Kapitel

• FrontmatterMüller-Putz, Gernot; Kostoglou, Kyriaki; Oberndorfer, Markus E.; Wriessnegger, Selina C.; 10.3217/978-3-99161-093-9-000
• Regulatory challenges of certification of implantable BCI medical devices in the EU and how to overcome themMachado, Ana Matos; 10.3217/978-3-99161-093-9-001Brain Computer Interface (BCI) technology has seen substantial growth over the past decade with increasing private investment into its research and development. With ongoing promising clinical studies, market entry of implantable devices is not far in the future and so regulatory discussions now start to become relevant. All medical devices in the EU fall under the scope of the Medical Device regulation 2017/745 (MDR) and potentially the AI Act 2024/1689. So how will certification work for this technology? What will be the expected challenges considering the current regulatory framework? With focus on the clinical assessment part of certification, this article will try to navigate these questions and provide clarity to the process.
• Sensorimotor EEG Activity During Brain-Computer Interface Therapy Predicts Motor Function in Parkinson’s DiseaseSieghartsleitner, Sebastian; Cho, Woosang; Sebastián-Romagosa, Marc; Ortner, Rupert; Schwarzgruber, Michael; Scharinger, Josef; Guger, Christoph; 10.3217/978-3-99161-093-9-002Parkinson’s disease (PD) is associated with motor impairment and altered sensorimotor oscillatory activity. Brain computer interface (BCI) based neurorehabilitation may provide both therapeutic benefits and a means to assess underlying neurophysiology. In this study, we investigated whether people with PD (pwPD) can effectively use a motor imagery (MI) based BCI and whether MI related neurophysiological markers allow prediction of motor function at the individual subject level. Electroencephalography (EEG) data were analyzed from pwPD performing BCI therapy based on MI, functional electrical stimulation and visual feedback using a 3D avatar. PwPD achieved a mean BCI accuracy of 74%, which was not related to motor symptom severity as quantified by the Movement Disorders Society Unified Parkinson Disease Rating Scale (MDS-UPDRS) Part III. Using cross validated regression models, beta band event related (de)synchronization (ERD/S) during MI of all four limbs reliably predicted MDS-UPDRS Part III on an individual subject level. These findings demonstrate that pwPD can control an MI based BCI independent of motor impairment severity and extend existing literature beyond group level differences by identifying beta ERD/S as an individualized neurophysiological marker of motor function in PD.
• Unsupervised Intent Gating: Improving Calibration and Online Reliability in Motor Imagery BCIsForin, Paolo; Tortora, Stefano; Menegatti, Emanuele; Tonin, Luca; 10.3217/978-3-99161-093-9-003Motor Imagery (MI) Brain-Computer In terfaces (BCIs) are frequently hampered by fluctuations in user attention, which introduce noisy labels during standard calibration. To address this, we propose an unsupervised gating architecture using a Gaussian Mix ture Model (GMM) to dynamically estimate Intentional Control (IC). By evaluating sensorimotor rhythm spa tial focality, the GMM automatically filters out non informative idling segments without manual pruning. Of fline validation (N=5) demonstrated sharper neurophysi ological topographies and higher class separability than standard methods. During real-time closed-loop control, the GMM probability dynamically weighted a continu ous soft-increment integrator. While baseline sample wise accuracy remained comparable (77.7% vs. 77.4%), the gating mechanism effectively converted neural uncer tainty into safe timeouts, boosting effective active deci sion accuracy to 92.7% and maintaining a 76.0% safety margin against false positives during rest. This intent gated paradigm offers a robust framework for safe, real world assistive BCI technologies.
• Confidence-weighted cumulative rCCA with post hoc re-analysis: unsupervised adaptive learning for calibration-free c-VEP BCIThielen, Jordy; 10.3217/978-3-99161-093-9-004Abrain–computerinterface(BCI)typically requires a calibration period to train a decoding model, which is a tedious and time-consuming process. Previous studies have shown high performance of calibration-free decoding approaches, like adaptive reconvolution canon ical correlation analysis (rCCA). rCCA can operate in an instantaneous mode, classifying trials independently without prior information, but can be improved by cumu latively learning from previously classified trials. Such cumulative strategy, however, relies on potentially un reliable pseudo-labels. In this study, two extensions to unsupervised rCCA are introduced. First, confidence weighting ensures that model updates are driven primar ily by high-confidence trials. Second, a post hoc re analysis allows misclassifications to be corrected using a later, more accurate model. Four rCCA variants were evaluated on an existing code-modulated visual evoked potential (c-VEP) dataset, showing no significant effect of confidence-weighting, but a significant improvement of post hoc re-analysis. These findings pave the way for robust and fully calibration-free c-VEP BCI systems.
• Realtime Class Balancing for Efficient Decoder Adaptation in Asynchronous Brain-Computer InterfaceSouriau, Rémi; Martel, Félix; Costecalde, Thomas; Karakas, Serpil; Aksenova, Tetiana; 10.3217/978-3-99161-093-9-005Brain-Computer Interfaces (BCIs) allow direct control of external effectors via brain activity de coding. For real-world use, BCIs must operate in an asyn chronous mode that enables users to initiate or withhold system use, while maintaining low false positive rates (FPR) and high decoding accuracy. Class balance is cru cial for efficient model training. The RSW-NPLS algo rithm has been proposed recently to balance classes dur ing online and incremental learning. However, perfect balancing may not be optimal for asynchronous BCIs, as they require a highly stable idle state. We propose a class balancing strategy that prioritizes the idle state to reduce its false activations of the other states, and evaluated it on four databases from tetraplegic and paraplegic users. It significantly reduced the FPR (between-55.36% and 70.08%) with minimal loss of accuracy (between-3.68% and-10.51%).
