Handbook of Clinical Neurology, Vol.

168 (3rd series)

Chapter 2

Brain-computer interfaces: Definitions and principles

1
National Center for Adaptive Neurotechnologies and Stratton VA Medical Center, Wadsworth Center, Albany, NY, United States
2
Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, United States
3
Department of Neurology, The University of Texas at Austin, Austin, TX, United States
4
Brain Center, University Medical Center Utrecht, Utrecht, The Netherlands

Abstract
Throughout life, the central nervous system (CNS) interacts with the world and with the body by activating
muscles and excreting hormones. In contrast, brain-computer interfaces (BCIs) quantify CNS activity and
translate it into new artificial outputs that replace, restore, enhance, supplement, or improve the natural
CNS outputs. BCIs thereby modify the interactions between the CNS and the environment. Unlike the
natural CNS outputs that come from spinal and brainstem motoneurons, BCI outputs come from brain
signals that represent activity in other CNS areas, such as the sensorimotor cortex. If BCIs are to be useful
for important communication and control tasks in real life, the CNS must control these brain signals nearly
as reliably and accurately as it controls spinal motoneurons. To do this, they might, for example, need to
incorporate software that mimics the function of the subcortical and spinal mechanisms that participate in
normal movement control. The realization of high reliability and accuracy is perhaps the most difficult and
critical challenge now facing BCI research and development.
The ongoing adaptive modifications that maintain effective natural CNS outputs take place primarily in
the CNS. The adaptive modifications that maintain effective BCI outputs can also take place in the BCI. This
means that the BCI operation depends on the effective collaboration of two adaptive controllers, the CNS and
the BCI. Realization of this second adaptive controller, the BCI, and management of its interactions with
concurrent adaptations in the CNS comprise another complex and critical challenge for BCI development.
BCIs can use different kinds of brain signals recorded in different ways from different brain areas. Deci-
sions about which signals recorded in which ways from which brain areas should be selected for which
applications are empirical questions that can only be properly answered by experiments.
BCIs, like other communication and control technologies, often face artifacts that contaminate or imi-
tate their chosen signals. Noninvasive BCIs (e.g., EEG- or fNIRS-based) need to take special care to avoid
interpreting nonbrain signals (e.g., cranial EMG) as brain signals. This typically requires comprehensive
topographical and spectral evaluations.
In theory, the outputs of BCIs can select a goal or control a process. In the future, the most effective BCIs
will probably be those that combine goal selection and process control so as to distribute control between
the BCI and the application in a fashion suited to the current action. Through such distribution, BCIs may
most effectively imitate natural CNS operation.
The primary measure of BCI development is the extent to which BCI systems benefit people with neu-
romuscular disorders. Thus, BCI clinical evaluation, validation, and dissemination is a key step. It is at the
same time a complex and difficult process that depends on multidisciplinary collaboration and manage-
ment of the demanding requirements of clinical studies.

*Correspondence to: Jonathan R. Wolpaw, M.D., DAI, PO Box 509, 120 New Scotland Ave, David Axelrod Institute, Rm 2085,
Albany, NY, 12208, United States. Tel: +1-518-473-3631, Fax: +1-518-486-4910, E-mail: [Link]@[Link]
16 J.R. WOLPAW ET AL.
Twenty-five years ago, BCI research was an esoteric endeavor pursued in only a few isolated labora-
tories. It is now a steadily growing field that engages many hundreds of scientists, engineers, and clinicians
throughout the world in an increasingly interconnected community that is addressing the key issues and
pursuing the high potential of BCI technology.

DEFINITIONS lateral sclerosis (ALS) to spell words that are then spoken
by a speech synthesizer (see also Chapter 7).
Brain-computer interface BCI output could restore lost muscle control.
The central nervous system (CNS) receives sensory A person in whom a spinal cord injury has paralyzed
inputs and produces appropriate motor outputs. Its natu- arms and hands might grasp objects by using a BCI that
ral outputs consist of muscle activity and hormones. stimulates the paralyzed muscles through implanted elec-
Brain-computer interfaces (BCIs) give the CNS novel trodes (see Chapters 13 and 22). Or, in someone in whom
outputs that are neither neuromuscular nor hormonal. multiple sclerosis has eliminated bladder function, a BCI
A BCI is a system that records CNS activity and translates might stimulate the peripheral nerves to the bladder to
it into artificial output that replaces, restores, enhances, produce urination.
supplements, or improves natural CNS outputs; it thereby BCI output could enhance natural CNS output. For a
modifies the interactions of the CNS with the rest of the person driving a vehicle or performing sentry duty, a BCI
body or with the external world. that detects features in brain activity that precede lapses
This definition is based on reviews published over the in attention might sound an alarm that restores attention.
past 15 years (Donoghue, 2002; Wolpaw et al., 2002; It could thereby enhance the individual’s normal capacity
Schwartz, 2004; K€ ubler and M€ uller, 2007; Daly and for continuous attention.
Wolpaw, 2008; Graimann et al., 2010a; Millán et al., BCI output could supplement natural CNS output.
2010); it was formalized in 2012 (Wolpaw and A person who is using a joystick to control the move-
Wolpaw, 2012). It includes the many different kinds of ments of a cursor might use a BCI to choose items the
BCI systems now being developed and the many differ- cursor reaches. Or a person might conceivably control
ent purposes to which they are being, or might conceiv- a third (robotic) arm with a BCI (see Chapter 13).
ably be applied. It also places BCIs within the theoretical Finally, BCI output could potentially improve natural
framework of modern neuroscience. The foundation of CNS output. In a person who has suffered a stroke that
this framework is the sensorimotor hypothesis, the impairs arm function, a BCI might measure activity in
hypothesis that the entire function of the CNS is to trans- the damaged cortical area during attempted movements
late sensory inputs into motor outputs (Young, 1990; and use it to stimulate muscles or control an orthotic
Wolpaw, 2002). BCIs are systems that translate brain device so as to improve arm movement. With repeated
signals into novel types of outputs. use, this strategy might guide activity-dependent plastic-
The signals that BCIs measure are due to the electro- ity that restores more normal movement control (see
physiologic, neurochemical, and metabolic phenomena Chapter 9).
(such as neuronal action potentials, synaptic potentials, Until recently, the first two BCI uses—replacement or
release of neurotransmitters, and oxygen uptake) that restoration of lost natural output—were the foci of almost
are continually occurring in the CNS. The signals are all BCI research and development; the fifth use—to
measured by using sensors on the scalp, on the surface improve natural output that has been impaired by trauma
of the brain, or within the brain to monitor electric or or disease—is now engaging much interest and effort
magnetic fields, blood flow, hemoglobin oxygenation, (Bockbrader et al., 2018; Lazarou et al., 2018). The third
or other phenomena. A BCI records these brain signals, and fourth uses—to enhance or supplement normal
derives particular measures (or features) from them, and output—are also receiving attention (see Chapters 10
translates the features into novel CNS outputs. Fig. 2.1 and 15).
shows the five kinds of uses to which a BCI output might The final part of the definition is a fundamental fea-
be applied. ture of any BCI: it changes the interactions between
BCI output could replace muscle control lost to injury the CNS and its external or internal environment. Nor-
or disease. A BCI might enable a person paralyzed by a mally, these interactions comprise CNS motor outputs
high-level spinal cord injury to control a motorized to the environment and sensory inputs from the environ-
wheelchair (see Chapter 8). Or a BCI might enable some- ment. By measuring CNS activity and translating it
one who cannot speak due to a brainstem stroke or a into new outputs that affect the environment, BCIs
degenerative neuromuscular disorder like amyotrophic change both the outputs and the inputs. The changes
 BRAIN-COMPUTER INTERFACES 17

Fig. 2.1. Brain-computer interface (BCI) design and operation. Electrical or other signals reflecting brain activity are recorded
from the scalp, the cortical surface, or within the brain. They are analyzed to measure signal features (e.g., single-neuron firing
rates, amplitudes of EEG rhythms) that indicate the BCI user’s intent. The features are translated into commands that operate appli-
cations that replace, restore, enhance, supplement, or improve natural CNS outputs. From Wolpaw, J.R., Wolpaw, E.W. (2012).
Brain-computer interfaces: something new under the sun. In: Wolpaw, J.R. and Wolpaw, E.W. (Eds.), Brain-computer interfaces:
principles and practice. Oxford University Press, 3–12.

in sensory input are commonly termed feedback. The terms dependent BCI and independent BCI were
Devices that only monitor brain activity without using coined in 2002 (Wolpaw et al., 2002) to define BCIs that
it to modify the interactions between the CNS and its differ in their dependence on normal (i.e., neuromuscu-
environment are not regarded as BCIs. lar) CNS outputs. A dependent BCI uses brain signals
that depend on muscle activity. The early BCI developed
by Vidal (1973, 1977) used a visual evoked potential
Related terms
(VEP) that depended on gaze direction, and thus on
BCIs are often called brain–machine interfaces (BMIs). the muscles that move the eyes. While it does not provide
While BCI and BMI are essentially synonymous terms, a new CNS output that is independent of natural outputs,
systems that use externally recorded signals (e.g., EEG, it can still be valuable (e.g., Sutter, 1992). An indepen-
Chapter 18) are commonly referred to as BCIs, and sys- dent BCI does not depend on normal CNS output; muscle
tems that use signals recorded by implanted sensors are activity is not needed for generating the brain signals that
often referred to as BMIs. In general, BCI might be con- the BCI measures. In BCIs that use sensorimotor rhythms
sidered the preferable term, because “machine” implies a (SMRs) (e.g., McFarland et al., 2010), actual muscle
fixed conversion of brain signals into outputs; thus, it activity is not needed; the brain signals alone are suffi-
does not recognize that the system and the brain are part- cient, even if they do not result in actual movement
ners in the interactive adaptive control that is essential for (e.g., because of spinal cord injury or loss of spinal
successful BCI (or BMI) function (see following text). motoneurons in ALS). For those with the most severe
18 J.R. WOLPAW ET AL.
neuromuscular disabilities, such as in ALS, independent requirement is apparent when the BCI operation is
BCIs are likely to be more valuable. At the same time, it addressed in terms of how the CNS normally operates.
