Machine Vision and Applications Computer Science Applications

An interpretable multi-channel CNN for cross-subject EEG-based driver drowsiness detection

Lim, G Wei, Pang, Yee Yong, Ismail, Nor Azman, Ghazali, Masitah, Sim, Hiew Moi, Fong, Cheng Weng
Springer Nature · 2026-11-16 · DOI 10.1007/s00138-026-01960-4

Abstract

Drowsiness is one of the major factors contributing to traffic accidents worldwide. Electroencephalography (EEG) is a useful physiological signal for drowsiness detection because it directly reflects neurophysiological activity in the brain. However, many EEG-based deep learning models rely on single-channel EEG or provide limited interpretability, which can reduce cross-subject generalization and trust in safety-critical applications. This study proposes a lightweight and interpretable multi-channel SE-CAM-CNN for cross-subject EEG-based drowsiness detection using four electrodes: Oz, Fz, Pz, and Cz. The model integrates electrode-level Squeeze-and-Excitation (SE) for input-channel recalibration, feature-level SE for feature-map refinement, and Class Activation Mapping (CAM) for temporal localization. Under leave-one-subject-out cross-validation repeated across ten random seeds, the proposed model achieved an average accuracy of 0.7679 and an F1-score of 0.7668, outperforming the evaluated single-channel baselines and maintaining competitive performance compared with a plain multi-channel CNN without SE modules. Interpretability analysis showed that Pz and Oz received relatively higher electrode-level attention weights. CAM-based spectral validation further showed that high-CAM windows had significantly higher theta and alpha power than low-CAM windows, suggesting that the model emphasized physiologically meaningful drowsiness-related EEG activity. These findings indicate that SE-CAM-CNN provides a compact and interpretable framework for cross-subject EEG-based drowsiness detection.

Read at publisher →