Добавил:
kiopkiopkiop18@yandex.ru t.me/Prokururor I Вовсе не секретарь, но почту проверяю Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз: Предмет: Файл:
Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_6027_Библиотеки_им_академика_М_И_Перельмана.pdf
Скачиваний:
0
Добавлен:
31.08.2026
Размер:
29 Мб
Скачать
25 From Brain Signals to Neuroadaptive Technology: BCIs for Human-Computer... 337
signicant challenges if they co-occur systematically with labels, especially in mobile or applied settings. Techniques such as Independent Component Analysis
(ICA) (Bell & Sejnowski,
struction (ASR) (Mullen et al., 2015; Chang et al., 2019) can help to suppress
Recon
these
sources but come at the cost of not necessarily being real-time capable, or risking removing brain activity from the data (see also Chap. situati
ons, however, preprocessing can only do so much, and good data begins at the source. Proper EEG setup and data quality monitoring during acquisition are vital to ensure the highest possible data quality for any downstream analysis.
Comparing BCIs with other forms of neuroimaging, one point needs to be emphasized: In real-time BCIs, preprocessing must operate with minimal latency. By this nature, BCI preprocessing, too, is minimal and favors lightweight, causal lters, and reduced-complexity artifact handling.
1995; Hyvärinen et al., 2001) or Artifact Subspace
15). As in many

25.3.4 Feature Extraction

BCI performance depends heavily on how neural activity is represented and which mental states are to be decoded, or elicited in the paradigm. As the neural signals in the EEG were categorized into event-related and oscillatory components, this dis­tinction is relevant for BCI feature extraction, as well. Time-domain features, such as ERP windowed mean amplitudes, capture transient components tied to specic events (e.g., error responses). Frequency-domain features are used to capture oscil­latory activity in frequency bands associated with continuous cognitive processes (e.g., workload). Spatial ltering methods, including Common Spatial Patterns (CSPs), enhance class separability in such frequency features by emphasizing spatial differences in activation patterns before extracting the band powers themselves.
A more recent family of representations uses Riemannian geometry, modeling covari
ance matrices on their natural manifold to capture stable spatial relationships
in EEG (Congedo et al., 2017; Yger et al., 2017). These approaches have shown
performance in both event-based and continuous decoding tasks, but require
strong additional processing steps.
25.3.5 Machine Learning and Classiers
BCIs employ a range of machine learning methods to translate features into pre­dictions. Linear Discriminant Analysis (LDA) and logistic regression remain widely used because they provide strong performance with minimal computational cost and work wel l with limited training data (Lotte et al.,
(SVMs)
linear SVMs are more commonly used. They may exhibit slightly improved perfor­mance but require more tuning and computational overhead. Importantly, in BCI,
can handle nonlinear boundaries, but this is often not necessary in BCI, and
2018). Support Vector Machines
338 M. Klug et al.
there is often a lack of data points in relation to the number of featuresor parametersthe classier is taking in. For example, in windowed-means approaches, the number of features can exceed one thousand, but the number of samples may be only one hundred per class. Hence, it is particularly important to take care to regularize the data accordingly, independently of the machine learning method employed.
Recently, new approaches employing convolutional neural nets, such as EEGNet (Lawhern et al., 2018), have emerged. These approaches combine the extraction of features and the separation of the classes in one algorithm to learn temporal or spatial structure directly from raw or minimally processed EEG, thus leveraging the power of both data-driven feature extraction and non-linear decoding (Schirrmeister et al.,
2017; Roy et al., 2019). While these methods are newer and initially showed
ise, deep learning approaches require more substantial computational
prom resources, are prone to overtting due to their size, and they do not show relevant improvements over classic approaches (Lotte et al., data,
these methods might become more powerful, however.
2018). With larger quantities of
25.3.6 Classier Validation
Validation assesses how well a trained model can decode neural states, and it must go beyond reporting the accuracy on the training data, which is inated because the model has already seen those examples. Cross-validation addresses this by partitioning data into train–test folds and estimating performance on unseen data, ideally preserving temporal structure to avoid leakage. A completely independent test set provides the most reliable measure of generalization, however. Importantly, it is not enough to assume that a decoding accuracy above chance level is sufcient. Instead, the statistical signicance level of exceeding chance level by chance needs to be computed for the given number of samples per class, and reported as a baseline that the decoding must exceed (Combrisson & Jerbi, et
al., 2008). Here, data imbalance can bias both models and accuracy estimates, and
gh some algorithms, such as LDA, tolerate imbalance reasonably well during
althou training, it is preferable to balance conditions for validation to obtain a proper estimate of the classiers performance.
Beyond numerical metrics, validation must include inspection of the neural
source
s driving classication to ensure that decoding reects genuine brain signals
rather than artifacts (Haufe et al., requi
re assessing generalizability to new tasks or contexts, since calibration is often
performed in one setting but deployment may occur in others (e.g., Zhang et al.,
2018).
2014; Krol et al., 2018a). Finally, especially pBCIs
2015; Mueller-Putz
25 From Brain Signals to Neuroadaptive Technology: BCIs for Human-Computer... 339