• Combining Virtual Reality Feedback and Transcranial Alternating Current Stimulation to Enhance Restorative Brain-Computer InterfacesSaraiva, Joana; Esteves, Daniela; Bour, Christel; Fernandes, Sofia; Vourvopoulos, Athanasios; 10.3217/978-3-99161-093-9-006Instrokerehabilitation, brain–computer in terfaces (BCIs) enable neurofeedback-based regulation of sensorimotor rhythms, promoting motor-related corti cal reorganization. Integrating virtual reality (VR) into BCI paradigms enhances this process through immer sive, multisensory feedback that strengthens sensorimo tor engagement. In parallel, transcranial alternating cur rent stimulation (tACS) may further modulate motor im agery(MI)–related oscillatory activity. Together, these approaches offer a multimodal framework to optimize neurorehabilitation outcomes. This study investigated the synergistic effects of a MI-based VR-BCI paradigm and tACS applied at the individual alpha frequency (IAF). Twelve healthy participants performed bimanual MI tasks in VR and non-VR conditions before and after sham or active tACS. Alpha event-related desynchronization (ERD) served as the primary outcome. VR elicited stronger contralateral and more widespread ERD, includ ing parietal and occipital regions. tACS effects were sub tle and, in VR, sometimes attenuated ERD.
• Efficient Decoding of the Code-Modulated Motion Visual Evoked Potential (c-MVEP) for BCIsScheppink, Hanneke A.; Bialonski, Stephan; Tangermann, Michael; Thielen, Jordy; Volosyak, Ivan; 10.3217/978-3-99161-093-9-007The code-modulated motion visual evoked potential (c-MVEP) paradigm visually stimulates targets using motion in pseudo-random intervals. Due to its code-modulated nature, c-MVEP shares aspects of the code-modulated visual evoked potential (c-VEP). This study investigated whether the c-MVEP can be decoded using similar classification approaches as used for c VEP. Several canonical correlation analysis (CCA)-based methods were evaluated, with increasingly constrained assumptions about the neural data, to reduce model com plexity. These included ensemble and global spatial fil ters, as well as template estimation by averaging across trials, cycles, classes, or stimulus events. Consistent with findings in c-VEP, we found that the event-based ap proaches, with events defined by the rising or falling edge of the stimulus sequence, achieved a higher decoding ac curacy with fewer training data for calibration than the commonly used trial-based approach. This study shows the potential for decoding the c-MVEP with high accu racy, without requiring extensive training.
• Advancing NIRS-Based Insights into Motor Imagery and Execution of Swallowing: Effects of Artifact Correction Methods and Hemodynamic ValidationKober, Silvia Erika; Engele, Veronika Isabella; Kanatschnig, Thomas; 10.3217/978-3-99161-093-9-008Motor imagery (MI) of swallowing has proven to be a promising mental strategy for brain computer interface (BCI) and neurofeedback applications aimed at treating swallowing difficulties. Previous near-infrared spectroscopy (NIRS) studies demonstrated comparable hemodynamic responses over the inferior frontal gyrus during motor execution (ME) and MI of swallowing, particularly for deoxygenated hemoglobin. However, due to the reliance on manual artifact rejection, it remained unclear whether these findings were influenced by motion artifacts. In this study, 33 healthy young adults were tested, and their hemodynamic responses during ME and MI of swallowing were assessed using NIRS. By applying advanced and automated motion artifact correction methods, such as wavelet filtering and short-distance channel regression, the results were largely consistent with previous studies that used manual artifact rejection. This validates earlier findings and confirms that the observed activation patterns in the hemodynamic response were not primarily caused by motion artifacts, demonstrating the robustness of prior conclusions. When directly comparing motion artifact correction methods (manual rejection, wavelet filtering, and short-distance channel regression), wavelet filtering effectively reduced concentration changes in oxygenated hemoglobin compared to manual correction, indicating fewer motion artifacts. In contrast, short-distance channel regression showed no significant effects, underscoring the need for further optimization of correction methods and the importance of selecting appropriate artifact correction techniques in NIRS studies investigating ME and MI of swallowing.
• Inter-Expert Variability in Common Spatial Pattern Component Selection: Implication for Neurofeedback ApplicationsDumas, Cassandra; Dussard, Claire; Corsi, Marie-Constance; George, Nathalie; 10.3217/978-3-99161-093-9-009Common Spatial Patterns (CSP) are widely used in motor imagery (MI)-based brain computer interfaces (BCIs). While classification pipelines combine multiple components to optimize performance, MI-based neurofeedback (MI-NF) requires selecting a single spatial filter to drive feedback based on physiologically interpretable sensorimotor activity. This critical step is often implicit and relies on expert judgment. We quantified inter-expert variability in CSP selection using data from 20 subjects in a right-hand MI NF experiment. Twenty-three BCI or neurophysiology experts independently selected the most physiologically relevant component among six CSP candidates per subject. Inter-expert agreement was fair overall (Fleiss’ κ = 0.256) but varied substantially across subjects. Consensus did not systematically favor the top-ranked component, and selections spanned all six candidates. Although higher-ranked components elicited stronger agreement, substantial variability persisted. These findings identify single-component CSP selection as an overlooked source of methodological heterogeneity in MI-NF systems, highlighting the need for explicit and reproducible selection criteria.
• Resting-State EEG Predictors for BCI Performance: A Scoping ReviewSettgast, Tomko; Patel, Rishan; Kübler, Andrea; 10.3217/978-3-99161-093-9-010This scoping review summarizes studies investigating the association between resting-state EEG and subsequent BCI performance. Following PRISMA ScR guidelines, we systematically mapped the literature to provide a comprehensive overview. We discuss the results within the context of BCI performance prediction, emphasizing the importance of resting-state brain activity, particularly for patient-centered research.
• Comparison of sequence, stimulus and decoding algorithm combinations in c-VEP-based BCIsMartín-Fernández, Ana; Martínez-Cagigal, Víctor; Santamaría-Vázquez, Eduardo; Hornero, Roberto; 10.3217/978-3-99161-093-9-011Brain-computer interfaces (BCIs) enable users to control external devices using brain activity. Among exogenous paradigms, code-modulated visual evoked potentials (c-VEP) stand out. These are neural responses elicited by flickering visual stimuli encoded with pseudorandom sequences, allowing the system to identify the target attended by the user. Optimizing c VEP performance requires understanding the combined effects of encoding sequences, stimulus patterns, and de coding algorithms. In this study, 26 healthy participants evaluated a 16-target speller with 6 c-VEP configurations, combining 2 sequences (m-sequences, burst codes) with 3 stimulus patterns (plain, grating, checkerboard) using 4 decoding algorithms. Results showed that burst codes with a checkerboard pattern and a bit-wise reconstruc tion decoder offered the best balance between accuracy (99.76%) and comfort (eyestrain score 4.81/10). How ever, several combinations also yielded reliable perfor mance, highlighting that optimal design may depend on the specific application needs. This work emphasizes the importance of jointly considering sequence type, stimu lus pattern, and decoding algorithm in c-VEP research.
• Wearable brain-computer interfaces and their applications: advancements from the ARHeMlab research groupEsposito, Antonio; Moccaldi, Nicola; Gargiulo, Ludovica; Duraccio, Luigi; De Benedetto, Egidio; Galasso, Enza; Di Marino, Lucrezia; Angrisani, Leopoldo; Arpaia, Pasquale; 10.3217/978-3-99161-093-9-012This work presents recent advancements in the development of wearable brain–computer interfaces (BCIs) bythe ARHeMlabresearchgroup, focusingonthe period from 2021 to 2025. The activities address the key challenges of translating BCIs from controlled laboratory settings to real-world, daily-life scenarios. The research integrates wearable and portable EEG hardware with cus tom signal processing pipelines and XR environments to enhance signal quality and user engagement. Exper imental contributions span reactive, passive, and active BCI paradigms, including SSVEP-based interfaces with adaptive XR stimulation, cognitive load and distraction monitoring using low-channel EEG, and motor imagery control with multimodal neurofeedback. The proposed solutions are validated through application-specific stud ies, supporting the feasibility of practical, scalable, and ecologically valid BCI systems. No original research is presented in this work but the abovementioned works are reviewed.
• Evaluating Baseline Correction Techniques for EEG Time-Frequency Feature ClassificationHons, Manuel; Kober, Silvia Erika; 10.3217/978-3-99161-093-9-013EEG-based classification is sensitive to preprocessing procedures. Although baseline correction is widely applied in EEG research, its impact on classification performance remains underexplored. We compared four baseline strategies: (I) amplitude domain correction, (II) normalized power-domain correction, (III) combined amplitude- and power domain correction, and (IV) no baseline correction. Analyses were conducted using both an empirical speech imagination dataset and simulated data. Across both datasets, power-domain baseline correction produced significantly lower classification accuracies than amplitude-domain correction and no baseline correction. The replication of this pattern in simulated data indicates that these differences reflect fundamental signal-processing properties rather than dataset-specific confounds. These results suggest that baseline correction applied after nonlinear power transformation can distort feature distributions, whereas amplitude-domain correction removes additive offsets prior to nonlinear expansion, thereby preserving more stable variance structures.
• Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data AugmentationPoļaka, Viktorija; De Jong, Ivo Pascal; 10.3217/978-3-99161-093-9-014This paper addresses the challenge of gen erating synthetic electroencephalogram (EEG) covari ance matrices for motor imagery brain-computer inter face (MI-BCI) applications. Objective. We aim to de velop a generative model capable of producing high fidelity synthetic covariance matrices while preserving their symmetric positive-definite nature. Approach. We propose a Riemannian geometry-preserving variational autoencoder (RGP-VAE) integrating geometric mappings with a composite loss function combining Riemannian distance, tangent space reconstruction accuracy and gen erative diversity. Results. The model generates valid, representative EEG covariance matrices, while learning a subject-invariant latent space. Synthetic data proves practically useful for MI-BCI, with its impact depend ing on the paired classifier. Contribution. This work introduces and validates the RGP-VAE as a geometry preserving generative model for EEG covariance matri ces, highlighting its potential for signal privacy, scalabil ity and data augmentation.
• A Feasibility Study of a Motor Imagery–Based Brain–Computer Interface for Post-Stroke PatientsMartínez-Cagigal, Víctor; Moreno-Calderón, Selene; Cisnal, Ana; Pérez-Velasco, Sergio; Martín-Fernández, Ana; Santamaría-Vázquez, Eduardo; Fraile, Juan C.; Pérez-Turiel, Javier; Hornero, Roberto; 10.3217/978-3-99161-093-9-015Strokeisamajorcontributor to chronic dis ability, and many survivors continue to experience pro nounced upper-limb weakness despite standard rehabil itation. By converting motor intent into real-time sen sory feedback, brain–computer interfaces (BCIs) could help re-engage the sensorimotor loop. However, find ings across clinical studies remain mixed, especially for robot-assisted approaches. In this feasibility study, we investigated whether a noninvasive motor-imagery (MI) BCI providing concurrent multimodal feedback (visual avatar and robot-assisted hand motion) yields additional improvements beyond usual care. Sixteen patients with a first-ever stroke were assigned to conventional therapy alone (CG, n = 6) or to the same therapy supplemented with 23 one-hour BCI sessions(EG,n=10). OnlineEEG from eight sensorimotor channels was used to classify left- versus right-hand kinesthetic MI, and motor, cog nitive, and functional tests were evaluated before and af ter the intervention. The EG showed significant within group gains in voluntary motor control and coordination of the affected hand, episodic memory and visuocon structive ability, as well as functional measures such as perceived general health and independence. These pre liminary findings suggest that BCI-based therapies may enhance post-stroke rehabilitation.
• Noninvasive self-paced BCI for lower limb exoskeleton control: a multi-session investigation of user learningFaro, Alessio Lo; Cortecchia, Tommaso; Mihalovic, Miroljub; Trombin, Edoardo; Menegatti, Emanuele; Tonin, Luca; Tortora, Stefano; 10.3217/978-3-99161-093-9-016Brain-Computer Interfaces (BCIs) repre sents a promising approach to incorporate the user’s motion intention into the control of assistive technolo gies. However, its efficacy for the control of real lower limb exoskeletons (LLE) remain largely overlooked, with most of the work in the literature focusing on single session studies. To overcome these limitations, this work proposes and investigates the usage of a noninvasive electroencephalography (EEG)-based BCI for self-paced control of a LLE across multiple sessions. The proposed framework and training protocol are explicitly designed to promote progressive familiarization through repetitive training, continuous feedback, and minimal decoder cal ibration, encouraging the emergence of more stable and physiologically interpretable neural patterns. Experimen tal results show that all participants achieved functional LLE control within three sessions, paired with classifica tion improvements and stabilization of cortical activation patterns. Overall, the findings support the feasibility of practical LLE-BCI systems while highlighting the impor tance of stable decoding strategies in online and multi session experiments promoting user learning, which are key to achieving a more reliable and effective BCI-driven LLE control.
• The Effect of Single Session Implicit Motor Learning on Finger Motor Imagery for EEG BCIDamm, Laura M.; Jiang, Dai; Demostheneous, Andreas; 10.3217/978-3-99161-093-9-017Motor imagery (MI) based brain computer interfaces (BCIs) rely on cue driven paradigms, yet little is known about how the order in which cues are presented influ ences neural discriminability. Here the effect of sequen tial versus random cue order on finger-MI is investigated. Ten healthy, right-handed participants (7 F, mean age = 31.9 ± 10.7 years) completed two runs of 300 trials each (30 cycles of 5-finger trials). Across subjects, the γ band integrals showed the most consistent cue-order ef fect. Depending on frequency band 7-8 subjects yielded significant effects. Participants who began with the se quential cues exhibited higher mean amplitudes and re duced variability across runs, suggesting that early ex posure to a predictable cue sequence facilitates implicit motor-learning and stabilises the forward model. These findings demonstrate that cue-order manipulation is a po tent factor influencing MI-BCI performance and under score the importance of designing training protocols that promote implicit learning.
• Diversity in EEG: Why is it important?Kelly, Merlin Angel; Patel, Rishan; Bryan, Peter; Kier, Lexi; Carlson, Tom; 10.3217/978-3-99161-093-9-018Electroencephalography (EEG)-based brain-computer interfaces (BCIs) are increasingly deployed across clinical and consumer contexts, yet the research foundations underpinning these systems reflect a profound lack of demographic diversity. Hardware limitations, particularly the inability of standard EEG electrode systems to achieve reliable scalp contact across curly, kinky, and afro-textured hair types, represent a concrete and under acknowledged barrier to inclusive data collection. We present findings from a mixed-methods survey of 31 EEG researchers and practitioners, finding that 74% reported hardware-related challenges with diverse hair types, with difficulties clustering at cap placement and preparation. Qualitative themes revealed patterns of signal degradation, extended setup burden, explicit participant exclusion, downstream data loss, and a systemic researcher knowledge gap. Existing hardware solutions only partially address this problem. We argue that electrode redesign capable of achieving reliable scalp contact across all hair types is a prerequisite for BCI systems that work equitably for all.
• Decoding Sleep and Wakefulness from Chronically Implanted Intracortical Signals in a Communication BCI UserOffenberg, Elena Charlotte; Freudenburg, Zachary V.; Branco, Mariana P.; Zimmermann, Jonas B.; da Cruz, Janir R.; Vansteensel, Mariska J.; Ramsey, Nick F.; 10.3217/978-3-99161-093-9-019As brain-computer interfaces (BCIs) move toward long-term home use, continuous operation across day and night introduces new challenges. Decoders trained during wakefulness may produce unintended activations during sleep due to large differences in neural activity between vigilance states. Users must be able to communicate at night, making reliable detection of sleep and wakefulness essential for safe and autonomous BCI operation. We investigated whether awake-sleep stages can be inferred from chronically implanted intracortical neural signals in a communication BCI user with complete locked-in syndrome. Simultaneous non-invasive electroencephalography (EEG) and electrooculography (EOG) recordings provided reference sleep labels over a 44-hour period. A convolutional neural network was trained to distinguish between wakefulness and N2 sleep using intracortical data and achieved high accuracy (96.1±1.6%). The model was applied to six independent BCI spelling sessions, where more sleep-classified epochs coincided with reduced spelling performance. These results demonstrate the feasibility of decoding vigilance state from intracortical signals and support the development of state-aware BCIs for continuous home use.
• Proteus Effect in VR-Based BCI: How Avatar Age Influences Motor Imagery Behavior and EEGEsteves, Daniela; Vasconcelos, Marta; Vourvopoulos, Athanasios; 10.3217/978-3-99161-093-9-020The Proteus Effect suggests that embodying avatars in virtual reality (VR) with specific characteristics can activate related stereotypes and alter behavior. While prior work has shown that avatar age influences motor imagery (MI) performance, its impact on MIrelated EEG activity remains unclear. This pilot study investigated whether avatar age modulates both behavioral and EEG markers relevant for MI-based Brain–Computer Interfaces (BCIs). Sixteen healthy young adults performed a walking MI task in VR while embodied in either a young or elderly gender-matched avatar. Agerelated beliefs, embodiment, mental chronometry, and Alpha event-related desynchronization/synchronization (ERD/ERS) were assessed. Participants endorsed agerelated physical decline stereotypes and reported comparable embodiment across conditions. MI duration tended to be consistently longer when embodied in the elderly avatar. However, ERD/ERS analyses revealed modest and highly variable responses, with no consistent neural modulation across conditions.
