NTSM: Neurophysiology-Guided Tri-Stream Selective State Space Model for Real-Time Driver Fatigue Detection with EEG-EOG Signals

Driver fatigue is a leading contributor to fatal traffic accidents globally, and Electroencephalogram (EEG)-Electrooculogram (EOG) multi-modal systems are the gold standard for objective fatigue detection. However, existing Transformer-based methods face critical bottlenecks of high quadratic complexity and oversimplified feature fusion that neglects intrinsic brain-eye neurophysiological couplings. To address these issues, this paper proposes NTSM, a neurophysiology-guided tri-stream architecture integrating the Mamba-based selective State Space Model (SSM) with four specialized modules, i.e., Functional Connectivity Bridge (FC-Bridge), Spectral Recalibration Unit (SRU), Ocular-Guided Attention (OGA), Artifact Disentanglement Gate (AD-Gate). Equipped with a bidirectional Mamba encoder, NTSM achieves linear complexity (reducing computational cost to $5\%$ of Transformer) while preserving cross-modal interaction performance. Evaluations on the SEED-VIG dataset via Leave-One-Subject-Out (LOSO) validation yield a Pearson Correlation Coefficient (COR) of $\mathbf {0.8738} \pm 0.10$ and Root Mean Square Error (RMSE) of $\mathbf {0.1294} \pm 0.06$, delivering $\mathbf {8.19}\times$ inference acceleration over the state-of-the-art Transformer-based EEG-EOG Cross-modal Fusion (E2CF). Furthermore, cross-dataset validation on the clinical MPDDF dataset achieved an exceptional accuracy of $92.95\%$, proving its robust generalization capability. Gradient-weighted saliency analysis confirms its neuroscientifically plausibility, with significant parieto-occipital activation during alertness and frontal-temporal slow-wave dominance during fatigue.