25.4 BCI Types Illustrated

25.4.1 Motor Imagery as an Active BCI Paradigm

Motor imagery (MI) BCIs convert imagined movements into commands by exploiting modulations of sensorimotor rhythms. When users imagine left- or right-hand actions, mu (8–13 Hz) and beta (13–30 Hz) power over the sensorimotor cortex shows reliable event-related desynchronization patterns, lateralized to the imagined limb (Pfurtscheller & Lopes da Silva, 1999; Pfurtscheller et al., 2006). A
ard MI pipeline uses band-pass ltering around the mu/beta range, spatial
stand ltering such as Common Spatial Patterns to enhance class separability (Ramoser et al., 2000), and a linear classier such as LDA. MI systems are a canonical example of volitional neural control and are widely used to study intentional BCI communi­cation, neurofeedback, and motor rehabilitation (Daly & Wolpaw, 2008; Blankertz et
al., 2008).

25.4.2 P300 and SSVEPs as Reactive BCI Paradigms

Reactive BCIs leverage stimulus-evoked activity that is modulated by attention. In the P300 speller, a matrix of letters is presented, where rows and columns ash at random. Flashes of the attended letter elicit a positive ERP around 300 ms poststimulus. By classifying which stimuli produce strong P300 responses, the system infers the intended selection, and over time, words and sentences can be spelled (Farwell & Donchin, 1988). Decoding typically uses ERP amplitude features with linear classiers. Steady-state visually evoked potential (SSVEP) BCIs instead use different continuous icker frequencies on visible objects, where xations on such objects induce frequency-tagged oscillations in the visual cortex detectable over occipital channels (Vialatte et al., the
EEG can then be used to identify the dominant icker frequency to determine the chosen item. Both paradigms provide high reliability with little user training, making them central for communication BCIs.
2010; Norcia et al., 2015). Spectral power in

