Participant-Level Cross-Fit Ensemble Disagreement for Reliability-Aware Vision-Based Infant Respiratory-Rate Estimation

Contactless video-based respiratory-rate (RR) estimates may be unreliable for individual infant recordings even when overall accuracy is acceptable. We examined whether disagreement among participant-separated models could rank estimation reliability in public AIR-125/AIR-400 data (15 infants, 374 approximately 60-second recordings). For each recording, four eligible models excluded its participant group from training; their mean defined the RR estimate and their sample standard deviation defined ensemble disagreement (ED). EfficientPhys and DeepPhys were the primary architectures, with VIRENet examined as a third-architecture sensitivity analysis. ED was associated with absolute error (Spearman rho 0.742, 0.756, and 0.778, respectively). At 70% retrospective coverage, retained-recording MAE decreased from 3.64 to 2.00, 4.00 to 2.15, and 3.62 to 2.27 breaths/min, respectively, relative to all recordings. Exploratory reanalysis identified exact agreement among all four models in 28.3%-32.9% of recordings, including occasional large errors, and instability at very low coverage. The findings establish an empirical error-ranking association in this small, home-recorded infant benchmark, not calibrated uncertainty, a clinically acceptable RR threshold, or prospective monitoring performance. Independent cohorts with physiological reference measurements and deployment evaluation are needed.