Source-linked AI summary
Heart Rate Tracking using Wrist-Type Photoplethysmographic (PPG) Signals during Physical Exercise with Simultaneous Accelerometry
Mahdi Boloursaz Mashhadi, Ehsan Asadi, Mohsen Eskandari, Shahrzad Kiani, Farrokh Marvasti
TL;DR
Wrist PPG signals during intensive exercise are strongly contaminated by motion artifacts, complicating heart-rate tracking. The paper combines acceleration-based artifact cancellation with high-resolution spectral analysis and peak selection. On benchmark recordings, it reports a 1.25 BPM average absolute error and high estimation accuracy under strong motion artifacts.
Problem
Wrist PPG recorded during intensive exercise is strongly contaminated by motion artifacts, making heart-rate tracking challenging.
Source: §I · p. 1Method
The algorithm uses SVD-derived acceleration references and adaptive filtering to cancel motion artifacts, then applies IMAT for higher-resolution spectrum estimation and selects heart-rate peaks.
Source: §I · p. 2, B. Spectral Analysis · p. 3Results
1.25 BPM average absolute error was achieved on the 12 subjects’ recordings, and the authors report high robustness against strong motion artifacts.
Source: B. Results · p. 4, §IV · p. 4Takeaways & Limitations
The results support wrist-type PPG heart-rate estimation during intensive physical exercise using simultaneous acceleration data.
Source: §IV · p. 4
Abstract
from arXiv · showhide
This paper considers the problem of casual heart rate tracking during intensive physical exercise using simultaneous 2 channel photoplethysmographic (PPG) and 3 dimensional (3D) acceleration signals recorded from wrist. This is a challenging problem because the PPG signals recorded from wrist during exercise are contaminated by strong Motion Artifacts (MAs). In this work, a novel algorithm is proposed which consists of two main steps of MA Cancellation and Spectral Analysis. The MA cancellation step cleanses the MA-contaminated PPG signals utilizing the acceleration data and the spectral analysis step estimates a higher resolution spectrum of the signal and selects the spectral peaks corresponding to HR. Experimental results on datasets recorded from 12 subjects during fast running at the peak speed of 15 km/hour showed that the proposed algorithm achieves an average absolute error of 1.25 beat per minute (BPM). These experimental results also confirm that the proposed algorithm keeps high estimation accuracies even in strong MA conditions.
I. INTRODUCTION
Wrist PPG during intensive exercise is strongly contaminated by motion artifacts, making heart-rate tracking difficult. The proposed approach uses simultaneous acceleration data for motion-artifact cancellation, then estimates a higher-resolution spectrum to identify heart-rate peaks; experiments report a 1.25 BPM average absolute error.
I. INTRODUCTION
- Wrist PPG during intensive exercise faces severe motion artifacts, a setting less addressed by prior work focused on small motions or fingertip and ear recordings.The paper targets wrist monitoring during fast running at 12–15 km/hour and uses simultaneous acceleration data.Source: §I · p. 1
- The algorithm decomposes 3D acceleration with SVD to generate motion-artifact references, then uses adaptive filtering to cleanse two PPG channels before spectral analysis.The spectral step applies IMAT sparse reconstruction for a higher-resolution spectrum and selects peaks corresponding to heart rate.Source: §I · pp. 1–2
- 1.25 BPM average absolute error was achieved on benchmark datasets.The paper says using the same datasets and performance metrics as prior work facilitates comparison with state-of-the-art techniques, -.Source: §I · p. 2
- The contribution includes extracting motion-artifact references from acceleration rather than PPG, jointly processing two PPG channels, and using IMAT instead of the higher-complexity MMV model.The authors also describe the number of extracted reference signals as a parameter trading estimation error against simulation time.Source: §I · p. 2
II. THE PROPOSED ALGORITHM
The proposed heart-rate tracking algorithm has two main stages: motion-artifact cancellation and spectral analysis. Figure 1 presents the algorithm’s block diagram.
II. THE PROPOSED ALGORITHM
- The algorithm is organized into motion-artifact cancellation followed by spectral analysis.The blocks are shown in the proposed algorithm’s block diagram and explained in the following subsections.Source: §II · p. 2
- Figure 1 presents a block diagram of the proposed heart-rate tracking algorithm.Source: §II · p. 2
- The following subsections explain the blocks used in the algorithm.Source: §II · p. 2
A. MA Cancellation
Motion-artifact cancellation first filters both PPG channels to the natural heart-rate band, then suppresses in-band artifacts using acceleration-derived references and successive adaptive filters.
