EEG EYE STATE CLASSIFICATION: INVARIANT AND INTERPRETABLE WITH CROSS DATASET ADAPTATION AND NEUROPHYSIOLOGICAL CONSTRAINTS
Taouli Sidi Ahmed
Biomedical Engineering Research Laboratory,
Department of Biomedical Genius, Faculty of Technology,
University Aboubekr Belkaid,
University Aboubekr Belkaid, P.O. Box 230, Chetouane, Tlemcen 13000, Algeria
Abstract: Electroencephalography (EEG)-based eye state classification is a key task in neuroscience and biomedical engineering, widely employed in cognitive monitoring and brain computer interface systems. While many studies report high classification accuracy within a single dataset, performance often deteriorates across different subjects or databases due to inter-subject variability and heterogeneous acquisition conditions. In this study, we introduce a subject-invariant and interpretable approach for EEG eye state classification that explicitly addresses cross-database generalisation. Our approach combines multi-domain EEG features with correlation alignment (CORAL) for domain adaptation and integrates neurophysiological constraints grounded in established EEG principles, including alpha-band reactivity, functional connectivity, and signal complexity. We evaluated the proposed method on three real-world universal EEG databases: PhysioNet, SPIS, and MNNIT. In cross-database evaluations, conventional models achieved accuracies of 63.4–66.1%, revealing limited generalisation. Incorporating CORAL improved performance to 72.1–74.8%, and the addition of neurophysiological constraints further increased accuracy to 81.5–83.2%. Notably, interpretability was maintained through SHAP-based feature importance analysis, identifying alpha-band power, occipital-parietal coherence, and entropy as the most influential features. These findings demonstrate that combining domain adaptation with physiologically informed constraints substantially enhances cross-database EEG eye-state classification, providing a robust and clinically interpretable solution for real-world EEG applications.
Keywords: Electroencephalography (EEG); Eye-state classification; CORAL; Neurophysiological constraints; Model interpretability.
VOLUME 10 ISSUE 04 2026: 1-13