ReliaGate: Reliability Routing for Low-Stakes Wearable Stress Prediction

Summary

It locks an upstream classifier (per-modality TCNs on BVP/EDA/TEMP, or BVP/EDA for UBFC-Phys), then trains a correctness scorer on ~90 reliability features: confidence, per-modality signal quality, cross-modality agreement, train- standardized atypicality, and train-fitted embedding geometry (Mahalanobis / k-NN distances). It evaluates on four datasets (WESAD, UBFC-Phys, EmpaticaE4Stress, PhysioNetE4) with subject-disjoint folds and reports not just coverage/risk but a per-subject "Reach" metric — explicitly because a selective system can look better in pooled numbers while starving specific people of any output.