25.4.3 Workload, Error, and Other Passive BCI Paradigms

Passive BCIs decode naturally occurring brain activity to estimate cognitive or affective state in real time, without requiring intentional control. Typical targets include workload, vigilance, attention, fatigue, and error perception, operationalized through task manipulations rather than explicit commands (Zander & Kothe, 2011; Arico
et al., 2018). Continuous states are commonly inferred from oscillatory
ers such as frontal theta or parietal alpha changes for cognitive workload,
mark
340 M. Klug et al.
whereas event-based states rely on ERPs at the level of single trials linked to feedback, surprise, or errors. Processing pipelines mirror those of other BCIs, with continuous states being decoded similarly to MI BCIs, and event-based transient states being decoded similarly to P300 BCIs. These paradigms are not used for communication or control, and do not require conscious attentional shifts or active imagination on the users side. Instead, they ground neuroadapt
ive systems or articial intelligence applications that can adjust to users based on moment-to­moment state detection.
In the literature, workload is one of the most extensively studied targets for passive BCIs. Here, changes in frontal theta and parietal alpha are commonly used to track cognitive demand in working-memory paradigms (Gevins & Smith, 2003; Gerjets have rema
et al., 2014; Brouwer et al., 2012). Carefully designed training paradigms
also been shown to be sensitive across tasks (Zhang et al., 2018), while
ining robust across posture and stimulation modality (at screen/VR) and using the same frontal theta/parietal alpha markers as previous studies (Gherman et al.,
2025).
Beyond workload, error and surprise responses provide powerful event-based
targe
ts. Error-r elated negativities and related components can signal when the user detects a mismatch or mistake, whether caused by themselves or by the system (Falkenstein et al., 2000). This has been exploited in passive BCIs that detect interacti
on errors and trigger automatic undoor corrective actions (Chavarriaga et al., 2014; Lopes-Dias et al., 2019; Xavier Fidêncio et al. 2022). Prediction-error respon
ses have also been used to detect unrealistic visuo-haptic interactions and implausible object physics in VR, enabling systems to identify when a simulation violates user expectations and adapt automatically to user preference (Gehrke et al.
2019, 2022, 2025).
Beyond workload and error, passive BCIs have been used to track a range of other mental states. Attention and task engagement are often monitored using spectral indices similar to workload, for example, frontal theta and parietal alpha ratios, or composite engagement indices(Pope et al., meas
ures have been applied to vigilance and drowsiness monitoring and safety-
1995; Berka et al., 2007). These
critical supervision, sometimes complemented by event-related components such as the P300, which is modulated by stimulus relevance, motivation, and attentional resources (Begleiter et al.,
1983; Polich, 2007; Acı et al., 2019; Vortmann et al., 2022; Pawlitzki et al., 2021). A closely related line of work targets relaxation, often
operationalized through increased alpha power or related spectral changes, for example, in neurofeedback or stress-reduction settings (Ewing et al., 2016; Krol et
al., 2017; Klug, 2022). Another distinct target is the intent to interact: passive
BCIs
can distinguish spontaneous from goal-directed xations by analyzing xation-related potentials, particularly when combined with an eye tracker. Studies have shown that EEG during target xations carries markers of intentional selection (Protzak et al.,
incidental xations and even smooth pursuit using EEG features (Shishkin
from et al.,
2016; Zhao et al., 2021). Together, these paradigms demonstrate that passive
BCIs
can access a broad spectrum of cognitive, affective, and intentional states
2013), and that gaze xations used for interaction can be separated
beyond workload and error.
25 From Brain Signals to Neuroadaptive Technology: BCIs for Human-Computer... 341
25.5 From Mental State Assessment to Neuroadaptive
Systems

25.5.1 Mental State Assessment as a First Stage

The most basic use of passive BCIs is mental state assessment (MSA). Here, brain signals are recorded and decoded to estimate states such as workload or vigilance, but these estimates are used mainly for monitoring or later analysis, not for direct control. A fatigue-m onitoring app that records your state while working and lets you review it afterwards is a typical example. Strictly speaking, this does not meet the classical BCI denition, which requires both neural input and a system-level output (Wolpaw & Wolpaw,
time
brain state decoding (RBSD) as an extension of cognitive monitoring (Zander &
Kothe, 2011), in line with neuroergonomics, which studies cognition at work and in everyda only interactive systems and is one of the most mature real-world uses of passive BCIs (Arico et al., 2018; Krol et al., 2018b).
y life (Parasuraman & Rizzo, 2007; Ayaz & Dehais, 2018). Although MSA
uses the front halfof the BCI loop, it relies on the same decoding methods as
2012). The original passive BCI paper instead framed real-

25.5.2 Open- and Closed-Loop Adaptation

The next step is to let systems respond to decoded states. Open-loop adaptation refers to one-off actions that do not directly change the triggering state, for example, a high-workload estimate could be used to trigger a notication to take a break or highlight a problematic interface element. Closed-loop adaptation changes task parameters to inuence the state itself, such as increasing automation or suppressing non-essential information when workload rises in a driving or piloting task (Kohlmorgen et al.,
cases, passive BCIs provide implicit input, but in closed-loop systems, the
both adaptation becomes part of the mechanism that shapes the users mental state.
2007; Zander & Jatzev, 2012; Andreessen et al., 2021). In
25.5.3 Autonomous Adaptation, User Models, and Cognitive
Probing
The most powerful use of passive BCIs is to monitor mental states over longer periods and relate them to contextual factors to build user models and enable automated adaptation that anticipates problematic states (Zander & Jatzev, 2012; Fairclo
ugh, 2017; Klaproth et al., 2020). Because decoding alone cannot specify why
state occurred, systems can actively manipulate their own settings to elicit diag-
a nostic responses, an approach termed cognitive probing (Krol et al.,
2020).
Over
342 M. Klug et al.
time, such probes rene the user model and suppor t autonomous behavior that increasingly approaches the goal state of the user (Zander et al., adapta
tions rely on internal states not explicitly offer ed as input, they must satisfy the user-interest criterion: actions should be justiable as being in the users best interest and remain transparent and overridable (Krol & Zander, 2022).
2016). Since these