• Deep learning based generalizable articulatory feature extraction for speech- BCI applicationsWu, Ruoling; Berezutskaya, Julia; Freudenburg, Zachary V.; Ramsey, Nick F.; 10.3217/978-3-99161-093-9-021Objective: Speech brain-computer interfaces (BCIs) often rely on paired neural-speech data, which is challenging to acquire from individuals with vocal tract paralysis. We aim to bridge this gap by proposing an across-subject framework that synthesizes intelligible speech using generalizable articulatory features extracted from an able-bodied cohort. Approach: We develop a subject-independent articulatory-to-speech model to extract generalizable articulatory features from real-time MRI (rtMRI) videos of a healthy cohort and use these features to synthesize text. Main Results: Our framework successfully synthesized text from the extracted generalizable articulatory features by achieving a Phoneme Error Rate (PER) of 18.9 % on unseen subjects. The results confirm that the extracted generalizable features remain robust and decodable across different subjects. Significance: By reducing the dependence on subject-specific articulatory data, our work offers a viable pathway for developing generalizable speech BCIs that could be used by those who can no longer articulate.
• Optimizing Bit-wise Linear Decoding for Imperceptible Code-Modulated Visual Evoked Potentials (I-c-VEP)Fodor, Milan A.; Tangermann, Michael; Thielen, Jordy; Volosyak, Ivan; 10.3217/978-3-99161-093-9-022Visual evoked potential (VEP)-based brain-computer interfaces (BCIs) are a well-established paradigm in the field, yet they remain limited by a fundamental usability constraint: they rely on perceptible low-frequency flicker, which induces visual fatigue and discomfort. To address this limitation, we previously proposed modulating an imperceptible high-frequency (60 Hz) carrier based on a pseudo-random binary code sequence. The resulting imperceptible code-modulated VEP (I-c-VEP) paradigm substantially reduces perceived flicker compared to conventional c-VEP approaches, while preserving key advantages over steady-state VEP (SSVEP) systems, namely efficient target scalability and robustness to narrowband interference. In this study, we explore whether simple, efficient linear decoding pipelines suffice for bit-wise I-c-VEP classification across 16 participants. By validating our pipelines on strictly independent, temporally held-out data, we demonstrated that approaches based on linear ridge regression and linear discriminant analysis can achieve over 97% mean trial accuracy. These findings provide a solid foundation for efficient and high-performance real-time implementations of the I-c-VEP paradigm.
• Ethical Considerations for User-Centric iBCI Studies: A Thematic ReviewThomas, Alex; Patel, Rishan; Podmore, Joshua James; Scherer, Reinhold; Carlson, Tom; 10.3217/978-3-99161-093-9-023Implanted brain-computer interfaces (iBCI) have been transformative for patients with paralysis, substantively restoring independent function in key domains. Despite this, the experimental nature and surgical complexity of implants means only 67 patients received long-term implants across published studies between 1998–2023. As the field grows, understanding what users actually want and need from these systems becomes critical to shaping future studies and applications. We performed a reflexive thematic review on the publicly available transcript of ’The Lived Experiences of BCI Users’ panel, conducted with three implanted BCI users, identifying 6 key themes: Purpose, Support, Understanding, Connection, Restriction, and Ethics. For each theme we discuss the core ideas with supporting quotes, derive best practice suggestions for future user-centred studies, and compare findings to existing literature, providing insight into the participant experience for researchers without direct exposure to user studies.
• EEG2EMG-Net: A Deep Learning Framework for Continuous EEG-to-EMG Estimation for Real-World ApplicationsMarcos-Martínez, Diego; Santamaría-Vázquez, Eduardo; Martínez-Cagigal, Víctor; Graz University of; Müller-Putz, Gernot; Hornero, Roberto; 10.3217/978-3-99161-093-9-024Estimating muscle activity from invasive neural recordings has been researched for decades with encouraging results. Such brain-computer interfaces (BCIs) are promising, as they offer an alternative pathway for motor control in patients with spinal cord injury. However, this application remains largely unexplored with non-invasive techniques, such as electroencephalography (EEG). Current non-invasive approaches typically rely on within-subject training, requiring both input (EEG) and output (electromyography, EMG) signals for calibration. This strategy is impractical for paralyzed patients unable to produce the required EMG. To address this, we developed EEG2EMG-Net a deep learning framework for continuous EEG-to-EMG estimation tailored for real-world clinical use. Our approach introduces a subject-adaptation strategy using synthetic EMG templates, eliminating the need for patient-specific muscle recordings during calibration. Results from 20 healthy participants show that the fine-tuning method allows accurate estimation of the EMG (pearson correlation coefficient: 0.69±0.32) while suppressing spurious activations during resting-state. These findings open new research avenues for non-invasive BCI alternatives aimed at assisting motor function in individuals with severe impairments.
• Neural and Clinical Variability in Immersive VR–BCI Training: A Case-Series Study with Stroke PatientsValente, Madalena; Oliveira, Ines; Fernandes, Jean-Claude; Almeida, Ana Isabel ; Figueiredo, Patricia; Vourvopoulos, Athanasios; 10.3217/978-3-99161-093-9-025Immersive virtual reality–based brain–computer interface (VR–BCI) systems have emerged as promising tools for post-stroke motor rehabilitation, providing embodied feedback and taskspecific training. However, reliable neurophysiological predictors of clinical responsiveness remain unclear. This case-series study investigated whether longitudinal modulation of event-related desynchronization (ERD) during VR–BCI training predicts upper-limb motor improvement in individuals with sub-acute and chronic stroke. Five participants completed 12 VR–BCI sessions over four weeks, with clinical assessments conducted pre- and post-intervention. ERD was extracted using individualized frequency bands, and linear mixed-effects models were applied to estimate subject-specific ERD trajectories, including baseline magnitude and sessionto- session change. These features were then examined in relation to motor recovery. Robust Alpha ERD was observed across participants, indicating consistent sensorimotor engagement. However, ERD progression across sessions did not significantly predict clinical gains, consistent with prior findings. Finally, BCI accuracy also showed substantial inter-subject variability and was not directly associated with motor improvement.