A. MA Cancellation
- Both PPG signals are filtered to 0.4–5 Hz to reject motion artifacts outside the natural heart-rate range.The in-band artifacts are then targeted with adaptive filters.Source: A. MA Cancellation · p. 2
- SVD decomposes the three acceleration signals to generate reference components within 0.4–5 Hz for adaptive cancellation.The method forms trajectory matrices, decomposes them into rank-one matrices, groups components with the same or harmonically related oscillations, and selects in-band groups.Source: A. MA Cancellation · p. 2
- Successive adaptive-filter stages use each prior residual and an SVD-derived reference component to remove periodic and random motion artifacts from both PPG signals.The resulting cleansed signal is passed to spectral analysis; Fig. 2 depicts cancellation by successive stages.Source: A. MA Cancellation · p. 2, Figure 2 · p. 2
B. Spectral Analysis
Spectral analysis applies IMAT to produce a higher-resolution, denoised spectrum, then selects heart-rate peaks using harmonic relationships and continuity across windows. The selection procedure handles dominant peaks, harmonic pairs, and an additional fallback based on concatenated windows.
B. Spectral Analysis
- IMAT sparse reconstruction produces a higher-resolution, denoised spectrum before peak selection.The subsequent selection mechanisms estimate the heart-rate frequency location in each time window.Source: B. Spectral Analysis · p. 3
- Peak selection uses heart-rate harmonic relationships and assumes heart rate varies smoothly without jumps between successive windows.The initial estimate uses the highest spectral peak in the first 8-second window, considered artifact-free because subjects reduce hand motions at recording onset.Source: B. Spectral Analysis · p. 3
- The method searches around the previous estimate and its second and third harmonics, using ranges defined by ε1, ε2, and ε3.Source: B. Spectral Analysis · p. 3
- When one of six peaks across the two PPG channels and three harmonic ranges is dominant, its corresponding fundamental frequency is selected as the current heart-rate index.The dominance threshold is predefined and optimized by simulations.Source: B. Spectral Analysis · p. 3
- If neither channel has a dominant peak, the method searches across both channels for a harmonically related peak pair and selects its fundamental frequency.Source: B. Spectral Analysis · p. 3
- If the earlier cases do not apply, the method joins the current and previous windows into a 10-second window and averages the available fundamental-frequency estimates.Source: B. Spectral Analysis · p. 3
A. Parameter Setting
The section specifies tuned settings for reference-signal generation, adaptive motion-artifact cancellation, spectral estimation, and peak selection.
A. Parameter Setting
- SSA uses d=100; the adaptive LMS filter uses order 25 and μ=0.005, while IMAT uses five iterations with α=0.1.These settings were selected for high-resolution, sufficiently accurate spectra of the cleansed PPG signals.Source: A. Parameter Setting · p. 3
- Peak-selection parameters T, δ, ε1, ε2, and ε3 are set to 0.6, 9, 60, 80, and 100, respectively.Source: A. Parameter Setting · p. 3
B. Results
The proposed method achieves a 1.25 BPM average absolute error across 12 subjects’ recordings. Results also describe motion-artifact removal and trade-offs involving reference-signal count and spectral estimation.
B. Results
- The Fig. 3 frequency-content example shows motion artifacts removed and the HR-associated spectral peak made more significant.Source: Figure 3 · p. 4, B. Results · p. 4
- 1.25 BPM average absolute error was achieved across 12 subjects’ recordings.Source: B. Results · p. 4
- Table 2 compares the proposed technique with state-of-the-art methods using the same datasets and performance measures.Source: Table 2 · p. 4, B. Results · p. 4
- Table 3 reports that n=100 is an optimum performance-complexity choice; reducing n can lower simulation time while increasing AAE.Directly using the three acceleration signals as references (n=3) hampers adaptive-filter convergence and degrades overall performance.Source: Table 3 · p. 4, B. Results · p. 4
- For Subject 1, IMAT significantly reduces simulation time while maintaining acceptable estimation accuracy compared with FOCUSS.Source: Table 4 · p. 4, B. Results · p. 4
IV. CONCLUSION
The conclusion presents a real-time heart-rate estimation algorithm for wrist PPG during intensive exercise, using simultaneous acceleration as the motion-artifact reference. It reports high robustness against strong motion artifacts.
IV. CONCLUSION
- The proposed algorithm targets real-time heart-rate estimation from wrist PPG during intensive exercise, using simultaneous acceleration as the motion-artifact reference.It consists of motion-artifact cancellation and spectral analysis and is reported to be robust against strong artifacts from intensive exercise.Source: §IV · p. 4