AN INTERPRETABLE MACHINE LEARNING APPROACH FOR EEG-BASED EYE-STATE CLASSIFICATION: A FOCUS ON EYES-OPEN VS. EYES-CLOSED
Mourad Kholkhal1, Linda Bellal2, Ridha Ilyas Bendjillali2,Mohamed Sofiane Bendelhoum2, Ali Abderrazak Tadjeddine2, Kamline Miloud3
1Biomedical Engineering Laboratory, Department of Biomedical Engineering, Faculty of Technology, Abou Bekr Belkaid University, Tlemcen, Algeria
2LSETER Laboratory, Technology Institute, Nour Bachir University Center, El-Bayadh, Algeria
3Department of Electrical Engineering, Tahri Mohammed University Bechar, Algeria
Abstract: This study introduces an interpretable machine learning pipeline for classifying Eyes-Open vs. Eyes-Closed states using electroencephalography (EEG) data. The proposed method integrates multi-domain feature sets, including time, frequency, non-linear, and wavelet features, with a hybrid stacking ensemble model. We evaluate the model on three diverse datasets: PhysioNet, MNNIT, and SPIS. While the model achieved 97.4% accuracy on the two-subject MNNIT dataset, we focus on more robust results from the larger SPIS (93.02%) and PhysioNet datasets (84.3%). Feature importance analysis using SHAP reveals that coherence, alpha-band power, and non-linear entropy measures are the most influential features, supporting the clinical interpretability of the model.
Keywords: Electroencephalography (EEG); Sensory processing; Eye state recognition; Machine learning; Interpretability.
VOLUME 9 ISSUE 11 2025: 153 – 162