Pseudo-label-assisted subdomain adaptation network with coordinate attention for EEG-based driver drowsiness detection

• Published in: Biomedical signal processing and control 2025-03, Vol.101, p.107132, Article 107132
• Main Authors: Feng, Xiao, Dai, Shaosheng, Guo, Zhongyuan
• Format: Article
• Language: English
• Subjects: Coordinate attention; Drowsiness recognition; EEG; Pseudo labeling; Separable convolution; Subdomain adaptation
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.107132

•We propose the first EEG feature extraction network framework with a coordinate attention mechanism, to capture spatial long-range dependencies and channel-wise relationships to efficiently augment the EEG feature representations.•We propose a subdomain adaptation method with a LMMD loss and dynamic weighted pseudo-label learning strategy, which can enhance cross-domain generalization capability for new target subjects.•We propose the interpretable convolutional coordinate attention network framework, giving an interpretable classification decision-making via the coordinate attention maps instead of black-box results.•Our proposed model is evaluated on two publicly available driver drowsiness dataset with remarkable classification performance.•Visualizing the learned attention map and feature distribution provides interpretable insights for model validation. Accurate detection of driver drowsiness using Electroencephalography (EEG) is crucial for reducing traffic accidents. Although recent deep learning-based approaches have shown promising results, two significant challenges still exist: how to explicitly model EEG features interdependencies and augment representations learning for better classification and interpretation; how to improve generalization performance on the calibration-free EEG system for new subjects in practice. To address these issues, we propose a pseudo-label-assisted subdomain adaptation network with coordinate attention (PASAN-CA) for EEG-based driver drowsiness detection. In our method, the feature extractor employs coordinate attention to enhance the discriminative features representation by effectively modeling long-range dependencies between features. For subdomain adaptation, a local maximum mean discrepancy (LMMD) is constructed to align the distribution of relevant subdomains with the same class between source and target domains, so as to learn domain-invariant discriminative features. In addition, a curriculum pseudo labeling (CPL) strategy is introduced to dynamically pick up high-quality pseudo labels of target subdomain for model training, assisting subdomain adaptation. Extensive experiments on two publicly available driver drowsiness datasets demonstrate that the proposed framework outperforms state-of-the-art baselines in overall performance. Moreover, visualizing the learned attention map and feature distribution provides interpretable insights for model validation.