An Agent Architecture with Tools and Memory for Selective Correction of Wrist PPG Heart Rate Estimates
Wrist PPG heart rate monitoring is vulnerable to motion artifacts, and naively applying corrections can lower average error while damaging individual windows. This paper proposes a selective correction framework organized as an agent around a fixed core estimator: signal-analysis modules extract pulse and motion evidence, a temporal memory module generates candidate heart rates, and a decision module predicts each candidate's benefit and harm before a separate risk gate chooses to retain, correct, or reject the output. The authors train the auxiliary modules on PPG-DaLiA and evaluate frozen weights on three additional public datasets plus a prototype wearable stream, reporting macro-averaged MAE reductions up to 44.5%. The core idea—explicitly trading off predicted gains against harms rather than universally replacing estimates—is a clear contribution to wearable heart rate analysis, though deployment readiness is weakened by calibration transfer failures and an extremely limited prototype.
The framework is a solid methodological contribution with rigorous internal structure and honest harm reporting, but its practical impact is tempered by failed threshold transfer and a vanishingly small prototype study. The authors correctly note that their system is a fixed pipeline rather than an autonomous agent, and they forthrightly report that source-calibrated coverage thresholds collapsed on external data. The large macro-averaged improvements across ten estimators are real in the evaluated setting, yet the near-zero gains for BeliefPPG and the missing code release make the work feel more like a well-engineered proof of concept than a deployable system.
The explicit modeling and reporting of correction harm is the strongest feature. The authors decompose total error reduction into numerical correction versus window selection, showing that correction on the same accepted windows accounts for most of the gain at 80% coverage. They also report beneficial and harmful correction frequencies separately, and the ablation study supports genuine contributions from motion evidence and temporal memory. The geometric constraints in M4 ensure internally consistent predictions of gain, harm, and absolute error, which is a thoughtful design choice.
The most serious flaw is that fixed source-domain thresholds fail to transfer: the nominal 80% coverage rule yielded only 42.29%, 45.48%, and 30.50% coverage on Wrist, Galaxy, and Wild, while the 5 bpm MAE target was exceeded in up to 91.45% of nonempty configurations. This means the system requires recalibration per deployment setting. The prototype evaluation is limited to four self-recordings from the first author without institutional ethics approval; the authors themselves state these do not establish generalization or clinical validity. Additionally, the agent framing may mislead readers because the system uses no language model, dynamic tool scheduling, or multi-estimator voting, and the authors admit all datasets had been accessed during earlier development, undermining independence claims.
The evidence supports average MAE reduction but shows highly uneven gains across estimators. BeliefPPG exhibited essentially zero change, indicating the auxiliary modules add little when the core estimator already captures similar spectral and uncertainty information. The ridge rejection baseline sometimes outperformed the framework's own risk ranking on the prototype, so the risk score is not universally dominant. The comparisons to same-risk rejection fairly isolate the value of numerical correction versus selection, and the authors appropriately caution that they measure auxiliary gains around fixed implementations rather than provide a fair ranking of the original published methods.
Independent reproduction is blocked by the absence of released code and weights. The authors state that public repository links are not provided in this version, and while hyperparameters and module definitions are documented in the supplementary material, the lack of accessible code makes independent verification impractical. Some core estimators rely on author-provided weights with incomplete training-participant provenance, and fixed-threshold transfer experiments show that source configurations cannot be reused without recalibration on the target domain. Participant-bootstrap intervals also exclude the uncertainty of full model retraining.
Wrist photoplethysmography (PPG) heart rate estimates are vulnerable to motion artifacts, and corrections that reduce mean error can worsen individual windows. We propose an agent architecture with tools and memory around a fixed heart rate estimator. Pulse, motion, and limited historical evidence generate candidate estimates, whose predicted benefits and harms guide selection. An independent risk gate determines whether to retain, correct, or reject each output. We trained the architecture on PPG-DaLiA and tested it on Wrist Exercise, GalaxyPPG, and WildPPG with frozen weights. Across these four public datasets, mean absolute error (MAE), macro-averaged over ten estimators, decreased by up to 44.5%, with the largest reduction on GalaxyPPG. Beneficial and harmful corrections decreased and increased absolute error by more than 1 bpm, respectively. Their ten-estimator macro-averaged proportions ranged from 16.06% to 36.64% and from 1.68% to 6.77%, respectively, across the four datasets. At 60% and 80% target coverage, numerical correction on the same accepted windows contributed more to total error reduction than window selection. We implemented real-time wearable acquisition, Bluetooth reception, and smartphone deployment of a selected estimator and the framework. In backend analysis of four prototype recordings from one participant, ten-estimator macro-averaged MAE decreased from 9.09 to 4.45 bpm without additional rejection. Under the evaluated conditions, the architecture improved average accuracy while quantifying correction benefits and harms, supporting wearable heart rate analysis that considers error, correction risk, and coverage.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.