A Lightweight Hybrid Temporal–Spectral Model with Gated Fusion for EEG-Based Driver Drowsiness Detection

Abstract

Driver drowsiness is a major contributor to road accidents worldwide. Electroencephalography (EEG) enables direct measurement of neural correlates of cognitive fatigue, allowing early detection before behavioral symptoms manifest. This paper proposes a lightweight hybrid temporal–spectral EEG model with gated adaptive fusion for cross subject driver drowsiness detection using a single frontal channel (Fp1). The temporal branch employs a two-layer 1-D CNN (kernel sizes 7 and 5, 32 filters each) with windowed multi-head self-attention (4 heads, window size 32) to capture local oscillatory dynamics. The spectral branch extracts eleven physiologically grounded features including absolute and relative band powers, drowsiness index ratios, and spectral entropy from three sub window Welch estimates. A sigmoid gated fusion module learns per sample weighting between both representations. Evaluated under Leave-One-Subject-Out (LOSO) cross-validation on the SEED-VIG dataset, the model achieves 82.37% mean accuracy, F1 is 0.8316, sensitivity is 0.8411, specificity 0.8063, and AUC-ROC is 0.922, with only 14,499 trainable parameters and 1.063 ms inference latency. Ablation across four model variants confirms that each component contributes measurably to overall performance. A multichannel comparison (17 channels, 90.35% accuracy) quantifies the single-channel trade-off. Quantitative explainability analysis demonstrates statistically significant gate class discriminability (t = −11.23, p < 0.0001). The improvement over a temporal-only baseline is statistically significant (Wilcoxon p = 0.0161).