DrowsyFormer-Gen: classical baselines match a triple-stream transformer with diffusion augmentation for driver drowsiness detection

Driver drowsiness detectors must be judged on minority-class sensitivity, calibration and alarm behaviour, not accuracy alone. DrowsyFormer-Gen combines physiological, vehicle and behavioural transformer streams with cross-modal fusion, an adaptive classification head and class-conditional diffusion augmentation. We re-implemented and evaluated it on shared causal windows and subject-disjoint folds against five classical baselines, using DMPS (225,000 samples; 15 drivers) and SC-DD (87,430 segments; 20 participants); DMPS drivers are a subset of SC-DD participants. Baselines received nested hyperparameter optimisation; thresholds were fixed on training data. Tuned logistic regression reached 99.747 ± 0.031% accuracy, 0.953 drowsy-class recall and expected calibration error 0.0008, ahead of every neural configuration. Each proposed component failed a controlled test: removing diffusion augmentation improved every metric, and class weighting alone reached 0.989 recall; zeroing the adaptive head’s reference halved calibration error while its refresh interval was inert; and the fusion ablation, over 25 runs, ranged from 0.000 to 1.000 recall at a fixed threshold despite stable ranking. A single sample reached 0.915 recall against 0.956 for the 32-sample window. We report a negative result, conditional on the supplied records being what the source publications describe.