SE-Driven Dynamic Convolution for Adaptive EEG-Based Driver Fatigue Detection Across Spectral, Spatial, and Temporal Domains

EEG-based driver fatigue detection faces three signal-level challenges: inter-subject spectral variability, coupled frequency–spatial–temporal dynamics that existing methods process independently, and dependence on a single labeling scheme. This paper presents DCAMNet, a lightweight CNN (12.3 K parameters) that addresses these challenges through three end-to-end blocks. An SE-driven dynamic convolution block adapts spectral sensitivity per sample via input-dependent kernel weighting—applied here for the first time to fatigue detection. A spatial convolution block encodes electrode-level cortical patterns, and a temporal attention block captures fatigue dynamics through windowed variance descriptors with group-wise attention scoring. DCAMNet was evaluated on SEED-VIG (PERCLOS labels) and MESD (reaction-time labels) under both subject-mixed and leave-one-subject-out (LOSO) protocols. Under LOSO cross-validation—the operationally relevant test that eliminates within-subject information leakage and simulates deployment on unseen drivers—DCAMNet achieved 85.43% accuracy on SEED-VIG with a 2.86-point advantage over the strongest baseline, and 79±5% accuracy on MESD with a 3-point advantage. As upper-bound estimates under the subject-mixed protocol, accuracy reached 97.47% (SEED-VIG) and 96.52% (MESD). With 1.35 ms inference latency on a standard GPU, the compact architecture suggests potential suitability for real-time embedded deployment, although on-device validation on representative automotive hardware remains necessary.