should be recognized that most BCIs are neither Research over the past two centuries, and particularly
completely dependent nor completely independent. in recent decades, has revealed two principles regarding
The output produced by a VEP-based BCI may reflect how the CNS produces its natural outputs.
the person’s attention rather than merely gaze direction The first principle is that the creation of natural out-
(e.g., Allison et al., 2008); and many SMR-based BCIs puts is distributed throughout the CNS, from cerebrum to
rely on the person having sufficient gaze control to watch spinal cord. No individual area is completely responsible
the results of the BCI’s outputs (e.g., cursor movements). for a natural output. The initiation, formulation, and pro-
The term hybrid BCI is employed in two different duction of behaviors such as walking, speaking, or play-
ways (Graimann et al., 2010b; M€ uller-Putz et al., ing the piano involve intricate interactions among
2015; Choi et al., 2017). It can be applied to a BCI that cortical regions, basal ganglia, thalamic nuclei, cerebel-
uses two different kinds of brain signals (e.g., VEPs and lum, brainstem nuclei, and spinal interneurons and moto-
SMRs (Ma et al., 2017)) to produce its outputs. Or it can neurons. Thus, while cortical areas initiate walking and
be applied to a system that combines a BCI output with a monitor its performance, the rhythmic high-speed senso-
natural muscle-based output (e.g., Gao et al., 2017). rimotor interactions that produce effective walking
depend largely on spinal circuitry (Ijspeert, 2008;
Adaptive neurotechnologies McCrea and Ryback, 2008; Guertin and Steuer, 2009;
Zehr et al., 2009). The result of this broadly distributed
BCIs fit within the broad class of adaptive neurotechnol- activity is appropriate excitation of the spinal (and/or
ogies. These systems bypass natural CNS input and out- brainstem) motoneurons that excite muscles to produce
put pathways (i.e., peripheral nerves and muscles) to the behavior. Although activity in the many contributing
create artificial interactions with the CNS that replace, CNS areas frequently correlates with the motor perfor-
restore, enhance, supplement, or improve natural interac- mance, the activity in a given area may vary widely from
tions. They adapt to optimize the new interaction and one instance (one performance of an action) to the next.
they often induce adaptive plasticity in the CNS, which Nevertheless, the coordinated activity in all the areas
also helps optimize the interaction. Adaptive neuro- ensures that the behavior itself is stable across multiple
technologies offer unprecedented opportunities to gain performances.
new scientific insights and realize new therapeutic The second principle is that the actions produced by
methods. Some act directly on the CNS (see natural CNS outputs are acquired and maintained by ini-
Chapter 25); thus, they contrast with BCIs, which enable tial and ongoing adaptations in all the CNS regions
the CNS to act directly on the world. At the same time, involved. Early in development and throughout life, neu-
some of these systems (e.g., those that stimulate cortical rons and synapses everywhere in the CNS change to mas-
or subcortical sensory areas) may be incorporated into ter new behaviors (new skills) and to maintain those
future BCI systems to improve their performance (Chap- previously mastered (e.g., Carroll and Zukin, 2002;
ters 13 and 22). Gaiarsa et al., 2002; Vaynman and Gomez-Pinilla,
2005; Saneyoshi et al., 2010; Wolpaw, 2010, 2018). Such
KEY ISSUES activity-dependent plasticity is responsible for mastering
Effective BCI research and development requires atten- and preserving common behaviors such as walking and
tion to a set of important issues. A number of the most talking, as well as special skills such as dancing and sing-
fundamental ones are summarized here. ing; it is guided by the results produced. Thus, as muscle
strength, limb length, and body weight change with
growth and aging, CNS regions adapt in order to preserve
BCIs create artificial CNS outputs
these behaviors. The CNS anatomy and physiology that
The CNS has evolved to produce effective muscular and underlies this ongoing adaptation is the product of evo-
hormonal outputs; and it adapts continually throughout lution guided by the necessity to produce effective
life to maintain them. BCIs give the CNS additional arti- actions (i.e., to appropriately control the spinal motoneu-
ficial outputs that are derived from brain signals. Thus, rons that excite the muscles).
