Contactless atrial fibrillation (AF) screening from facial-video remote photoplethysmography (rPPG) is an attractive opportunistic-screening modality, but deployment requires the model to abstain on segments it cannot classify safely. Standard selective-prediction methods rank samples by model confidence alone, ignoring the physical signal-quality information available from the rPPG pulse waveform itself. We study post-hoc deferral policies that combine model confidence with spectral signal-to-noise ratio (SNR) in the cardiac band. We first show that a naïve linear combination achieves an apparent 18 % relative improvement in area under the risk- coverage curve (AURC) over uncertainty-only deferral on a 45,064-segment synthetic-rPPG benchmark derived from the PhysioNet/CinC 2017 AF Challenge — but this aggregate gain is clinically harmful: AF recall at 50 % coverage collapses from 0.71 to 0.04 because AF’s irregular rhythm is itself a low-SNR signature in the cardiac band, so the deferral rule systematically rejects positives. We then introduce Learned Class-Conditional Signal-Quality Deferral (LW-CCSD), a constrained-grid-search procedure that estimates per- predicted-class quality weights on validation data subject to a configurable AF-recall floor. LW-CCSD makes the safety / accuracy trade-off explicit and tunable: across the Pareto frontier we measure +4.1 % AURC improvement at ≤ 0.01 AF-recall cost, scaling to +15.6 % AURC at 8 percentage-point AF cost. The method generalises across three uncertainty quantification methods (deterministic single-pass softmax, MC Dropout, deep ensembles). Strikingly, LW-CCSD applied to a single-forward-pass 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.

If you use this software in academic work, please cite it as below.