25.6 Practical Challenges

25.6.1 Mobility, Artifacts, and Non-stationarity

Real-world BCIs must operate under movement, changing environments, and vary­ing user states. Mobile EEG introduces additional artifacts from motion, muscle activity, and shifting electrode contact that may result in suboptimal data quality (Gramann et al.,
is approaches, such as ICA, can help reduce this issue for MSA, real-time
analys capable artifact attenuation approac hes, such as ASR (Mullen et al., requi
red for fully real-time applications in the wild. Even with good preprocessing, the statistical properties of EEG change over time due to factors such as fatigue, learning, electrode drift, or changing posture, a phenomenon commonly referred to as non-stationarity (Blankertz et al., 2008; Samek et al., 2014). Classiers trained on short,
clean calibration sessions can degrade at long-term use or in different contexts. Robust pBCIs therefore require not only better artifact handling but also models that can track slow drifts and rapid changes in data distributions, for example, through adaptive lters, online covariance updates, or periodic recalibration (Lotte et al.,
2018).
2011; Klug & Gramann, 2021; Klug et al., 2024). While ofine
2015), may be

25.6.2 Cross-User and Cross-Session Generalization

Most current BCIs are still subject-specic: they require a dedicated calibration session for each user and often for each new session or task. This calibration burden is problematic for applied and consumer scenarios. Cross-session and cross-user generalization are difcult because of individual differences in anatomy, cognitive strategy, and noise proles (Lotte et al., domai
n adaptation addresses this by aligning feature spaces across users or sessions, re-weighting samples, or learning shared representations that are less sensitive to distributional changes (Jayaram et al.,
ing, and few-shot adaptation are promising directions (Yger et al., 2017), but
learn robust
zero-calibration BCIs rema in rare in practice. For passive BCIs in particular, where the goal is to deploy workload or error decoders across tasks and contexts, improving generalization is one of the central open challenges.
2018). Research on transfer learning and
2016). Riemannian alignment, multitask
25 From Brain Signals to Neuroadaptive Technology: BCIs for Human-Computer... 343

25.6.3 Evaluatio n in Real Settings

BCI performance is often reported as ofine accuracy on well-controlled datasets, but real-world evaluation requires broader metrics. In applied scenarios, task-rele- vant outcomes such as driving performance, error rates, learning gains, or user workload reduction are more informative than raw classication accuracy. Longitu­dinal and eld studies are needed to assess robustness over days or weeks, as well as user acceptance and trust. For pBCIs, cross-task application is particularly critical: decoders trained in one paradigm are often applied to different tasks or environ­ments, where signal characteristics and label structure change. Similarly, applying event-based classiers (e.g., error reaction decoders) continuously to live-streamed data is nontrivial, as the live data may not have event markers at the same temporal accuracy level as the training data, or no markers at all (Pan et al., framewo rates, and the impact of neuroadaptive inte rventions on actual user behavior.
rks must therefore account for streaming performance, latency, false-alarm
2024). Evaluation

25.6.4 Ethics, Privacy, and Neurorights

BCIs and pBCIs raise distinctive ethical questions because they access signals that are closely linked to perception, intention, and affect. Even simple workload or error decoders can reveal information that users might prefer to keep private, such as lapses of attention, fatigue, or preference patterns (Krol & Zander, 2022; Ienca & Ando
rno, 2017). Responsible design requires transparency about which signals are
ed, how long they are stored, who can access them, and how adaptations are
record triggered. Informed consent should cover not just data acqu isition but also the possible inferences and system acti ons. Recent discussions of neurorights have proposed protections for cognitive liberty, mental privacy, and mental integrity in response to emerging neurotechnology (Ligthart et al., 2023; Yuste et al., 2017). For neuroad user-interest criterion is essential to maintain trust and ensure that powerful decoding methods are used in ways that genuinely benet users (Krol & Zander,
aptive systems based on passive BCIs, respecting these principles and the
2022).