they ask the CNS to produce entirely new kinds of out- Given these two principles—the numerous regions
puts. For example, sensorimotor cortical regions, which that contribute to normal CNS outputs and the ongoing
normally work together with subcortical and spinal areas adaptation of these regions—BCI use is a unique chal-
to control muscles, are now asked instead to control spe- lenge for a CNS that has evolved and continually adapts
cific brain signals (e.g., EEG rhythms, single-neuron fir- to acquire and maintain actions produced by the natural
ing patterns). The fundamental implications of this CNS outputs. In contrast to natural CNS outputs, BCI
 BRAIN-COMPUTER INTERFACES 19
outputs are not spinal motoneuron activations that lead to microelectrodes recording from single cells may support
muscle contractions, rather they are signals that reflect detailed control of a robotic arm, but they may not func-
activity in a particular CNS region (such as the sensori- tion reliably over months or years. On the other hand,
motor cortex). In natural actions, activity in this region is electrodes on the cortical surface may provide only
only one of the many regions that collaborate to produce relatively simple control, such as selecting items on a
appropriate motoneuron activation. However, when this computer screen or moving a cursor, but they may remain
region’s signals control a BCI, they become the CNS out- robust stable signals for long periods. This trade-off
put: the region producing the signals that the BCI uses is may improve with maturation of translation methods
given the role that is normally performed by spinal (e.g., artificial intelligence). Nevertheless, it is important
motoneurons (i.e., it produces the final product, the to realize that successful BCI use requires effective
CNS output). How effectively the chosen region (e.g., the interaction between two adaptive controllers: the CNS
sensorimotor cortex) performs this new role is deter- and the BCI. Enabling such effective interaction between
mined by how well the multiple CNS regions that nor- CNS adaptations and concurrent BCI adaptations is
mally adapt to control spinal motoneurons can adapt to among the most difficult tasks of BCI research
control the relevant cortical neurons and synapses. For (McFarland and Krusienski, 2012; Perdikis et al., 2018).
instance, a BCI requires that the cerebellum (which nor-
mally contributes to ensuring that motoneurons activate
Selecting signal types and CNS regions
muscles so that movement is smooth, rapid, and accu-
rate) now help to ensure that those cortical neurons Brain signals recorded by a number of electrophysiologic
recorded by a microelectrode array produce patterns of and metabolic methods can serve as BCI inputs. These
action potentials that move a prosthetic limb smoothly, signal types differ widely in spatial resolution, frequency
rapidly, and accurately. The extent to which the cerebel- content, region of generation, and technical demands.
lum and other key regions can adapt to this new role is For example, the spatial resolution of electrophysiologic
unclear. The capabilities and usefulness that BCIs can methods range from the centimeter scale of electroen-
achieve depend to a considerable degree on this issue. cephalography (EEG), through the millimeter scale of
At the same time, it should be noted that even the very electrocorticography (ECoG), to the tens of microns
simple control that reliable BCIs are currently limited scale of neuronal action potentials. Each method has
to can benefit people with severe disabilities (e.g., unique advantages and disadvantages. Which will prove
Birbaumer et al., 1999; Vansteensel et al., 2016; most successful for which purpose is not yet well
Wolpaw et al., 2018). defined; the answers hinge on scientific, technical, clin-
The results up to the present indicate that the adapta- ical, and commercial factors.
tion needed to control activity in the CNS regions that Questions about signal selection are empirical ques-
produce the signals BCIs use is possible, but imperfect. tions to be answered by experiment, not by a priori
As a rule, BCI outputs are much less smooth, rapid, and assumptions about the theoretical superiority of one sig-
accurate than natural CNS outputs, and their moment- nal type or another. For BCIs, the crucial issue is which
to-moment, day-to-day, and week-to-week variation is signals are the best and most reliable indicators of the
high. These problems, especially the poor reliability, user’s intent (i.e., which signals comprise the best lan-
are major issues in BCI research and development. guage for communicating to the BCI the output that
the user desires). This issue can be resolved only by
experimental results; thus, prevailing sentiments are
BCI operation requires the effective
likely to shift as the field matures.
interaction of two adaptive controllers
Selection of the best CNS regions from which to
The activity-dependent plasticity that acquires and main- record the signals is also an empirical question. Up to
tains effective natural CNS outputs occurs mainly in the the present, most work has focused on signals from sen-
CNS. In contrast, effective BCI outputs depend on adap- sorimotor (and visual) cortical areas. Exploration of
tation in both the CNS and the BCI. Beyond adapting to other regions (e.g., Janssen and Scherberger, 2015;
the amplitudes, frequencies, and other characteristics of a Ming et al., 2017) is important, particularly because
user’s brain signals, a BCI might also adapt to enhance the sensorimotor cortices of many potential BCI users
the correlation between its outputs and the user’s inten- have been damaged by injury or disease and/or their
tions, to enhance the effectiveness of CNS adaptations, vision may be impaired. CNS regions may well differ
and/or to guide CNS adaptations. In general, there is a in their adaptive capabilities and in other factors that
trade-off between the complexity of the brain signals might affect their suitability as the origin of signals for
and the difficulty of reliably translating them into the BCIs. Finally, on a positive note, it appears that cortical
outputs desired by the user. For instance, intracortical areas that have been deprived of their normal function for
20 J.R. WOLPAW ET AL.
years (e.g., by high-level spinal cord injury or ALS) can results can be difficult to interpret and of questionable
still provide effective signals for a BCI (e.g., Hochberg significance.
et al., 2006; Vansteensel et al., 2016; Pels et al., 2019).