• Learning brain connectivity features for improving BCI performancePavaux, Marion; de Vico Fallani, Fabrizio; La Rocca, Daria; 10.3217/978-3-99161-093-9-026Brain computer interfaces (BCIs) rely on the decoding of information embedded in brain signals for applications ranging from neurofeedback to device control. While power spectral features are widely used in motor imagery-based BCIs, recent evidence suggests that considering network interactions between sensors could improve performance. In practice however, connectivity between different brain signals is impaired by the high variability of classical estimators for short noisy segments. To address this issue, we propose a deep learning framework trained to recover ground-truth imaginary coherence from complex simulated signals. Here, we show that our framework learns to reconstruct smooth coherence spectra, designed as Gaussian mixtures. Evaluation on an electroencephalography dataset across three motor execution and imagery tasks, shows that connectivity matrices derived from the proposed method consistently outperform those obtained withWelch’s estimator. These results demonstrate the feasibility of using deep learning to obtain stable informative coherence estimations in sparse and noisy contexts, supporting the use of functional brain networks for BCI applications.
• A Systematic Overview of EEG-Detectable Control-Related Mental States and Processes in HCIMatthias, Eidel; Settgast, Tomko; Silveira, Sarita Teja; Krol, Laurens R.; Wagner, Johanna; Kübler, Andrea; 10.3217/978-3-99161-093-9-027We report our endeavor to systematically map and describe all EEG-detectable mental states and processes (MS) with a focus on Human-Computer Interaction (HCI) from existing literature. Methods: Literature was collected in two steps. First, we performed a large-scale query against three relevant literature databases to systematically retrieve all articles dealing with EEG-detectable MS within a HCI context. Second, we broadened our scope to encompass all MS without the restriction to HCI. We obtained and screened the full list of paradigms from the #EEGManyLabs consortium and performed a second database query, looking specifically for influential reviews and metaanalyses about EEG-detectable MS. Results: We provide a limited overview, focusing on mental states relating to control processes, based on 1024 screened articles, thus far. Specifically, we describe the different MS within this category and their respective elicitation paradigms and neural correlates. Significance: While originally created for the NAFAS project, this overview may serve as a reference for anyone working with EEG-detectable mental states.
• Finger abduction trajectory prediction from high-density ECoGFaes, Axel; Merino, Eva Calvo; Mirsaeedi, Mani; Van Hoylandt, Anaïs; Keirse, Elina; Theys,Tom; Van Hulle, Marc M.; 10.3217/978-3-99161-093-9-028A case study is conducted to demonstrate that not only finger flexion but also finger abduction trajectories can be decoded from high density electrocorticography (ECoG) recordings. Two patients, temporarily implanted with high-density ECoG grids as part of their clinical workup, performed three cued tasks: single finger flexion, finger abduction, and sign language alphabet gestures. The extended Block-Term Tensor Regression (eBTTR) model is trained on simultaneous ECoG and data glove recordings to predict continuous finger trajectories. Performance on the flexion task serves as a baseline for comparison with abduction decoding. The sign language task involves realistic, complex finger movements combining both flexion and abduction. We show that finger abduction trajectories can be decoded from ECoG with precision comparable to single finger flexion, even during sign language gestures that involve joint finger flexions and abductions. To our knowledge, this is the first demonstration that finger abduction trajectories can be decoded from ECoG.
• EEG-Based Detection of Continuous and Discrete Hand Movements for Cursor ControlCrell, Markus R.; Müller-Putz, Gernot R.; 10.3217/978-3-99161-093-9-029Continuous cursor control for non-invasive brain-computer interfaces (BCIs) has recently received widespread attention. However, the usability of such systems for patients with e.g., neuromuscular diseases remains limited while a selection mechanism is not simultaneously provided. We therefore attempt to asynchronously detect and discriminate between continuous and discrete hand movements to enable intuitive cursor control and click possibilities for cursor BCIs. Our results show that discrimination between discrete and continuous motions is possible with high accuracy and a low rate of false positive detections. We propose that the introduced method can be applied to non-invasive cursor control BCIs to add a click functionality to the cursor movement.
• Riemannian approaches for ECoG-based Motor Imagery BCI decoding using spatial covariance matrices : a comparative studyEtienne, Mathis; Martel, Felix; Bonnet, Stéphane; Aksenova, Tetiana; 10.3217/978-3-99161-093-9-030Symmetric positive definite (SPD) covariance matrices have demonstrated strong potential for motor imagery (MI) decoding in EEG-based brain-computer interfaces (BCIs), yet their application to electrocorticography (ECoG) data remains largely unexplored. In this study, we evaluate and compare multiple classification approaches leveraging the Riemannian geometry of SPD matrices for ECoG-based MI decoding, across two paradigms: upper limb (3-class) and lower limb (3- class) motor imagery. Spatial covariance matrices were estimated from signals filtered in two task-specific frequency bands and used as input features. Three families of classifiers were compared: tangent space mapping (TSM) combined with machine learning models (LR, LDA, RSW-NPLS, LightGBM) and a multilayer perceptron (MLP); classifiers operating directly on the Riemannian manifold (MDM, FgMDM); and an SPD-aware neural network (SPDNetBN). Models were trained and tested offline on data from two implanted patients using lagcorrected balanced accuracy as the performance metric. FgMDM and TSM+RSWPLS achieved the highest performance for the upper limb paradigm (up to 91.6% and 89.5%, respectively), while SPDNetBN demonstrated superior accuracy for the lower limb paradigm (up to 69.4%).