25.7 Conclusions and Outlook

25.7.1 Key Takeaways

This chapter has outlined how BCIs transform neural activity into meaningful system behavior. We distinguished active, reactive, and passive BCIs based on whether signals are intentionally produced, stimulus-locked, or arise spontaneously during ongoing cognition (Zander & Kothe, 2011). We then described the core
344 M. Klug et al.
pipeline from experimental design and signal acquisition through preprocessing, feature extraction, machine learning, and validation (Lotte et al., and
oscillations providing the main signal families exploited for decoding.
2018), with ERPs
Within this landscape, passi ve BCIs occupy a special role. They do not replace explicit input but add an implicit channel that tracks workload, attention, error perception, surprise, or intent to interact in real time. We reviewed paradigms and markers for these states and showed how they can be combined with context information to support neuroadaptive systems. Finally, we discussed a spectrum of interactivityfrom mental state assessment and open-loop notications to closed­loop and autonomous adaptations driven by user models and cognitive probing underpinned by the user-interest criterion as a guardrail for responsible design (Krol et al.,
2018b; Krol & Zander, 2022).

25.7.2 Future Trajectories

Several technical and conceptual developments are likely to shape the next gener­ation of BCIs and neuroadaptive systems. On the hardware side, progress in wear­able EEG, ear-EEG, and VR-integrated devices will improve comfort and ecological validity, making long-term and mobile applications more realistic. On the algorith­mic side, cross-session, cross-subject, and cross-task classier generalization are the most relevant frontiers at the moment. Both hardware and software improvements are building towards a more user-friendly application of pBCIs, allowing their widespread use.
Perhaps most intriguingly, passive BCIs can provide a rich feedback channel for adapti
ve AI systems. Real-time estimates of workload, uncertainty, surprise, prefer­ence, or agreement could serve as implicit rewards or constraints for interactive learning, helping AI systems to align their behavior with human goals and limits more directly than behavioral logs alone. In such scenarios, pBCIs would not read thoughtsto replace speech but offer continuous, noisy measurements of how well an AI is supporting the user in context. Combining these capabilities with neurorights-informed safeguards for mental privacy and agency may enable articial assistants that are not only more capable, but also more genuinely understanding and aligned with the people they serve.

