Source-linked AI summary
Robust PPG-Exclusive Heart Rate Estimation During Maximal Exercise Using Population-Based Priors
Elisabetta De Giovanni, Iñigo Urteaga, Matteo Croci, Tomas Teijeiro
TL;DR
Motion artifacts can undermine PPG heart-rate estimates during intense exercise, so this study tests whether population-based HR priors can support PPG-only estimation. It combines PPG spectra with priors learned from other subjects and reports a 63.3% median MAE reduction versus a PPG+ACC benchmark.
Problem
Motion artifacts during intense activity compromise PPG HR-estimation reliability, motivating the question of whether heavily contaminated PPG can yield reliable estimates without auxiliary sensors.
Source: §1 · p. 1Method
PPG-Prior combines the PPG-derived PSD with a population-based HR prior to estimate a posterior and select an HR trajectory.
Source: §2 · p. 2, §2.3 · p. 2Results
63.3% median MAE reduction versus BeliefPPG with PPG+ACC was reported for PPG-Prior on maximal exercise tests.
Source: §3.1 · p. 3, §5 · p. 4Takeaways & Limitations
The results highlight that exercise-specific prior knowledge can reduce motion-artifact effects without auxiliary motion sensors.
Source: §5 · p. 4Validation is restricted to incremental maximal and submaximal exercise protocols.
Source: §4 · p. 4
Abstract
from arXiv · showhide
PPG is the main technology for wearable heart rate (HR) monitoring. However, motion artifacts during intense physical activity reduce its reliability due to spectral overlap with the cardiac signal. This study investigates whether severely contaminated PPG signals can yield reliable HR estimates without auxiliary sensors by leveraging prior knowledge of HR dynamics. We acquired simultaneous PPG and ECG signals from 21 subjects during a maximal cycloergometer test. Using Leave-One-Subject-Out cross-validation, we built HR evolution priors from the ground-truth ECG and applied Dynamic Time Warping (DTW) to align normalized HR curves and estimate a barycenter average. DTW aligns normalized HR curves with minimal error across all subjects, validating robust population-based priors. For each test subject, we combined the PPG Power Spectral Density (PSD) with the prior distribution to generate a "posterior PSD". We identified and ranked candidate HR trajectories using heuristic search within the "posterior PSD" and selected the most likely path. Our approach reduced mean absolute error (MAE) in BPM relative to the ground-truth ECG by a median of 63.3% (IQR: 47.3%-74.6%) compared to a PPG+accelerometer benchmark. These results show that physiological priors and DTW-based alignment enable robust, PPG-exclusive HR monitoring during high-intensity exercise.
1. Introduction
Intense-exercise motion artifacts can compromise PPG heart-rate estimates, motivating investigation of whether heavily contaminated PPG can still support reliable estimates without auxiliary sensors. This study addresses that question using population-based HR-dynamics priors.
1. Introduction
- Motion artifacts during intense activity can corrupt PPG cardiac signals and compromise HR-estimation reliability.PPG remains a key wearable technology because of its low cost and user comfort.Source: §1 · p. 1
- Many existing approaches rely on accelerometer data to remove motion artifacts, despite work spanning classical, machine-learning, and hybrid methods.The cited benchmarks report that deep-learning methods consistently outperform classical ones, with BeliefPPG reported as the most robust.Source: §1 · p. 1
- The study presents a PPG-exclusive method using population-based priors for high-intensity exercise and compares it with BeliefPPG on maximal and submaximal tests.The authors also state that they examine the approach’s applicability and limitations.Source: §1 · p. 1
2. PPG-exclusive HR estimation
PPG-Prior estimates HR by combining a time-varying PPG spectrum with a population-derived HR prior, then selecting a plausible trajectory from the resulting posterior. The prior is built from other subjects’ HR dynamics, with soft-DTW alignment used to form a representative trend.
