Open research questions in EEG and Brain-Computer Interfaces

What the literature leaves open

• Future research should than general equivalence focus on configuration-specific optimization rather testing. On the hardware side, improvements in electrode-skin coupling stability, shielding, and multi-contact intra-auricular arrays may enhance spatial sensitivity without sacrificing wearability. On the methodological side, artifact modeling approaches tailored to the ear environment and paradigm designs calibrated to expected effect sizes may expand the range of reliably measurable signals.

  Signal-specific performance of in-ear EEG: strengths and limitations · 2026 · DOI

• Existing supervised and self-supervised EEG models mainly learn discriminative or reconstructive representations within individual segments, while the transition information between adjacent EEG segments remains underexplored.

  Multimodal EEG World Model: Self-Supervised Latent Transition Learning for Wearable EEG Seizure Detection · 2026 · DOI

• the performance gains The current study also focuses mainly on disease classification, while more fine-grained clinical questions remain unexplored. Therefore, the generalization ability of the proposed framework needs to be further examined on larger multicenter cohorts and independent clinical datasets. Although the results on APAVA, ADFTD, and TDBrain suggest that ScaleSpecter can provide effective representations for different neurodegenerative disease- related EEG classification tasks, the current evidence is still limited by the scale and heterogeneity of the available data.

  ScaleSpecter: a frequency-aware multi-scale patch framework for robust physiological classification under non-stationarity · 2026 · DOI

• 255 zeuspress.org ; Computers and Artificial Intelligence; Vol.3, No.3 2026 This research focuses on signal preprocessing, model construction, parameter setting and experimental result analysis through the classification and identification of EEG electroencephalogram signals as the research object. By establishing a deep learning classification model and conducting feature learning and classification judgment of input signals, the experimental results show that the method can identify the signal characteristics of different categories to a certain extent, which has a good classification effect. On the whole, this article completes the basic process from data processing to model training and result analysis, which provides a certain reference for the subsequent further optimization of the intelligent diagnosis method of EEG signals. Follow-up research can be further verified by combining real clinical experimental data. By introducing more sources and larger-scale clinical electroencephalogram data, the stability and generalization ability of the model in actual medical scenarios can be tested. At the same time, it can also be compared and analyzed with the doctor‘s diagnosis results, evaluate the reliability of the model in auxiliary diagnosis, and provide support for the application of intelligent medical auxiliary diagnosis system.

  Automatic Detection of Epileptic EEG Signals Based on Channel Attention and Bidirectional Temporal Modeling · 2026 · DOI

• Future research could address this gap by incorporating datasets with more detailed channel data, allowing for a more comprehensive understand- ing of the neurophysiological differences between ASD and TD populations.

  Robust and Interpretable Deep Learning on EEG Spectrograms for Autism Spectrum Disorder Detection · 2026 · DOI

• Given that experiments relating cognitive processing and EEG complexity are still scarce, this work is a narrative review of studies in which non-clinical cognitive processing, such as memory, perception, or attention, is addressed using complexity measures.

  Cognitive Processing and EEG Complexity · 2026 · DOI

• The results highlight the potential of SCP-based decoding for directional control and motivate further investigation of speed-related neural signatures.

  Exploring Kinematics Information Decoding from EEG Slow Cortical Potentials During Movement Imagination and Observation · 2026 · DOI

• Specifically, these concerns relate to the description of EEG data acquisition; inconsistencies between reported statistical results and narrative interpretations; discrepancies in classification accuracy values; incomplete or inconsistent sample and protocol information; insufficient details of ethical approval; and limited information on the deep learning methodology and data transparency.

  Expression of Concern: Examining Cognitive Shifts Through EEG: Insights from Resting State to Neurofeedback Game Engagement<b></b> · 2026 · DOI

• Although laboratory experiments have demonstrated physiological indicators associated with cognitive load, it remains to be established whether these patterns can consistently report on cognitive load in a real-world environment. Future work with larger samples and more trials per condition will be needed to confirm that this opposing trend reflects a genuine contextual difference rather than an artifact of limited data.

  Multimodal quantification of cognitive load using a printed wearable facial bio-potential system · 2026 · DOI

• While electroencephalography (EEG) provides a promising avenue for pain assessment, it remains unclear whether phase- or power-based neural integration drives pain-state discrimination.

  Predicting Pain: Electroencephalography Signatures of Neural Integration During Experimental Tonic Thermal Pain · 2026 · DOI

• Current human-machine collaboration (HMC) systems rely on environment-facing sensors to observe visible actions and scene states, but the internal perceptual, intention-related, and state-related processes of operators remain insufficiently integrated into machine perception.

  DS-MTNet:Structured Multi-Task EEG Decoding for Human-Machine Collaboration · 2026

• Future studies should address these limitations through larger and more diverse participant cohorts, higher-density EEG systems, cross-dataset generalization experiments, and deep learning architectures that can learn directly from raw EEG signals. Future research should investigate adaptive or hybrid feature therefore features, whereas [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] pp.

  Redundancy-Aware Feature Selection using mRMR and F-Test for EEG Emotion Classification · 2026 · DOI

• Conclusion This exploratory within-subject study provides preliminary evidence that the mechanism-driven, realtime respiratory modulation protocol may be associated with reductions in negative affective states and changes in cortical activity. At the behavioral level, post-intervention assessments utilizing the POMS questionnaire revealed significant reductions in Tension, Anger, Fatigue, Depression, Confusion, and Total Mood Disturbance (TMD) scores, accompanied by a marked increase in self-esteem.

  A mechanism-driven real-time respiratory modulation framework for rapid affective regulation via prefrontal EEG computational phenotyping · 2026 · DOI

• Although standardized protocols such as the SINBAR criteria provide guidance for PSG scoring [12], inter-rater variability in RBD diagnosis has not been systematically quantified.

  ActiTect: a generalizable machine learning pipeline for REM sleep behavior disorder screening through standardized actigraphy · 2026 · DOI

• Multi-Modal Integration: Integration of EEG with other modalities (neuroimaging, genetic, clinical, cognitive as- sessments) represents an underexplored opportunity for en- hanced diagnostic accuracy through complementary informa- tion sources. Independent external validation is scarce and generally performed using data from different laboratories.

  Identification of Alzheimer's Disease based on EEG Signal using Artificial Intelligence Techniques · 2026 · DOI

• The duration of illness among the included pDoC participants varies widely, which may affect their ability to perform motor imagery and their responsiveness to brain-computer interface tasks. Additionally, further investigation of quantitative EEG feature analysis across different brain regions is warranted.

  Effects of motor imagery brain-computer interface task on quantitative EEG features in patients with prolonged disorders of consciousness · 2026 · DOI

• Despite these advantages, standardised data analysis frameworks specifically tailored to OPM technology are still lacking, leading to variability in processing choices and reduced reproducibility across laboratories and hardware platforms.

  OPM-FLUX: A Pipeline for OPM MEG Data Analysis · 2026 · DOI

• Whether the P300 event-related potential, an established marker of attention and cognitive processing, can be elicited as an incidental byproduct of genuine gameplay, recorded with a minimal wearable EEG system under unsupervised home conditions, has not been established.

  Wearable EEG during gameplay captures a robust P300 cognitive signal in unsupervised home settings · 2026 · DOI

• However, visually-evoked EEG datasets remain scarce, leading existing methods to align neural signals mainly with abstract text, a lossy translation that may discard fine-grained perceptual information encoded in brain activity.

  Visualizing the Invisible: Generative Visual Grounding Empowers Universal EEG Understanding in MLLMs · 2026

• Limited dataset information Limited discussion on computational efficiency 1D CNN, BiLSTM UCI LSTM, MSA-DCNN UoB Precision, Sensitivity, Specificity, F1 Sensitivity, specificity…

  Deep Learning for EEG-based epilepsy seizure detection and prediction: a comprehensive review of datasets, architectures, and clinical challenges · 2026 · DOI

• Multimodal signal fusion (EEG plus EMG plus fNIRS), while theoretically offering complementary advantages, remains an open engineering challenge for the efficient joint decoding of multimodal features on embedded real-time platforms.