References

Acı, Ç.İ., Kaya, M., & Mishchenko, Y. (2019). Distinguishing mental attention states of humans via
an EEG-based passive BCI using machine learning methods. Expert Systems with Applications, 134, 153–166. Available at:
9303926
https://www.sciencedirect.com/science/article/abs/pii/S095741741
25 From Brain Signals to Neuroadaptive Technology: BCIs for Human-Computer...
Andreessen, L. M., et al. (2021). Toward neuroadaptive support technologies improving digital
reading: Passive BCI-based assessment mental workload imposed text difculty presentation speed during reading. User Modeling and User-Adapted Interaction, 31, 75– 104.
Arico, P., et al. (2018). Passive BCI beyond the lab: Current trends and future directions. Physio-
logical
Ayaz, H., & Dehais, F. (Eds.). (2018). NeuroergonomicsThe brain at work and in everyday life
Begleiter, H., et al. (1983). P3 and stimulus incentive value. Psychophysiology, 20(1), 95–101.
Bell, A. J., & Sejnowski, T. J. (1995). An information-maximization approach to blind separation
Berka, C., et al. (2007). EEG correlates of task engagement and mental workload in vigilance,
Biasiucci, A., Franceschiello, B., & Murray, M. M. (2019). Electroencephalography. Cell. https://
Blankertz, B., et al. (2008). The Berlin brain-computer interface. In Lecture notes in computer
Brouwer, A.-M., et al. (2012). Estimating workload using EEG spectral power and ERPs in the
Chang, C.-Y., et al. (2019). Evaluation of artifact subspace reconstruction for automatic artifact
Chavarriaga, R., Sobolewski, A., & Millán, J. D. R. (2014). Errare machinale est: The use of error-
Combrisson, E., & Jerbi, K. (2015). Exceeding chance level by chance:
Congedo, M., Barachant, A., & Bhatia, R. (2017). Riemannian geometry for EEG-based brain-
Daly, J. J., & Wolpaw, J. R. (2008). Brain-computer interfaces in neurological rehabilitation. Lancet
Debener, S., et al. (2015). Unobtrusive ambulatory EEG using a smartphone and exible printed
Donoghue, T., & Voytek, B. (2022). Automated
Ewing, K. C., Fairclough, S. H., & Gilleade, K. (2016). Evaluation of an adaptive game that uses
Fairclough, S. H. (2017. Physiological computing and intelligent adaptation. In Emotions and affect
Falkenstein, M., et al. (2000). ERP components on reaction errors and their functional signicance:
Farwell, L. A., & Donchin, E. (1988). Talking off the top of your head: Toward a mental prosthesis
Measurement, 39(8). https://doi.org/10.1088/1361-6579/aad57e
(1st ed.). Elsevier Academic. https://doi.org/10.1016/C2016-0-02196-4
https://doi.org/10.1111/j.1469-8986.1983.tb00909.x
blind deconvolution. Neural Computation, 7(6), 1129–1159. https://doi.org/10.1162/neco.
and
1995.7.6.1129
learning, B244.
doi.org/10.1016/j.cub.2018.11.052
science
n-back
i crossref.27952d535478677b655a40d95df34b6d
components Engineering, 67(4), 1114–1121. https://doi.org/10.1109/TBME.2019.2930186
related
frontiersin.org/article/10.3389/fnins.2014.00208/abstract
chance levels in brain signal classication and statistical assessment of decoding accuracy. Journal of Neuroscience Methods, 250, 126–136.
01.010
computer
doi.org/10.1080/2326263X.2017.1297192
Neurology,
electrodes
literature. Scientic Reports, 12(1), 1867. https://doi.org/10.1038/s41598-022-05939-9
EEG
in Human Neuroscience, 10, 223. https://doi.org/10.3389/fnhum.2016.00223
in human factors and human-computer interaction (pp. 539–556). Elsevier. https://doi.org/10.
1016/b978-0-12-801851-4.00020-3
A
61/
utilizing 70(6), 510–523.
and memory tasks. Aviation, Space, and Environmental Medicine, 78(5 Suppl), B231–
https://www.ncbi.nlm.nih.gov/pubmed/17547324
(pp. 79–101). Springer. https://doi.org/10.1007/978-3-540-68860-0_4
task. Journal of Neural Engineering, 9(4), 045008. http://stacks.iop.org/1741-2552/9/
¼4/a¼045008?key¼
removal in multi-channel EEG recordings. IEEE Transactions on Bio-Medical
potentials in brain-machine interfaces. Frontiers in Neuroscience, 8, 208.
The caveat of theoretical
https://doi.org/10.1016/j.jneumeth.2015.
interfaces; a primer and a review. Brain-Computer Interfaces, 4(3), 155–174. https://
7(11), 1032–1043. https://doi.org/10.1016/S1474-4422(08)70223-0
around the ear. Scientic Reports, 5, 16743. https://doi.org/10.1038/srep16743
meta-analysis of the event-related potential (ERP)
measures validated during the design process as inputs to a biocybernetic loop. Frontiers
