Beyond Average Accuracy: Motion-Related Failure Patterns in Wrist-Worn Photoplethysmographic Heart-Rate Monitoring

Abstract

Wrist-worn photoplethysmography (PPG) enables unobtrusive heart rate monitoring, but motion-related errors tend to cluster under specific conditions and are often masked by aggregate accuracy metrics. This study characterized those failure patterns and assessed whether wrist-sensor features could identify unreliable estimates in previously unseen participants. PPG-DaLiA recordings from 15 participants, acquired with an Empatica E4 and synchronized chest electrocardiography during daily-life activities, were segmented into 8-s windows with a 2-s step. Heart rate was estimated using a transparent dominant-frequency PPG method. A large error was defined a priori as an absolute deviation of more than 10 beats per minute (bpm) relative to the ECG-derived reference. Motion intensity, acceleration-band power, PPG-acceleration spectral overlap, peak dominance and spectral entropy were used as reliability predictors. Of 47,182 non-transition windows, 14,622 (31.0%) met the large-error criterion. Failure prevalence differed substantially by activity, from 5.8% during sitting to 63.8% during stair climbing. In leave-one-subject-out validation, a motion-only model achieved an AUROC of 0.743, wrist-signal features increased AUROC to 0.851, and activity context increased it to 0.864. Wrist PPG reliability is therefore condition dependent, and an interpretable signal-based reliability flag may complement point heart-rate estimates. External prospective validation is required before clinical or consumer deployment.