity from the central nervous system directly; a system driven by residual eye movements or muscle twitches uses the ordinary neuromuscular pathway the interface is meant to bypass (Wolpaw et al., 2002).
- “These systems are ready for everyday consumer use.”
- The high-performance results come from a small number of research participants under expert supervision, and chronic stability, independent home use, and regulatory approval remain open challenges. The clearest benefit so far is for people with severe paralysis, not the general public (Vansteensel et al., 2016).
Glossary
- Brain-computer interface.
- A system that records central-nervous-system activity and translates it in real time into commands for an external device, creating an output pathway that bypasses peripheral nerves and muscles.
- Closed loop.
- The arrangement in which the result of each decoded command is fed back to the user, who adjusts the next intention, allowing both the user and the decoder to improve with practice.
- Decoder.
- The algorithm that maps a recorded neural signal onto an intended command, often retrained on the individual user’s activity to track changes in the signal.
- Electrocorticography (ECoG).
- Recording of electrical activity from electrodes placed directly on the surface of the cortex, offering finer detail than scalp recording without penetrating brain tissue.
- Electroencephalography (EEG).
- Non-invasive recording of the brain’s summed electrical activity from scalp electrodes; inexpensive and portable but limited in spatial resolution by the intervening skull.
- Event-related potential.
- A stereotyped voltage change in the electroencephalogram time-locked to a specific event; the P300, a positive deflection to a rare attended stimulus, is widely used for control.
- Information transfer rate (ITR).
- The standard measure of a brain-computer interface’s throughput, in bits per minute, combining the number of choices, the accuracy of each selection, and the time per selection.
- Intracortical array.
- A grid of microelectrodes inserted into the cortex to record the firing of individual neurons and small ensembles, yielding the richest signal at the cost of surgery and long-term instability.
- Locked-in syndrome.
- A condition of near-total paralysis with preserved awareness, in which a person cannot move or speak yet remains cognitively intact; a principal clinical target for brain-computer interfaces.
- Matrix speller.
- A communication interface in which the rows and columns of a character grid flash in turn, the target letter being identified from the event-related potentials evoked by the attended flashes.
- Motor imagery.
- The mental rehearsal of a movement without executing it, which modulates sensorimotor rhythms over motor cortex and so can serve as a voluntary control signal.
- Neuroprosthesis.
- A device that restores a lost neural function; in this context, a brain-computer interface that returns communication or movement, such as a decoded-speech or robotic-arm system.
- Sensorimotor rhythm.
- Oscillatory activity over motor cortex in the mu and beta bands whose power drops during real or imagined movement, providing a signal a trained user can modulate at will.
- Slow cortical potential.
- A gradual shift in cortical voltage over seconds that a user can learn to control, used in early spelling interfaces for people who are locked in.
Key Researchers
Niels Birbaumer
(born 1945). Neuroscientist and psychologist at the University of Tübingen; a pioneer of slow-cortical-potential and electroencephalographic interfaces, whose Thought Translation Device gave locked-in and ALS patients an early means of written communication (Birbaumer et al., 1999). Faculty · Google Scholar · Wikipedia
Edward F. Chang
(living). Neurosurgeon and neuroscientist at the University of California, San Francisco; leader of speech-decoding research and senior author of a neuroprosthesis that decodes attempted speech into text, synthesised voice, and avatar animation (Metzger et al., 2023). Faculty · Wikipedia
Leigh R. Hochberg
(living). Neurologist and neuroengineer at Brown University and Massachusetts General Hospital; director of the BrainGate clinical trials and senior author of the intracortical studies that first let people with tetraplegia control a cursor and a robotic arm (Hochberg et al., 2006; Hochberg et al., 2012). ORCID · Faculty · Google Scholar · Wikipedia
Miguel A. L. Nicolelis
(born 1961). Neuroscientist and Professor Emeritus at Duke University; a pioneer of neural-ensemble recording and brain-machine interfaces, and co-author of a comprehensive review of the field from basic science to neurorehabilitation (Lebedev & Nicolelis, 2017). Faculty · Google Scholar · Wikipedia
Krishna V. Shenoy
(1968–2023). Electrical engineer and neuroscientist at Stanford University; director of the Neural Prosthetic Systems Lab and co-senior author of the intracortical handwriting and speech neuroprostheses that decode attempted movement at high speed (Willett et al., 2021; Willett et al., 2023). ORCID · Faculty · Google Scholar · Wikipedia
Jonathan R. Wolpaw
(living). Neurologist and director of the National Center for Adaptive Neurotechnologies; author of the canonical review that defined the field’s terms and its throughput measure, and a pioneer of sensorimotor-rhythm interfaces (Wolpaw et al., 2002). Faculty · Google Scholar
Frequently Asked Questions
What is a brain-computer interface?
