MSANet: a multi-scale spatiotemporal adaptive network for EEG-based driver fatigue recognition.

Electroencephalography (EEG)-based driver fatigue recognition remains challenging because fatigue-related patterns are distributed across multiple temporal scales, inter-channel relationships vary across samples, and inter-individual differences complicate cross-subject transfer. To address these issues, a multi-scale spatiotemporal adaptive network (MSANet) is proposed for EEG-based driver fatigue recognition. The framework introduces a multi-granularity patch temporal gate (MGPTG) to model locally continuous temporal contexts, stable graph node enhancement (SGNE) to enhance channel-node features before dynamic graph construction, and an adaptive classifier gate (ACG) to combine cosine-based and affine decision geometries. MSANet was evaluated on two public simulated-driving EEG datasets using a cross-subject transfer protocol, in which the target-subject samples were stratified with a partition ratio of 0.5 into a labeled adaptation subset and a held-out test subset. MSANet achieved an accuracy (ACC) of 83.64% and an F1-score (F1) of 83.74% on Dataset I, and 74.64% ACC and 72.92% F1 on Dataset II. The results indicate strong numerical performance on Dataset I and competitive performance on Dataset II. Subject-level statistical analysis was additionally conducted to distinguish mean performance differences from differences that are consistently supported across subjects. Ablation results showed lower mean performance after removing MGPTG, SGNE, or ACG under the evaluated settings. These findings support further evaluation of the proposed stage-specific temporal, graph-node, and classification mechanisms for cross-subject transfer in EEG-based driver fatigue recognition.