  Advances in mechanisms of neuroplasticity induced by multimodal closed-loop brain–computer interfaces after stroke · 2026 · DOI

• 6.5.1 Sample size and statistical power The most pervasive limitation across the BCI rehabilitation literature is small sample size. The majority of published RCTs involve fewer than 20 participants per group, providing limited statistical power to detect moderate effect sizes. Meta-analytic pooling by Cervera et al. (2018) across 476 patients provided the most statistically robust estimates to date (overall SMD = 0.53), but the high heterogeneity across included studies limits the preci- sion of subgroup analyses. Future research requires adequately powered multicenter RCTs with pre-registered sample size calculations. 6.5.2 Outcome measure heterogeneity Substantial heterogeneity exists across studies in the selection, timing, and analytical methods applied to both neuroplasticity and clinical outcome measures. Establishing a standardized multidi- mensional assessment battery—incorporating at minimum one electrophysiological index (EEG ERD or TMS MEP), one neuro- imaging index (fMRI functional connectivity or DTI FA), and one validated clinical outcome measure (FMA-UE)—assessed at stan- dardized time points (baseline, post-treatment, 3-month follow- up, 12-month follow-up) should be adopted as a field-wide standard. 6.5.3 Control condition design The majority of BCI rehabilitation studies employ passive control conditions that do not adequately isolate the specific contribution of closed-loop neural feedback from non-specific therapeutic effects. Active sham-controlled designs, in which control participants receive matched sensory stimulation delivered non-contingently or at random intervals, are essential for establishing that neuroplasticity effects are specifically attributable to the Hebbian temporal contingency mechanism. 6.5.4 Population representativeness The predominance of chronic stroke populations (>6 months post-stroke) in published BCI rehabilitation studies limits the gener- alizability of findings to the subacute phase. Prospective studies spe- cifically targeting the subacute phase, with longitudinal neuroimaging assessment, are urgently needed. Additionally, the underrepresenta- tion of patients with severe motor impairment, aphasia, and signifi- cant cognitive deficits limits the applicability of current evidence. 6.5.5 Offline-to-online generalization gap The substantial performance gap between offline benchmark results and online clinical performance remains a critical unresolved challenge. Prospective validation of decoding algorithms under eco- logically valid conditions—including variable electrode impedance, movement artifacts, fatigue-related signal drift, and real-world noise environments—is an essential prerequisite for safe and effective clini- cal translation. 6.5.6 Long-term maintenance of effects Evidence for the long-term maintenance of BCI-induced neuro- plasticity beyond 6 months post-intervention remains limited. Whether the structural and functional plasticity changes induced by multimodal closed-loop BCI intervention are permanent, progressive, or subject to decay in the absence of continued training is a funda- mental question with direct implications for rehabilitation program design. Large-scale, well-designed long-term follow-up RCTs extend- ing to 12 months or beyond are an urgent priority. 6.5.7 Future directions Building on the limitations identified above, the highest-priority future research directions include: (1) reinforcement learning-based personalized adaptive closed-loop systems that autonomously opti- mize training parameters based on individual neural responses in real time; (2) multimodal biomarker-guided patient stratification frame- works integrating baseline EEG complexity, CST integrity, MI ability, and lesion characteristics; (3) multicenter large-sample RCTs with standardized outcome batteries, active sham controls, and long-term follow-up; and (4) fully home-deployable multimodal BCI rehabilita- tion platforms combining low-density wearable EEG, portable FES, and consumer-grade VR with single-calibration decoding algorithms and remote monitoring capabilities. A critical engineering design principle derived from this review: feedback latency must be main- tained below 100 ms (ideally ~100 ms post-MRCP peak) to reliably activate STDP-based Hebbian plasticity; system designs exceeding 200 ms total latency should be considered outside the effective neuro- plasticity induction window.

  Advances in mechanisms of neuroplasticity induced by multimodal closed-loop brain–computer interfaces after stroke · 2026 · DOI

• This work is evaluated only on the CHB-MIT corpus, so generalisation across acquisition devices, channel montages and clinical populations is not yet established. Energy efficiency is reported from GPU power telemetry and a single Raspberry Pi 4 measurement; deployment on a dedicated low-power inference accelerator may show a different power profile. Predictive performance varies between patients – for Patient 02 and Patient 09 the model does not statistically outperform the random predictor – and the headline sensitivity is therefore a mean over subjects rather than a per-subject guarantee. 184 Journal of Edge Computing, 2026, Vol. 5, Iss. 1, pp.