• Why Simple Models Still Matter: Adaptive sLDA for Non-Stationary EEG DecodingEgger, Johanna; Müller-Putz, Gernot R.; 10.3217/978-3-99161-093-9-031Electroencephalography (EEG)-based brain-computer interfaces (BCIs) are strongly affected by non-stationarities, causing covariate shifts between calibration and deployment and leading to performance degradation over time. While deep learning approaches address this issue, they typically require large training datasets that are often unavailable in practical BCI scenarios. We therefore investigate lightweight online adaptation strategies for shrinkage linear discriminant analysis (sLDA), a simple yet robust classifier for lowdata EEG settings. We evaluate three unsupervised methods: exponential moving average feature normalization (FNorm), pooled mean adaptation of the classifier bias (PMean), and pooled mean with covariance adaptation (PMean+Cov). Across six sessions from 20 participants, without FP/min control, true-positive rates rose by 14.27% (FNorm), 29.13% (PMean) and 5.87% (PMean+Cov), while falsepositives per minute changed to 3.71 (+2.19) for FNorm, 11.95 (+10.44) for PMean, and 1.05 (-0.47) for PMean+Cov (not statistically significant). PMean+Cov achieved the best balance between sensitivity and false positives, making it most suitable for real-time BCI speller applications.
• Wrapped One-Class Riemannian EEG Classifier for BCI-Detection of Anesthetic StatesCueva, Valérie Marissens; de Surrel, Thibault; Laurent Bougrain; Bidgoli, Seyed Javad; Cheron, Guy; Alvarez, Ana Maria Cebolla; Meistelman, Claude; Lotte, Fabien; Yger, Florian; Rimbert, Sébastien; 10.3217/978-3-99161-093-9-032Brain-computer interfaces based on ElectroEncephaloGraphy (EEG) often represent efficiently brain signals as covariance matrices and leverage Riemannian geometry for classification of mental states. However, current approaches typically require multiple classes for training. In many applications such as intraoperative monitoring of consciousness during anesthesia, only data from one class, such as the awake state, may be available, requiring one-class classification methods. We propose a novel Riemannian one-class classifier called the One-Class Wrapped Gaussian. Unlike existing methods that rely solely on the Riemannian mean, our approach incorporates second-order statistical information by using an anisotropic Gaussian-like distribution on the manifold of covariance matrices. We validated our method on EEG data from 19 patients undergoing general anesthesia. Results show that our classifier significantly outperforms state-of-the-art one-class methods for distinguishing between awake and anesthetized states, for clinically relevant electrode numbers. We further demonstrate the robustness of our approach by testing configurations with reduced electrode numbers, confirming its feasibility for real-world surgical settings where electrode placement is constrained.
• Filling the gap: sex-specific differences in post-task interval dynamicsLeitner, Michael; Wriessnegger, Selina C.; 10.3217/978-3-99161-093-9-033Electroencephalography (EEG) is widely used to investigate neural correlates of cognitive work load, fatigue, and stress through changes in oscillatory band power. While many studies have focused on neu ral activity during task execution, the post-task resting period following cognitive effort remains comparatively underexplored, despite its potential relevance for under standing how cognitive resources are restored. More over, little is known about whether neural dynamics dif fer between male and female participants. In this study, we analyzed EEG band power during the post-task rest ing interval following a cognitively demanding working memoryparadigm. Relativebandpower(rBP)wasexam ined across multiple frequency bands, scalp regions, and rest epochs to characterize temporal and spatial patterns. Significant sex-related differences were observed in the theta, alpha, and beta bands. Females exhibited higher theta activity, particularly in central regions, whereas males showed elevated alpha band power. Across rest ing epochs, alpha activity displayed the most pronounced increase, suggesting progressive reallocation of atten tional resources. Topographic analyses further revealed region-specific differences in central, parietal, and occip ital areas. These findings suggest that post-task interval dynamics exhibit distinct oscillatory characteristics that vary between sexes. Future studies might consider such baseline phases leading to improved EEG-based detection of cognitive states and inform the design of neuro adaptive systems.
• Is User-Independent MI-BCI Performance Influenced by Gender Proportion in Training Users?Kojima, Simon; Lotte, Fabien; 10.3217/978-3-99161-093-9-034In motor-imagery brain–computer inter faces (MI-BCI) research, how user gender influences BCI classification performance and ERD/S responses is not yet fully understood. Thus, investigating the impact of gender on BCI systems is an important step toward real izing BCIs that are accessible to all users. In this study, we examined how varying the proportion of female users in the training data in cross-user MI-BCI affects BCI clas sification performance. The results (on 139 users from 2 data bases) revealed positive correlations (0.33–0.71) be tween the proportion of female users in the training data and BCI performance, and showed that including more female users in the training data was associated with per formance improvements of up to 5.5%. These findings indicate that machine-learning models in MI-BCI may be influenced by gender-related biases and highlight the im portance of considering gender distribution in the design of future BCI research.
• Non-invasive measurement of somatosensory evoked spinal cord potentialsGolser, Bernhard; Oberndorfer, Markus; Kostoglou, Kyriaki; Müller-Putz, Gernot R.; 10.3217/978-3-99161-093-9-035The spinal cord is more than a mere re lay station between the brain and the peripheral nervous system. Recent research indicates that complex process ing already occurs at the spinal level. The non-invasive recording of spinal cord activity acquired with high density surface electrode arrays may advance our under standing of spinal network dynamics and the integration of spinal data could enhance existing brain–computer in terfaces. In the present study, electroencephalography and electrospinography were used to measure cortical and spinal somatosensory evoked potentials elicited by median and tibial nerve stimulation in 10 healthy indi viduals. In contrast to previous studies, participants were seated upright with no external support of the head and stimulus intensities were kept strictly below the move ment threshold. Under these more physiologically real istic settings, two of the three well-characterized spinal cord components (N13 and P22) were successfully de tected, indicating a broader variety of potentially success ful experimental setups.