It is a system that records activity from the brain, decodes it in real time, and turns it into commands for an external device such as a computer, a speller, or a robotic arm. The defining feature is that it creates an output pathway directly from neural activity, bypassing the peripheral nerves and muscles that normally carry out an action.
Who benefits most from a brain-computer interface?
The primary clinical beneficiaries are people with severe motor impairment who retain cortical function, such as those with amyotrophic lateral sclerosis, brainstem stroke, or high spinal cord injury. For someone who can no longer move or speak, the interface can restore a channel for communication or control that the body can no longer provide.
What is the difference between invasive and non-invasive interfaces?
Non-invasive interfaces, chiefly scalp electroencephalography, require no surgery and are inexpensive, but the skull blurs the signal and limits performance. Invasive interfaces place electrodes on or in the cortex, capturing far richer activity and enabling higher performance, at the cost of surgery and the difficulty of keeping electrodes stable over time.
How do P300 spellers work?
A grid of letters has its rows and columns flash in sequence. When the row or column containing the letter a user is attending to flashes, it evokes a P300 response, a brain potential arising about 300 milliseconds after an attended rare event. The system identifies the target letter as the intersection of the row and the column that produced this response.
What is information transfer rate?
It is the standard measure of how much a brain-computer interface can convey, expressed in bits per minute. It combines the number of possible choices, the accuracy of each selection, and the time each selection takes, capturing the trade-off between speed, precision, and the range of options a system offers.
Can a brain-computer interface read my private thoughts?
No. Existing systems decode specific trained signals, such as an imagined hand movement or an attempted word, not unconstrained inner thought. They recognise deliberate patterns of motor or attentional activity within the narrow repertoire on which the decoder was trained, and they do not work outside it.
Can these systems restore speech?
Recent intracortical and cortical-surface systems decode attempted speech into text and even into synthesised voice with facial animation, reaching rates approaching natural conversation in people who can no longer speak. These are advanced research results in a small number of participants rather than widely available clinical devices.
Are brain-computer interfaces available to the general public?
Not for everyday use. The most capable systems operate in research settings with expert support and a small number of implanted participants, and challenges of long-term stability, independent home use, and regulatory approval remain. The clearest established benefit is for people with severe paralysis.
