Confidence Does Not Travel: Cross-Dataset Validation of Selective AI for Photoplethysmographic Respiratory-Rate Monitoring Before Breathing-Centered Rehabilitation Deployment

This is the authors' abstract. We don't add key points for this paper.

Background/Objectives: Photoplethysmography (PPG) could support unobtrusive respiratory-rate monitoring during breathing-centered rehabilitation. Accuracy measured in one cohort, however, says little about how the same estimator will behave after deployment elsewhere. Confidence estimates may also fail. We tested whether physiology-informed artificial intelligence and a learned error-prediction gate retained their performance across two clinically distinct public datasets. Methods: We conducted a retrospective, subject-separated benchmark using 53 BIDMC intensive-care recordings and 42 CapnoBase anesthesia recordings. From 32 s PPG windows advanced every 8 s, we derived four respiratory modulations and 51 rate, agreement, morphology, and signal-quality features. Six estimators were evaluated by five-fold grouped cross-validation and by bidirectional external transfer. A nested gradient-boosting model predicted each estimate's absolute error to support selective reporting. Performance was summarized with subject-bootstrap confidence intervals, Bland–Altman agreement, and risk–coverage analysis. Results: The analysis included 5262 windows from 95 participants. Within-dataset mean absolute error (MAE) was lowest for random forest on BIDMC (1.65 breaths/min) and Extra Trees on CapnoBase (3.55 breaths/min). The physiology-consensus model achieved 1.70 and 4.02 breaths/min, respectively. Its learned gate reduced MAE at 50% coverage to 0.93 and 2.01 breaths/min. Under external transfer, however, its MAE increased to 6.19 breaths/min for BIDMC→CapnoBase and 4.90 breaths/min for CapnoBase→BIDMC. Uncertainty–error Spearman correlation fell from 0.48 and 0.41 internally to 0.15 and 0.09 externally, and gating yielded only modest improvement. A fixed raw-spectral baseline outperformed the learned model in both transfer directions. Conclusions: Confidence measured within the development cohort did not establish safety after transfer. PPG-derived respiratory rate needs external calibration before it can guide rehabilitation feedback, even when point accuracy appears acceptable internally. Public benchmarks should report bidirectional transfer and risk–coverage behavior alongside random or subject-wise internal splits.