  Optimising seizure prediction with reduced computational resources using depthwise CNN · 2026 · DOI

• techniques can Multimodal analysis using diverse input data such as EEG signals, ECG signals, accelerometry signals, and others gives more insight into the patient's health state. The use of deep improve seizure detection by learning incorporating prior knowledge from related tasks. Domain adaptation can be another promising technique in this context, which uses information from relevant domains. In particular, researchers should explore explainable AI techniques to make epileptic seizure detection models explainable. Continued technological advancements and increased interdisciplinary collaborations make the future of epileptic seizure detection promising.

  EEG-Based Epileptic Seizure Detection and Classification Using Machine Learning · 2026 · DOI

• Future work could explore adaptive loss-weighting strategies to dynamically balance the contribution of DML components during training for further optimization. Nevertheless, a key limitation lies in sensitivity, as misclassifying seizure segments as NSZ—especially in cases with limited seizure data—may result in loss of data utility and missed diagnoses.

  M ² -Net: A Multiscale Multitask Neural Network for EEG-Based Seizure Detection · 2026 · DOI

Most-cited papers in EEG and Brain-Computer Interfaces

• The neurophysiological bases of <scp>EEG</scp> and <scp>EEG</scp> measurement: A review for the rest of us · Psychophysiology · 2014 · 246 citations
• The Maryland analysis of developmental EEG (MADE) pipeline · Psychophysiology · 2020 · 199 citations
• An Accurate and Rapidly Calibrating Speech Neuroprosthesis · New England Journal of Medicine · 2024 · 198 citations
• Residue iteration decomposition (RIDE): A new method to separate ERP components on the basis of latency variability in single trials · Psychophysiology · 2011 · 177 citations
• Classification of Hand Movements From EEG Using a Deep Attention-Based LSTM Network · IEEE Sensors Journal · 2019 · 175 citations
• Deep Learning for Patient-Independent Epileptic Seizure Prediction Using Scalp EEG Signals · IEEE Sensors Journal · 2021 · 171 citations
• Standardized measurement error: A universal metric of data quality for averaged event‐related potentials · Psychophysiology · 2021 · 165 citations
• Impedance and Noise of Passive and Active Dry EEG Electrodes: A Review · IEEE Sensors Journal · 2020 · 162 citations
• CTNet: a convolutional transformer network for EEG-based motor imagery classification · Scientific Reports · 2024 · 140 citations
• A survey on methods and challenges in EEG based authentication · Computers & Security · 2020 · 139 citations

Most recent work

• Enhancing Unseen Driver State Detection: An EEG-Based Framework with Brain Connectivity and Depthwise Separable Convolutional Neural Networks · Biomedical Signal Processing and Control · 2026
• Toothy: an interactive platform for dentate spike curation · bioRxiv · 2026
• NeuroPath: Practically Adopting Motor Imagery Decoding through EEG Signals · 2026
• Optimising seizure prediction with reduced computational resources using depthwise CNN · Journal of Edge Computing · 2026
• Spectral Validity and Spindle Detection of Wearable Frontal EEG: A Per-Subject Calibration Framework and Systematic Validation Against Polysomnography Using the Wearanize+ Dataset · medRxiv · 2026
• Signal Quality Screening and Automated Sleep Stage Agreement in Home EEG: A Systematic Comparison of Dreamento and YASA on the Wearanize+ Dataset · medRxiv · 2026
• Brain–Computer Interface: From Human thought to Machine Control · International Journal of Advanced Research in Science Communication and Technology · 2026
• Emergence of Artificial Intelligence in Multiple Domains of Neurology: A Review · British Journal of Healthcare and Medical Research · 2026
• EEG-based Imagined Speech Analysis using Functional Connectivity · International Journal of Soft Computing and Engineering · 2026
• Zynq-Optimized EEG Artifact Removal and Sleep Stage Monitoring Using AI and Edge Acceleration · International Research Journal of Modernization in Engineering Technology and Science · 2026

Related topics in Neuroscience