2. PPG-exclusive HR estimation
- PPG-Prior combines the PPG power spectral density with a population-based prior to form a posterior and select an HR trajectory.Figure 1 depicts the PPG spectrum, prior, and posterior with selected trajectory for an example subject whose PPG is highly corrupted.Source: §2 · pp. 1–2, Figure 1 · p. 2
2.1. Power spectral density over time
- The method bandpass-filters and detrends PPG, then estimates its online PSD in overlapping 8-second windows and restricts frequencies to 40–220 BPM.The PSD can show transient bursts and periodic noise traces associated with motion artifacts.Source: §2.1 · p. 2
2.2. Population-based prior
- A leave-one-subject-out prior uses smoothed, peak-normalized ECG-derived HR series, soft-DTW alignment, and a multivariate kernel-density estimate.The aligned series yield a barycenter trend, while the density estimate supplies a prior distribution on the PSD’s normalized-frequency and time grid.Source: §2.2 · p. 2
2.3. Posterior HR estimation
- The posterior combines PSD and prior; iterative Viterbi search extracts physiologically constrained, diverse candidate paths, selected by cumulative score or distance to the soft-DTW barycenter.Min-DTW is intended for cases where the posterior remains highly corrupted and greater reliance on the prior is needed.Source: §2.3 · pp. 2–3
3. Experimental setup and results
The evaluation uses ECG-referenced PPG-Prior estimates from maximal and submaximal cycloergometer tests, comparing them with BeliefPPG. Results favor PPG-Prior in maximal exercise, while submaximal performance varies with the prior and test protocol.
3. Experimental setup and results
- The MAX-HIE evaluation used 21 subjects and leave-one-subject-out priors; SUBMAX-HIE included five volunteers and priors from MAX-HIE or ACTES.Second-by-second MAE was computed against ECG and compared with BeliefPPG, which reports estimates every two seconds.Source: §3 · p. 3
3.1. Maximal exercise results
- 63.3% median MAE drop, up to 85.4%, versus BeliefPPG with PPG+ACC was reported for PPG-Prior on MAX-HIE.Table 1 summarizes the maximal-exercise results.Source: Table 1 · p. 3, §3.1 · p. 3
- Top-Traj produced MAE below 3 BPM in 28.6% of cases and below 5 BPM in 38.1%, while BeliefPPG never fell below 3 BPM and reached below 5 BPM in 14.3%.The paper reports BeliefPPG with PPG-only but excludes a direct comparison because it consistently underperformed PPG+ACC.Source: §3.1 · p. 3
- For example MAX-HIE subjects, Min-DTW improved MAE by up to 5.5 BPM in approximately 33% of the dataset; high MAE cases included gaps in the original PSD observations.The examples include Sub7, whose highly corrupted PSD still yielded the best outcome, and Sub21, where the reported MAE drop was 62.3%.Source: §3.1 · p. 3, Table 2 · p. 3
3.2. Submaximal exercise results
- On SUBMAX-HIE, MAX-HIE priors lowered MAE by up to 49.3% in three of five cases, whereas ACTES priors lowered it by up to 91.2%; PPG-Prior had higher errors in two cases.BeliefPPG had MAE below 6 BPM in four of five cases; the table covers healthy volunteers and elite basketball players using different priors.Source: §3.2 · p. 4, Table 3 · p. 4
4. Limitations and future work
The method’s validation and learned priors are limited to available datasets and incremental exercise protocols. The authors identify adaptive candidate selection, broader protocol testing, physiological trend modeling, and personalized priors as future directions.
4. Limitations and future work
- Validation is restricted to incremental maximal and submaximal protocols, leaving non-incremental tasks and real-world sports training for future evaluation.Prior learning is limited to available datasets; the authors also identify adaptive segment-level path selection, trend forecasting, and personalized priors as future work.Source: §4 · p. 4
5. Conclusion
The PPG-exclusive method uses population-based priors and was validated on maximal and submaximal exercise tests. It achieved a 63.3% median MAE reduction versus a state-of-the-art method requiring ACC data.
5. Conclusion
- 63.3% median MAE reduction versus a state-of-the-art method requiring ACC data.The method was evaluated for PPG-exclusive HR estimation using population-based priors.Source: §5 · p. 4
- The population-prior method was validated on maximal and submaximal exercise tests.Source: §5 · p. 4
- Future work will integrate physiological trend forecasting and adaptive path selection for broader, non-incremental exercise contexts.Source: §5 · p. 4