BCI output commands can either select goals
Detecting and avoiding artifacts or control processes
Like conventional communication and control systems, In theory, BCI output commands fall into two classes:
BCIs have problems with artifacts that may obscure or commands that select goals or commands that control
contaminate the signals that provide output commands. the processes that achieve goals (Wolpaw and
BCI artifacts may come from the environment, e.g., Wolpaw, 2012).
electromagnetic noise from power lines or appliances; In a goal-selection BCI protocol, the user employs the
from the body, e.g., muscle (electromyographic BCI to send the goal (i.e., the user’s intent) to the soft-
(EMG)) activity, eye movement (electrooculographic ware in the application; the application then creates the
(EOG)) activity, cardiac (electrocardiographic (EKG)) process that realizes that intent. For example, the BCI
activity, body movements; or from the BCI hardware might send the goal of moving into the kitchen to face
(e.g., electrode/tissue interface instability, amplifier the refrigerator. The application (e.g., a BCI-controlled
noise) or software (e.g., shadows in rereferenced EEG wheelchair) then performs the actions (e.g., two-
topographies). dimensional translational movements, turning, braking)
Artifact recognition is particularly important for BCIs that together bring the wheelchair to the intended loca-
that record brain signals noninvasively (e.g., EEG, tion safely and efficiently. The wheelchair software uses
fNIRS). The first prerequisite for a BCI study is to make ongoing detailed feedback to continually adjust its
sure that it is an actual BCI study, i.e., that it derives its actions to avoid fixed or variable obstacles (e.g., walls
output commands from brain signals, not from other or people) and dangers (e.g., stairs). A goal-selection pro-
kinds of signals. Systems that use other biologic signals, tocol puts the burden of rapid complicated interactive
such as EMG, can be very useful, but they are not BCIs. control on the application. The user and the BCI merely
Nonbrain signals like EMG can easily masquerade as select and communicate the goal.
brain signals. Scalp electrodes can detect EMG from cra- In contrast, in a process-control protocol, the user and
nial muscles or EOG that equals or exceeds actual EEG in the BCI manage all the details of the process that
magnitude and overlaps it in frequency. People can easily achieves the user’s intent. Thus, for the wheelchair-to-
control cranial EMG or EOG; they may not even realize kitchen example described earlier, they produce contin-
that they are doing so. Such nonbrain activity can con- ual sequences of movement commands that the wheel-
taminate or even dominate the signals recorded by a chair merely executes (e.g., movements in x and y
BCI. The result can be that supposed BCI outputs are directions, turning, braking). The user employs the ongo-
actually produced in part, or even entirely, by nonbrain ing visual feedback to modify the BCI’s commands
signals. Serious and effective BCI research and develop- appropriately. The user and the BCI provide every step
ment is not possible in these circumstances. (In fact, the in the process that puts the user in front of the refrigerator;
scientific literature includes unfortunate instances of the wheelchair merely does what it is told to do. In short,
putative BCI studies in which the outputs reflect cranial in a goal-selection protocol, the BCI tells the application
muscle (i.e., EMG) control rather than brain signal (i.e., what to do; in a process-control protocol, the BCI tells the
EEG) control.) Often, commercial devices (e.g., for gam- application how to do it.
ing) sold as BCIs do not distinguish EEG from EMG or For the user and the BCI, goal selection is relatively
other nonbrain signals. If the results of BCI research easy. They only need to provide the goal (i.e., the user’s
studies are to be useful for people whose severe disabil- intent), which is the one feature of the desired action that
ities have eliminated their control of nonbrain signals, it the application itself cannot supply. Once provided with
is essential that the control signals that drive their outputs the goal, the application can achieve it quickly and con-
reflect brain activity. sistently. Goal-selection protocols are typically most
To avoid contamination by nonbrain signals, EEG- appropriate for applications with limited and fully
based BCI studies need to incorporate topographical defined sets of possible commands, such as word proces-
and frequency analyses that are comprehensive enough sing or movement in a fixed environment. For applica-
to differentiate between EEG and nonbrain signals; tions with many possible and perhaps not fully defined
fNIRS studies should incorporate analogous precautions. goals, for applications that may encounter unexpected
EEG studies that record from only one or two locations, complications (e.g., navigation in variable environments
or focus on a single narrow frequency band, cannot con- or multidimensional operation of a robotic arm), or for
fidently differentiate between EEG and EMG; thus, their users who want to regain the maximum possible control
 BRAIN-COMPUTER INTERFACES 21
(e.g., many people with severe disabilities), process con- neuromuscular disabilities, that is, the individuals who
trol protocols may be preferable, despite the fact that they constitute the primary rationale for the entire field and
make considerably greater demands on the user and the for the attention and resources that it is accorded. In short,
BCI. Such demands can be reduced by the use of shared it is critically important to develop and disseminate BCI
control approaches, where the output of the BCI is exe- systems that are clinically valuable.