• Towards thermal imagery-based brain-computer interfacesLefeuvre, Théo; Savalle, Emile; Petit, Jimmy; Macé, Marc J-M.; Lécuyer, Anatole; Pillette, Léa; 10.3217/978-3-99161-093-9-036Mental task-based brain-computer inter faces (BCIs) enable users to control a system by volun tarily modulating their brain activity without overt move ment. Research in the field has focused mainly on motor imagery, whereas tactile imagery, defined as the imagi nation of tactile stimuli on the skin, remains compara tively underexplored. Although pressure and vibratory imagery have received growing attention, other modal ities such as thermal imagery remain largely neglected, despite their potential clinical relevance. In this EEG based BCI study, healthy participants learned to con trol a BCI by imagining a warm sensation in the right hand, representing, to our knowledge, the first thermal imagery-based BCI user training. To facilitate mental representation, participants received visual, thermal, or combined visuothermal instructions. Thermal imagery elicited event-related desynchronization, primarily over left sensorimotor regions, and online performance im proved significantly across runs. Comparisons between instruction modalities suggested trends favoring visual and thermal guidance. Together, these findings support the feasibility of thermal imagery as a promising alterna tive control paradigm for BCI applications.
• Movement-Onset-Aligned EEG for Generalizable Arm and Hand Movement Decoding: A Large-Scale Multi-Experiment DatasetSterk, Kathrin; Kostoglou, Kyriaki; Müller-Putz, Gernot R.; 10.3217/978-3-99161-093-9-037This work presents a curated, movement onset-aligned electroencephalography (EEG) dataset comprising nine experiments with 188 able-bodied par ticipants performing hand and arm movement execution tasks as well as rest conditions. Movement onsets were aligned using auxiliary sensors. All recordings were uni formly processed using causal filtering to ensure com patibility with online brain-computer interface (BCI) ap plications. Further processing steps include resampling, independent component analysis (ICA) for removal of ocular and noise components, and strict epoch rejec tion. Across experiments, the datasets include recordings with different channel configurations, covering 94 dis tinct EEG channels. In total, the dataset comprises over 95,000 annotated trials with detailed movement labels and is provided in a standardized HDF5 structure to fa cilitate flexible data selection in machine learning work flows. Signal analyses confirm the presence of physiolog ically plausible movement-related patterns over sensori motor regions. This resource aims to support the devel opment of cross-participant and cross-experiment gener alizing movement decoders and transfer learning in EEG based BCIs. The data will be released in the near future.
• Benchmarking Machine Learning strategies to classify Steady-State Auditory Evoked Potentials.da Silva, Henrique Lefundes; Guého, Lenaïg; Plapous, Cyril; Bougrain, Laurent; Hénaff, Patrick; Nicol, Rozenn; 10.3217/978-3-99161-093-9-038This study benchmarks classification pipelines for auditory Brain–Computer Interfaces (BCIs), with the objective of detecting the modulation frequency of amplitude-modulated sinusoidal auditory stimuli in electroencephalographic recordings. For this purpose, state-of-the-art Steady-State Visually Evoked Potential (SSVEP) classification algorithms are adapted to the au ditory domain. This choice is motivated by the neuro physiological similarities between SSVEP and Steady State Auditory Evoked Potential (SSAEP), particularly their shared mechanism of neural entrainment to peri odic stimuli, despite differences in stimulation frequen cies and the cortical regions involved. In this context, both preprocessing strategies and classification models are systematically assessed to identify optimal parame ter configurations. Among the filtering techniques exam ined, a narrow-band Bessel filter yielded the highest ac curacies, while the effect of temporal windowing was less pronounced under narrow-band conditions but remained beneficial for broader bandwidths. Riemannian-based methods, particularly tangent-space classifiers, consis tently outperformed alternative approaches. Neverthe less, cross-subject classification proved more challeng ing due to pronounced inter-subject variability. These findings provide practical recommendations for auditory BCIs.
• Capacity-normalized Shannon graph complexity as an EEG biomarker for event-level cognitive load estimation in adaptive BCIsPascual-Roa, Beatriz; Santamaría-Vázquez, Eduardo; Marcos-Martínez, Diego; Ruíz-Gálvez, C. Rubén; Martínez-Cagigal, Víctor; Hornero, Roberto; 10.3217/978-3-99161-093-9-039Cognitive load (CL) reflects the mental resources required to perform a task. Adaptive brain computer interfaces (BCIs) could benefit from EEG biomarkers that estimate CL with sufficient temporal granularity to support online adaptation. However, many EEG-based CL studies still define CL according to pre defined task levels or objective task difficulty, which may not impose comparable cognitive demands across users. This assumption is especially problematic when individ uals differ in working-memory capacity. We propose an event-level and capacity-normalized framework based on EEG functional connectivity in frontal and parietal re gions. During the Corsi Block-Tapping Test, each en coded item was labeled according to its position rela tive to each participant’s individual maximum working memory capacity, allowing cognitive demand to be ex pressed as a percentage of individual capacity. Functional connectivity was quantified using orthogonalized am plitude envelope correlation (AEC) across theta, alpha, low- and high-beta, and gamma frequency bands, while network organization was summarized through Shannon graph complexity (SC). The approach was evaluated in 62 cognitively healthy older adults (mean age = 70.4±2.9 years). Across all frequency bands, SC showed a pro gressive increase as task demands approached individ ual working-memory capacity. Capacity-normalized CL stages differed significantly after FDR correction (p < 0.05), with mostly large effect sizes (r >0.5). These find ings suggest that SC, combined with capacity-normalized event-level labeling, may provide a compact and online compatible EEG biomarker for CL-aware BCI adaptation in cognitively healthy older adults.