Event-preserving residual filtering with benefit-aware bypass for motion-corrupted photoplethysmography (PulseGuard)

PulseGuard treats motion filtering as an *action decision*: train a residual correction with waveform + clean-identity + fixed pulse-event objectives; a benefit-aware gate applies the residual only when expected to improve the segment. Reports 2.2 insertions / 2.6 deletions per 100 reference pulses vs 4.4 / 5.1 for a waveform-only residual U-Net; HR MAE 3.80 bpm (PPG-DaLiA) and 3.30 bpm (WESAD). Emphasizes that poor input quality ≠ always apply a filter.

• Form factor: wrist

Background and Objective: Motion-corrupted photoplethysmography (PPG) filtering can suppress artifacts while inserting, deleting, or shifting pulse events. Conversely, poor input quality does not imply that a specific correction will improve a segment. This work addresses filtering as an event-preserving action decision rather than an unconditional denoising map. Methods: PulseGuard trains a residual correction candidate with waveform, clean-identity, and fixed pulse-event objectives. The event representation separates cumulative event timing from total event mass so that pulse displacement and pulse insertion/deletion receive distinct signals. A benefit-aware gate, trained from cross-fitted changes in event error, applies the candidate residual only when it is expected to improve the segment. Evaluation separates controlled clean-reference event fidelity from ECG-referenced heart-rate estimation on natural wearable data. Results: PulseGuard yields 2.2 insertions and 2.6 deletions per 100 reference pulses, compared with 4.4 and 5.1 for the waveform-only residual U-Net. Heart-rate mean absolute errors are 3.80 beats/min on PPG-DaLiA and 3.30 beats/min on WESAD. Removing the event-count term increases insertion and deletion rates, while replacing benefit gating with quality gating worsens the reported event and rate errors. Conclusions: The results support combining event-preserving correction with a learned bypass, so filtering is applied according to its expected downstream benefit under the evaluated signal and reference conditions.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://www.uspto.gov/ip-policy/copyright-policy/copyright-basics

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