Calibrated Multimodal Fatigue Detection
- PyTorch
- physiological signals
- LoRA
- mixture-of-experts
What it is
A calibrated multimodal framework that estimates fatigue from three physiological signals — EEG, heart-rate variability (from ECG), and electrodermal activity — and outputs well-calibrated probabilities suitable for decision support.
The interesting engineering: it fuses modalities with Low-Rank Adaptation (LoRA) + Mixture-of-Experts, training only ~15% of the parameters of full fine-tuning while improving accuracy. It also introduces a bounded Physiological Fatigue Index, audits temporal confounding (time-on-task vs. affective-cognitive load), and applies post-hoc isotonic calibration. Evaluated with nested LOSO (58 folds) on the ASCERTAIN dataset.
Stack
PyTorch · LoRA · Mixture-of-Experts · isotonic calibration · nested LOSO · signal processing (Welch, Pan-Tompkins, cvxEDA)