References
Birbaumer, N., Ghanayim, N., Hinterberger, T., Iversen, I., Kotchoubey, B., Kübler, A., Perelmouter, J., Taub, E., & Flor, H. (1999). A spelling device for the paralysed. Nature, 398(6725), 297–298. https://doi.org/10.1038/18581
Chaudhary, U., Birbaumer, N., & Ramos-Murguialday, A. (2016). Brain-computer interfaces for communication and rehabilitation. Nature Reviews Neurology, 12(9), 513–525. https://doi.org/10.1038/nrneurol.2016.113
Collinger, J. L., Wodlinger, B., Downey, J. E., Wang, W., Tyler-Kabara, E. C., Weber, D. J., McMorland, A. J. C., Velliste, M., Boninger, M. L., & Schwartz, A. B. (2013). High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet, 381(9866), 557–564. https://doi.org/10.1016/S0140-6736(12)61816-9
Farwell, L. A., & Donchin, E. (1988). Talking off the top of your head: Toward a mental prosthesis utilizing event-related brain potentials. Electroencephalography and Clinical Neurophysiology, 70(6), 510–523. https://doi.org/10.1016/0013-4694(88)90149-6
Hochberg, 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. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature, 442(7099), 164–171. https://doi.org/10.1038/nature04970
Hochberg, L. R., Bacher, D., Jarosiewicz, B., Masse, N. Y., Simeral, J. D., Vogel, J., Haddadin, S., Liu, J., Cash, S. S., van der Smagt, P., & Donoghue, J. P. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature, 485(7398), 372–375. https://doi.org/10.1038/nature11076
Lebedev, M. A., & Nicolelis, M. A. L. (2017). Brain-machine interfaces: From basic science to neuroprostheses and neurorehabilitation. Physiological Reviews, 97(2), 767–837. https://doi.org/10.1152/physrev.00027.2016
Metzger, S. L., Littlejohn, K. T., Silva, A. B., Moses, D. A., Seaton, M. P., Wang, R., Dougherty, M. E., Liu, J. R., Wu, P., Berger, M. A., Zhuravleva, I., Tu-Chan, A., Ganguly, K., Anumanchipalli, G. K., & Chang, E. F. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature, 620(7976), 1037–1046. https://doi.org/10.1038/s41586-023-06443-4
Musk, E., & Neuralink. (2019). An integrated brain-machine interface platform with thousands of channels. Journal of Medical Internet Research, 21(10), e16194. https://doi.org/10.2196/16194
Pandarinath, C., Nuyujukian, P., Blabe, C. H., Sorice, B. L., Saab, J., Willett, F. R., Hochberg, L. R., Shenoy, K. V., & Henderson, J. M. (2017). High performance communication by people with paralysis using an intracortical brain-computer interface. eLife, 6, e18554. https://doi.org/10.7554/eLife.18554
Pfurtscheller, G., & Neuper, C. (2001). Motor imagery and direct brain-computer communication. Proceedings of the IEEE, 89(7), 1123–1134. https://doi.org/10.1109/5.939829
Vansteensel, M. J., Pels, E. G. M., Bleichner, M. G., Branco, M. P., Denison, T., Freudenburg, Z. V., Gosselaar, P., Leinders, S., Ottens, T. H., Van Den Boom, M. A., Van Rijen, P. C., Aarnoutse, E. J., & Ramsey, N. F. (2016). Fully implanted brain-computer interface in a locked-in patient with ALS. New England Journal of Medicine, 375(21), 2060–2066. https://doi.org/10.1056/NEJMoa1608085
Vidal, J. J. (1973). Toward direct brain-computer communication. Annual Review of Biophysics and Bioengineering, 2, 157–180. https://doi.org/10.1146/annurev.bb.02.060173.001105
Willett, F. R., Avansino, D. T., Hochberg, L. R., Henderson, J. M., & Shenoy, K. V. (2021). High-performance brain-to-text communication via handwriting. Nature, 593(7858), 249–254. https://doi.org/10.1038/s41586-021-03506-2
Willett, F. R., Kunz, E. M., Fan, C., Avansino, D. T., Wilson, G. H., Choi, E. Y., Kamdar, F., Glasser, M. F., Hochberg, L. R., Druckmann, S., Shenoy, K. V., & Henderson, J. M. (2023). A high-performance speech neuroprosthesis. Nature, 620(7976), 1031–1036. https://doi.org/10.1038/s41586-023-06377-x
Wolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G., & Vaughan, T. M. (2002). Brain-computer interfaces for communication and control. Clinical Neurophysiology, 113(6), 767–791. https://doi.org/10.1016/S1388-2457(02)00057-3