tutorial. Biological Psychology, 51(2–3), 87–107. https://pubmed.ncbi.nlm.nih.gov/106863
event-related brain potentials. Electroencephalography and Clinical Neurophysiology,
https://doi.org/10.1016/0013-4694(88)90149-6
http://journal.
345
346 M. Klug et al.
Fischer, G. (2001). User modeling in human– computer interaction. User Modeling and User-
Adapted
Gehrke, L., et al. (2019). Detecting visuo-haptic mismatches in virtual reality
error negativity of event-related brain potentials. In Proceedings of the 2019 CHI conference on human factors in computing systems. CHI 19 (pp. 111). Association for Computing Machin- ery.
Gehrke, L., et al. (2022). Neural sources of prediction errors detect unrealistic VR interactions.
Journal
Gehrke, L., et al. (2025). Neuroadaptive haptics: A proof-of-concept comparing reinforcement
learning from explicit ratings and neural signals for adaptive XR systems. Frontiers in Virtual Reality, 6, 1616442.
Gerjets, P., et al. (2014). Cognitive state monitoring and the design of adaptive instruction in digital
environments: computer interface approach. Frontiers in Neuroscience, 8(DEC), 1–21.
3389/fnins.2014.00385
Gevins, A., & Smith, M. E. (2003). Neurophysiological measures of cognitive workload during
human
https://doi.org/10.1080/14639220210159717
Gherman, D. E., et al. (2025). An investigation of a passive BCIs performance for different body
postures
doi.org/10.1088/2057-1976/adb58b
Gramann, K., et al. (2011). Cognition in action: Imaging brain/body dynamics in mobile humans.
Reviews
Haufe, S., et al. (2014). On the interpretation of weight vectors of linear models in multivariate
neuroimaging.
Hyvärinen, A., Karhunen, J., & Oja, E. (2001). Independent component analysis. Wiley. https://doi.
org/10.3156/jsoft.16.150_1
Ienca, M., & Andorno, R. (2017). Towards new human rights in the age of neuroscience and
neurotechnology.
017-0050-1
Jayaram, V., et al. (2016). Transfer learning in brain-computer interfaces. IEEE Computational
Intelligence
Klaproth, O. W., et al. (2020). A neuroadaptive cognitive model for dealing with uncertainty in
tracing
onlinelibrary.wiley.com/doi/10.1111/tops.12515
Klimesch, W. (1999). EEG alpha and theta oscillations reect cognitive and memory performance:
A
org/10.1016/s0165-0173(98)00056-3
Klug, M. (2022). Real Virtual Magic–Modifying a VR game with a BCI to enhance immersion. In
Neuroadaptive
Klug, M., & Gramann, K. (2021). Identifying key factors for improving ICA-based decomposition
of 54(12), 8406–8420. https://doi.org/10.1111/ejn.14992
Klug, M., Berg, T., & Gramann, K. (2024). Optimizing EEG ICA decomposition with data cleaning
in stationary and mobile experiments. Scientic Reports, 14(1), 14119. https://doi.org/10.1038/
s41598-024-64919-3
Kohlmorgen, J., et al. (2007). Improving human performance in a real operating environment
through computer interfacing (pp. 409–422). MIT Press. https://play.google.com/store/books/details?
id V88swGX83ecC
Krigolson, O. E., et al. (2017). Choosing MUSE: Validation of a low-cost, portable EEG system for
ERP
Interaction, 11(1/2), 65–86.
https://doi.org/10.1145/3290605.3300657
of Neural Engineering, 19(3), 036002.
https://doi.org/10.3389/frvir.2025.1616442
Lessons learned from cognitive workload assessment using a passive brain-
–computer interaction. Theoretical Issues in Ergonomics Science, 4(1–2), 113–131.
and presentation modalities. Biomedical Physics & Engineering Express, 11(2). https://
in the Neurosciences, 22(6), 593–608. https://doi.org/10.1515/RNS.2011.047
NeuroImage, 87, 96–110.
Life Sciences, Society and Policy, 13(1), 5.
Magazine, 11(1), 20–31.
pilotscognitive state. Topics in Cognitive Science, 12(3), 1012–1029. https://
review and analysis. Brain Research. Brain Research Reviews, 29(2–3), 169–195. https://doi.
technologies 2022 conference proceedings.
EEG data in mobile and stationary experiments. The European Journal of Neuroscience,
real-time mental workload detection. In G. Dornhege, et al. (Eds.), Toward brain-
¼
research. Frontiers in Neuroscience, 11, 109. https://doi.org/10.3389/fnins.2017.00109
https://doi.org/10.1023/a:1011145532042
using the prediction
https://doi.org/10.1088/1741-2552/ac69bc
https://doi.org/10.
https://doi.org/10.1016/j.neuroimage.2013.10.067
https://doi.org/10.1186/s40504-
https://doi.org/10.1109/MCI.2015.2501545