A study introduces DrowsyFormer-Gen, a model consisting of three domain-specific transformer encoders, sequential fusion and an adaptive classification head, with a class-conditional denoising diffusion probabilistic model (DDPM) providing additional windows of the minority class in training. The research defines and assesses a multimodal framework with a common windowing and subject-disjoint evaluation protocol for five classical baselines and the neural variants, includes a direct assessment of synthetic-window quality, and reports component ablations and calibration measurements. The introduction notes that statistics of 461,312 road accidents and 168,491 fatalities in 2022 are not estimates of fatigue-related accidents, and that a useful evaluation should separate overall classification accuracy from sensitivity to rare positive events, the reliability of predicted probabilities and the timing of alarms. The authors state that the benefit of the components on safety relevant metrics can be empirically determined, not inferred from architectural complexity, and that the conditioned generation of ECGs using structured state-space diffusion and Diffusion-TS suggests the fidelity of generated drive windows should be empirically assessed.