Cross-subject drowsiness diagnosis (CDD) based on electroencephalography (EEG) is essential for mitigating drowsiness-related risks, particularly in high-stakes driving scenarios. Accurate diagnosis can effectively prevent traffic accidents. However, individual variability induces spatiotemporal isopositional value heterogeneity (SIVH) in EEG signals, which triggers a dynamic domain shift in the underlying drowsiness feature distributions, resulting in suboptimal CDD performance. In this paper, we propose MADNet, a novel CDD framework designed to enhance accuracy by strengthening feature distribution constraints and improving cross-subject generalization. To enhance the distribution constraints of drowsiness features, we propose a modality-augmented feature constraint (MAFC) module that leverages modality augmentation to generate complementary EEG-image feature pairs. Furthermore, we design a domain-adversarial feature-gating (DAFG) mechanism to extract domain-invariant drowsiness features while suppressing noise and subject-specific artifacts. In addition, we present a multi-user evaluation strategy (LTSO) designed to more faithfully reflect real-world deployment conditions. Extensive experiments show that MADNet achieves the state-of-the-art performance on Unbalanced-SADT and SADT datasets. Notably, under the rigorous LTSO, our method improves CDD accuracy by 3.08% and 1.57%, respectively, underscoring its robustness in real-world applications.
Li et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: