Adding accelerometry raises task-conditioned recoverability in wearable photoplethysmography

Separates model limitation from measurement limitation via task-conditioned recoverability on Brno University of Technology Smartphone PPG Database v2.0.0 (82,966 balanced ECG-QRS event/non-event samples; 3,840 recordings; 38 subjects). Compact ML zoo (logistic regression, hist gradient boosting) under subject-held-out evaluation; shows adding accelerometry increases recoverable information for the inference task.

• Form factor: other

Wearable photoplethysmography (PPG) is easy to collect, but its downstream value depends on whether the recorded signal preserves the distinctions required by the inference task. This study tests whether task-conditioned recoverability separates model limitation from measurement limitation and whether an added physical sensor increases recoverable information. The analysis uses the Brno University of Technology Smartphone PPG Database v2.0.0 and contains 82,966 balanced ECG-QRS event/non-event samples from 3,840 recordings and 38 subjects. A compact machine-learning model zoo comprising logistic regression and histogram gradient boosting establishes demonstrated subject-held-out performance; cross-subject nearest-neighbor geometry over prespecified representations supplies a comparative recoverability estimate. PPG alone yields balanced accuracy of 0.5018 and recoverability of 0.5454. Expert-rated good recordings yield higher recoverability than expert-rated poor recordings (0.5706 versus 0.5488), with the same ordering in all eight perturbation conditions. On the same ACC-capable records, adding accelerometry raises recoverability from 0.5520 to 0.6148 and balanced accuracy from 0.5007 to 0.5957. The recoverability quantity is comparative and task-conditioned. The results show that changing the physical measurement can change both the performance an observer reaches and the information available for the declared task.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0/

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.