Drowsiness reduces efficiency in perceptual processing, reaction time, and executive control, posing risks in safety-critical domains such as driving and long-duration monitoring tasks. EEG-based fatigue detection has emerged as a powerful approach for quantifying early neurophysiological signs of vigilance decline, yet many proposed algorithms are insufficiently evaluated in strictly subject-independent conditions. To address this gap, we systematically compare classical machine learning models, Riemannian geometry-based classification, and compact deep neural architectures on a publicly available electroencephalography (EEG) dataset containing 11 subjects. We employ a rigorous leave-one-subject-out (LOSO) protocol, ensuring that no individual contributes information simultaneously to the training and test sets. The study evaluates logistic regression, support vector machines with radial-basis kernels, random forests, a Log-Euclidean Riemannian classifier, EEGNet, a transformer encoder, and a bidirectional long short-term memory (BiLSTM) with temporal attention. Across folds, accuracy and macro-F1 scores were calculated and summarized with mean and standard deviation. The BiLSTM-attention model achieved the highest performance (accuracy ; macro-F1 ) but only moderately exceeded EEGNet and the classical baselines. Wilcoxon signed-rank tests revealed no significant difference between EEGNet and BiLSTM (), although BiLSTM significantly outperformed the transformer model (). Analysis of error structure demonstrated a notable asymmetry with 295 false positives and 184 false negatives aggregated across folds. Band-specific analysis revealed theta activity as the strongest contributor to class separation, followed by delta and alpha rhythms. Channel-importance analysis indicated that posterior and paracentral regions were consistently more informative. These findings highlight that model complexity does not guarantee superior performance in small datasets with large inter-subject variability. The study provides a transparent, fully reproducible baseline for future fatigue-classification research and demonstrates the practical relevance of compact architectures and Riemannian geometry in low-data conditions.
28 October 2026 in Traumatology and Orthopаedics of Kazakhstan
Objective. EEG stress classifiers are usually trained and tested on a single dataset, and accuracies above 90% are common in the literature. We asked how much of this performance holds for participants the model has never seen, and for a dataset recorded with other hardware and another stress protocol.Approach. Two open datasets were processed with one pipeline: SAM40 (40 participants; Stroop, arithmetic and mirror-image tasks versus relaxation; 32-channel Emotiv, 128 Hz) and EEGMAT (36 participants; serial subtraction versus rest; 19-channel clinical system, 500 Hz). Four compact classifiers(logistic regression, random forest and RBF-SVM on band power, and a Riemannian tangent-space classifier) were evaluated with leave-one-subject-out (LOSO) validation, with cross-corpus transfer in both directions on ten shared electrodes, and with two unsupervised alignment methods, Riemannian re-centring and CORAL. Main results. Within SAM40, the best LOSO macro-F1 for one stressor against relaxation was 70.3% (best AUROC 0.804). Cross-corpus transfer lowered macro-F1 significantly in 22 of 24 model–direction–stressor settings (pairedWilcoxon test, ???? < 0.01), by up to 37.6 percentage points. Transfer was asymmetric: classifiers trained on EEGMAT reached an AUROC below 0.5 on SAM40 in 9 of 12 settings, so the learned decision direction was reversed. Re-centring improved within-corpus performance (EEGMAT macro-F1 from 0.690 to 0.728) but not transfer, whereas CORAL raised the EEGMAT→SAM40 AUROC of band-power models by 0.04–0.19. Removing ocular components withICA reduced band-power macro-F1 by only 1.1–3.0 points. Significance. Stress labels defined by the task carry protocol-specific information that does not survive a change of dataset. Cross-corpus testing should come before any claim that an EEG stress detector is ready for everyday or biofeedback use.