Cross-subject EEG driving fatigue detection: early personal spectral reference and causal long-sequence residual encoding

Background In cross-subject electroencephalography (EEG)-based driving-fatigue detection, most existing decoders are designed around second-scale short windows, whereas fatigue is a slow drift; short-window spectral features provide little direct evidence of that slow drift, limiting fatigue decisions. More critically, each individual’s spectral baseline level and fatigue-change pattern differ, so simply lengthening the window still leaves models struggling to predict across individuals from observed spectral features. This work aims to enable cross-subject fatigue decisions under shared parameters, without per-driver specialized models. Methods We propose early personal spectral reference with causal long-sequence residual encoding: an early personal spectral reference is estimated only from a short session-opening spectral window; subsequent spectral observations are mapped into residual space; state-conditioned temporal modeling accumulates and updates a residual state over a long causal sequence; and fatigue is read out from the terminal state. Results On leave-one-subject-out binary classification of the Multimodal Phenotyping Dataset of Driving Fatigue (MPD-DF), five-seed (2026–2030) macro-averaged F1 is 0.655 ± 0.005 (0.660 at seed 2026), outperforming the best matched-protocol absolute-spectrum baseline (≈0.587). On continuous vigilance regression of the SEED Vigilance dataset (SEED-VIG) under a subject-LOSO protocol, root mean square error (RMSE) is 0.209 ± 0.005 under 5Bands and 0.200 ± 0.003 under 2 Hz, both below matched-protocol baselines. Conclusion An early personal spectrum estimated from a short opening window can reduce interference from intrinsic individual spectral level in fatigue decisions and provides a practical premise for cross-subject fatigue decoding. Accumulating and updating residual state over a causal long sequence in residual space further improves cross-subject decisions. Under shared parameters without per-driver models, the framework outperforms matched absolute-spectrum baselines on both fatigue classification and vigilance regression.