cuted by the intelligent device (e.g., a wheelchair or robot Fulfilling this requirement depends on strong interdis-
arm) that exploits contextual information (e.g., obstacles ciplinary collaborations and effective management of the
or inferred goals) to compensate for any missing or mis- complex administrative and clinical demands of human
taken parameter in the BCI output (Carlson and Millán, research, including the distinctive ethical issues involved
2013; Millán, 2015). in BCI research and development. To be useful, BCI sys-
Normal (i.e., muscle-based) CNS outputs are the com- tems need to operate reliably and effectively in compli-
bined product of activity in many areas from cortex to cated and unpredictable settings. They need to support
spinal cord; and the distribution of control depends on useful applications and to be easily operated by nonex-
the action being produced. For example, the cortex has perts with only minimal technical support. The transla-
a much greater role in fine finger control (e.g., performing tion of laboratory BCI devices into robust systems that
surgery) than it does in gross finger movements (e.- satisfy these essential criteria is a lengthy and demanding
g., hand grasp) (Porter and Lemon, 1993). For some fin- endeavor; nevertheless, it is the goal and validation of the
ger movements, the cortex is heavily involved in entire BCI research and development process (e.g.,
controlling the process; for others it simply provides Birbaumer et al., 1999; Vansteensel et al., 2016; Saeedi
the goal, the details are delegated to subcortical regions et al., 2017; Perdikis et al., 2018; Wolpaw et al., 2018).
and spinal cord. Clinically validated BCIs still face challenges regard-
The best BCIs will probably be those that emulate the ing their dissemination to the people who need them. For
action-specific distribution of control typical of natural new medical technologies, dissemination is usually a
CNS outputs. That is, BCIs ought to combine goal selec- commercial endeavor; thus it needs to be profitable.
tion and process control in an action-appropriate manner. The number of people who could benefit from the limited
Thus, for reaching and grasping an object with a robotic communication and control capabilities of current BCIs
arm, the user and the BCI could control hand movement, is relatively small. Nevertheless, as BCIs become more
orientation, and grasp; the application software could capable and reliable, they may benefit many more people
control movement of individual limb segments and the with less severe disabilities. Another solution to the cur-
details of wrist rotation and finger flexion. Distributed rently small market for BCIs may lie in well-designed
designs of this kind, which impose lesser demands on initiatives aimed at both the limited group of people with
the user and the BCI, are also more realistic given the cur- severe disabilities and the many other people who might
rent state of BCI development. As BCIs improve, as they use BCIs for other, nonmedical purposes. At the same
incorporate more detailed and timely feedback from the time, the population who might benefit from the rapidly
evolving action to the user (e.g., perhaps cutaneous and developing rehabilitation applications of BCIs is quite
proprioceptive feedback as well as visual), goal selection large; it comprises the numerous patients with strokes,
and process control may be combined to enable BCIs to brain and spinal cord injuries, or other chronic neuromus-
imitate more closely the reliability, rapidity, and ease of cular disorders who need better therapies to enhance their
actions produced by natural CNS outputs. functional recovery. The complex practical issues related
to BCI dissemination and their potential solutions are
addressed in Chapters 24–26.
Creating and disseminating important BCI
applications
CONCLUSIONS
BCI research and development is necessarily multidisci-
plinary; it involves basic neuroscience, electrical, bio- Throughout life, the CNS interacts with the world and
medical, and materials engineering, applied math, with the body by exciting muscles and excreting
computer science, clinical neurology, neurosurgery, hormones. In contrast, BCIs measure CNS activity and
and physiatry, assistive communication and control, translate it into new artificial outputs that replace, restore,
and human factors expertise. In contrast, many research enhance, supplement, or improve the natural CNS out-
groups concentrate on a single aspect, such as electrode puts. Thus, BCIs modify the interactions between the
design or signal analysis. This unitary focus is logical, CNS and the environment. Unlike the natural CNS out-
and often essential for producing significant advances. puts that come from spinal and brainstem motoneurons,
Nevertheless, successful BCI development hinges on BCI outputs come from brain signals that reflect activity
creating systems useful to people with severe in other CNS areas, such as the sensorimotor cortex.
22 J.R. WOLPAW ET AL.
If BCIs are to be useful for important communication and REFERENCES
control purposes in real life, the CNS must control these
Allison BZ, McFarland DJ, Vaughan TM et al. (2008).
brain signals nearly as reliably and accurately as it con-
Towards an independent brain-computer interface using
trols spinal motoneurons. To do this, BCIs might, for steady state visual evoked potentials. Clin Neurophysiol
example, need to incorporate software that mimics the 119: 399–408.
function of the subcortical and spinal mechanisms that Birbaumer N, Ghanayim N, Hinterberger T et al. (1999).
contribute to normal movement control. The realization A spelling device for the paralyzed. Nature 398: 297–298.
of high reliability and accuracy may be the most difficult Bockbrader MA, Francisco G, Lee R et al. (2018). Brain
and critical challenge now facing BCI research and computer interfaces in rehabilitation medicine. Phys Med
development. Rehabil 10: S233–S243. [Link]
The continuing adaptive modifications that maintain 2018.05.028.
effective natural CNS outputs take place mainly in the Carlson TE, Millán JdR (2013). Brain-controlled wheelchairs:
a robotic architecture. IEEE Robot Autom Mag 20: 65–73.
CNS. The adaptive modifications that maintain effective
Carroll RC, Zukin RS (2002). NMDA-receptor trafficking and
BCI outputs can also take place in the BCI. Thus, the BCI
targeting: implications for synaptic transmission and plas-
operation depends on the effective collaboration of two ticity. Trends Neurosci 25: 571–577.
adaptive controllers, the CNS and the BCI. Creation of Choi I, Rhiu I, Lee Y et al. (2017). A systematic review of
this second adaptive controller, the BCI, and manage- hybrid brain-computer interfaces: taxonomy and usability
ment of its interactions with concurrent adaptations in perspectives. PLoS One 12: e0176674. [Link]
the CNS, constitute another complex and critical chal- 10.1371/[Link].0176674.
lenge for BCI development. Daly JJ, Wolpaw JR (2008). Brain-computer interfaces in neu-
BCIs can use different types of brain signals recorded rological rehabilitation. Lancet Neurol 7: 1032–1043.
in different ways from different brain areas. Choices Donoghue JP (2002). Connecting cortex to machines: recent
about which signals recorded in which ways from which advances in brain interfaces. Nat Neurosci 5: 1085–1088.
Gaiarsa JL, Caillard O, Ben-Ari Y (2002). Long-term plastic-
brain areas should be selected for which applications are
ity at GABAergic and glycinergic synapses: mechanisms
empirical questions that can only be adequately
and functional significance. Trends Neurosci 25:
answered by experiments. 564–570.
BCIs, like other communication and control technol- Gao Q, Dou L, Belkacem AN et al. (2017). Noninvasive
ogies, often encounter artifacts that contaminate or electroencephalogram based control of a robotic arm for
imitate their chosen signals. Noninvasive BCIs (e.g., writing task using hybrid BCI system. Biomed Res Int
EEG- or fNIRS-based) must take special care to avoid 2017: 8316485. 8 pages, [Link]
interpreting nonbrain signals (e.g., cranial EMG) as 8316485.
brain signals. This usually requires comprehensive Graimann B, Allison B, Pfurtscheller G (2010a).
topographical and spectral evaluations. Brain-computer interfaces: a gentle introduction. In:
In theory, BCI outputs can select a goal or control a B Graimann, B Allison, G Pfurtscheller (Eds.), Brain-
computer interfaces. Springer, Berlin, pp. 1–27.
process. In the future, the most effective BCIs will likely
Graimann B, Allison B, Pfurtscheller G (Eds.), (2010b). Brain-
be those that combine goal selection and process control
computer interfaces. Springer, Berlin p. 21. et passim.
so as to distribute control between the BCI and the appli- Guertin PA, Steuer I (2009). Key central pattern generators of
cation in a fashion suited to the current action. Through the spinal cord. J Neurosci Res 87: 2399–2405.
this distribution, BCIs may most effectively imitate nat- Hochberg LR, Serruya MD, Friehs GM et al. (2006). Neuronal
ural CNS operation. ensemble control of prosthetic devices by a human with
The primary measure of BCI development is the tetraplegia. Nature 442: 164–171.
extent to which BCI systems benefit people with neuro- Ijspeert AJ (2008). Central pattern generators for locomotion
muscular disorders. Thus, BCI clinical evaluation, control in animals and robots: a review. Neural Netw 21:
validation, and dissemination is an essential step. It is 642–653.
also a complex and difficult process that requires multi- Janssen P, Scherberger H (2015). Visual guidance in control of
grasping. Annu Rev Neurosci 38: 69–86.
disciplinary collaboration and effective management of
K€ubler A, M€ uller KR (2007). An introduction to brain-computer
the demanding requirements of clinical studies.
interfacing. In: G Dornbhege, JdR Millan, T Hinterberger,
Twenty-five years ago, BCI research was an esoteric DJ McFarland, K-R M€ uller (Eds.), Toward brain-computer
endeavor that engaged a few isolated laboratories. It is interfacing. MIT Press, Cambridge, MA, pp. 1–26.
now a burgeoning field engaging many hundreds of Lazarou I, Nikolopoulos S, Petrantonakis PC et al. (2018). EEG-
scientists, engineers, and clinicians throughout the world based brain–computer interfaces for communication and
in an increasingly interconnected community addressing rehabilitation of people with motor impairment: a novel
the key issues and pursuing the high potential of BCI approach of the 21st century. Front Hum Neurosci 12: 14.
technology. [Link]
 BRAIN-COMPUTER INTERFACES 23
Ma T, Li H, Deng L et al. (2017). The hybrid BCI system for central nervous system by using neurotrophins.
movement control by combining motor imagery and mov- Neurorehabil Neural Repair 19: 283–295.
ing onset visual evoked potential. J Neural Eng 14: 026015 Vidal JJ (1973). Towards direct brain–computer communica-
(12 pp). [Link] tion. Annu Rev Biophys Bioeng 2: 157–180.
