Interpretable Reliability Control for Wearable PPG Heart Rate Estimation

Abstract

Wearable photoplethysmography (PPG) can return a plausible heart rate (HR) after motion or poor optical coupling has made the measurement unreliable. We treat HR reporting as a reliability-control problem and introduce an interpretable signal quality index (SQI) layer that reports coverage and residual error. On PPG-DaLiA, electrocardiography (ECG) supplied offline targets only; inference used PPG and accelerometer (ACC) features. The Cascade Veto defined three reliability states, while a HistGradientBoosting (HGB) regressor ranked continuous error risk. In 15-fold leave-one-subject-out testing, random forest classification attained a mean macro-F1 of 0.8995. Its static High-only policy achieved 1.479 bpm mean absolute error (MAE) with subject-adaptive coverage averaging 19.95%. Strict and permissive label configurations shifted High-only coverage to 16.54% and 22.79%, with MAEs of 1.257 and 1.666 bpm. At matched 10%, 20%, and 40% coverage, HGB achieved 1.006, 1.428, and 2.699 bpm, compared with 5.464, 5.605, and 7.043 bpm for ACC-energy ranking. Within activities at 20% coverage, HGB achieved 3.957 bpm versus 11.537 bpm for ACC and 16.714 bpm for random selection. Accepted windows clustered over time: the median subject-level maximum HGB blackout was 1356, 684, and 312 s. On an independent device, HGB trained on PPG-DaLiA reduced MAE by 61.2% relative to the full-coverage baseline at 10% coverage without external refitting. HGB provided the lower fixed-coverage MAE, whereas Cascade retained interpretable release states. The matched controls indicate that PPG/ACC features contain reliability information across subjects, activities, and the independent device.