Learned Class-Conditional Signal- Quality Deferral for Selective rPPG- Based Atrial Fibrillation Screening

Selective-prediction methodology for contactless atrial- fibrillation (AF) screening from remote-photoplethysmography (rPPG) signals derived from face video. Introduces Learned Class-Conditional Signal-Quality Deferral (LW-CCSD), a post-hoc model-agnostic deferral policy that learns per-predicted-class weights combining model confidence with spectral signal-to-noise ratio, subject to a configurable per-class recall floor. On a 45,064-segment synthetic-rPPG benchmark derived from the PhysioNet/CinC 2017 AF Challenge, LW-CCSD achieves a tunable Pareto frontier between selective accuracy and AF-recall safety (+4.1 percent AURC at essentially zero AF-recall cost, scaling to +15.6 percent AURC at 8 percentage-point AF cost). A Clopper- Pearson conformal extension provides finite-sample distribution- free coverage guarantees with essentially zero operational penalty (identical operating point at 90 percent confidence; 0.6 percent relative AURC cost at 95 percent confidence). LW-CCSD applied to a single-pass deterministic classifier achieves better selective performance than a five-model deep ensemble without LW-CCSD, suggesting the method can substitute for compute-expensive ensembling in deployed clinical screens. Cross-UQ replication (deterministic, MC Dropout, deep ensembles) and orthogonal SNR-tertile and HR-tertile stratifications confirm the cross-regime mechanism. Repository includes the full reproducibility chain: pipeline, configurations, trained checkpoints, per-method evaluation artefacts, paper LaTeX and HTML sources, and figure generation scripts.