McCrea DA, Ryback IA (2008). Organization of mammalian Vidal JJ (1977). Real-time detection of brain events in EEG.
locomotor rhythm and pattern generation. Brain Res Rev Proc IEEE 633–664: 65 [Special issue on Biological
57: 134–146. Signal Processing and Analysis].
McFarland DJ, Krusienski DJ (2012). BCI signal processing: Wolpaw JR (2002). Memory in neuroscience: rhetoric versus
feature extraction. In: JR Wolpaw, EW Wolpaw (Eds.), reality. Behav Cogn Neurosci Rev 1: 130–163.
Brain-computer interfaces: principles and practice. Wolpaw JR (2010). What can the spinal cord teach us about
Oxford University Press123–146. learning and memory? Neuroscientist 16: 532–549.
McFarland DJ, Sarnacki WA, Wolpaw JR (2010). Wolpaw JR (2018). The negotiated equilibrium model of spi-
Electroencephalographic (EEG) control of three- nal cord function. J Physiol 596: 3469–3491.
dimensional movement. J Neural Eng 7: 036007. Wolpaw JR, Wolpaw EW (2012). Brain-computer interfaces:
Millán JdR (2015). Brain-machine interfaces: the perception- something new under the sun. In: JR Wolpaw, EW Wolpaw
action closed loop. IEEE Syst Man Cybern Mag 1: 6–8. (Eds.), Brain-computer interfaces: principles and practice.
Millán JdR, Rupp R, M€uller-Putz G et al. (2010). Combining Oxford University Press, pp. 3–12.
brain-computer interfaces and assistive technologies: state- Wolpaw JR, Birbaumer N, McFarland DJ et al. (2002). Brain-
of-the-art and challenges. Front Neurosci 4: 161. computer interfaces for communication and control. Clin
Ming B-K, Chavarriaga R, Millán JdR (2017). Harnessing pre- Neurophysiol 113: 767–791.
frontal cognitive signals for brain-machine interfaces. Wolpaw JR, Bedlack RS, Reda DJ et al. (2018). Independent
Trends Biotechnol 35: 585–597. home use of a brain-computer interface by people with
M€uller-Putz GR, Leeb R, Tangermann M et al. (2015). amyotrophic lateral sclerosis. Neurology 91: e258–e267.
Towards non-invasive hybrid brain-computer interfaces: [Link]
framework, practice, clinical application and beyond. Young RM (1990). Mind, brain and adaptation in the nine-
Proc IEEE 103: 926–943. teenth century, Oxford University Press.
Pels EGM et al. (2019). Stability of a chronic implanted brain- Zehr EP, Hundza SR, Vasudevan EV (2009). The quadrupedal
computer interface in late-stage ALS. Clin Neurophysiol nature of human bipedal locomotion. Exerc Sport Sci Rev
130: 1798–1803. 37: 102–108.
Perdikis S, Tonin L, Saeedi S et al. (2018). The cybathlon BCI
race: successful longitudinal mutual learning with two tet- FURTHER READING
raplegic users. PLoS Biol 16: e2003787.
Porter R, Lemon R (1993). Corticospinal function and volun- Fetz EE (1969). Operant conditioning of cortical unit activity.
tary movement, Clarendon Press, Oxford. Science 163: 955–958.
Saeedi S, Chavarriaga R, Millán JdR (2017). Long-term stable Fetz EE, Finocchio DV (1971). Operant conditioning of spe-
control of motor-imagery BCI by a locked-in user through cific patterns of neural and muscular activity. Science
adaptive assistance. IEEE Trans Neural Syst Rehabil Eng 174: 431–435.
25: 380–391. Joseph AB (1985). Design considerations for the brain-
Saneyoshi T, Fortin DA, Soderling TR (2010). Regulation of machine interface. Med Hypotheses 17: 191–195.
spine and synapse formation by activity-dependent intracel- Vaughan TM, Wolpaw JR (2006). The third international
lular signaling pathways. Curr Opin Neurobiol 20: 108–115. meeting on brain-computer interface technology: making
Schwartz AB (2004). Cortical neural prosthetics. Annu Rev a difference (Editorial). IEEE Trans Neural Syst Rehabil
Neurosci 27: 487–507. Eng 14: 126–127.
Sutter EE (1992). The brain response interface: communica- Wolpaw JR (2007). Brain-computer interfaces as new brain
tion through visually induced electrical brain responses. output pathways. J Physiol 579: 613–619.
J Microcomput Appl 15: 31–45. Wolpaw JR, Birbaumer N (2006). Brain-computer interfaces
Vansteensel MJ, Pels EGM, Bleichner MG et al. (2016). Fully for communication and control. In: ME SelzerS Clarke,
implanted brain-computer interface in a locked-in patient LG Cohen, P Duncan, FH Gage (Eds.), Textbook of
with ALS. New Engl J Med 375: 2060–2066. neural repair and rehabilitation: neural repair and
Vaynman S, Gomez-Pinilla F (2005). License to run: exercise plasticity. Cambridge University Press, Cambridge,
impacts functional plasticity in the intact and injured pp. 109–125.