Source-linked AI summary
Online Heart Rate Prediction using Acceleration from a Wrist Worn Wearable
Ryan McConville, Gareth Archer, Ian Craddock, Herman ter Horst, Robert Piechocki, James Pope, Raul Santos-Rodriguez
TL;DR
Because PPG heart-rate sensing uses far more power than accelerometers and the heart-rate–acceleration relationship can change over time, this paper studies online prediction from wrist acceleration. Its active-learning approach requests PPG measurements when needed; across three patients, it reports MAE between 3 and 4 while requesting heart rate around 15%–20% of the time.
Problem
PPG heart-rate sensors use substantially more power than accelerometers, while factors outside acceleration and changing heart-rate relationships complicate prediction over time.
Source: §1 · p. 1Method
An online active-learning approach predicts heart rate from streaming acceleration and requests PPG measurements when prediction uncertainty warrants them.
Source: §3.2 · p. 3, §3.3 · p. 3, §4 · p. 4Results
MAE was between 3 and 4 across the tested uncertainty settings, with heart rate requested around 15%–20% of the time and significant improvement over the baseline.
Source: §3.5 · p. 4Takeaways & Limitations
The approach substantially reduces heart-rate sensor use with only a small performance decrease, supporting the study’s goal of energy savings while tracking changes over time.
Source: §3.5 · p. 4, §4 · p. 4The evaluation covered three patients with approximately four weeks of free-living data collected over two phases months apart.
Source: §4 · p. 4
Abstract
from arXiv · showhide
In this paper we study the prediction of heart rate from acceleration using a wrist worn wearable. Although existing photoplethysmography (PPG) heart rate sensors provide reliable measurements, they use considerably more energy than accelerometers and have a major impact on battery life of wearable devices. By using energy-efficient accelerometers to predict heart rate, significant energy savings can be made. Further, we are interested in understanding patient recovery after a heart rate intervention, where we expect a variation in heart rate over time. Therefore, we propose an online approach to tackle the concept as time passes. We evaluate the methods on approximately 4 weeks of free living data from three patients over a number of months. We show that our approach can achieve good predictive performance (e.g., 2.89 Mean Absolute Error) while using the PPG heart rate sensor infrequently (e.g., 20.25% of the samples).
1 INTRODUCTION
Wrist wearables can use low-power acceleration sensing, while heart-rate sensors consume substantially more energy. The paper motivates predicting heart rate from acceleration despite factors and changes over time that acceleration may not capture.
1 INTRODUCTION
- Accelerometers are widely used for wrist-worn activity detection, such as step counting, because they have low cost and power usage.Source: §1 · p. 1
- Health watches provide measurements such as activity, location, respiration, and heart rate for consumer and clinical applications.Clinicians use these devices to understand patient behavior over longer periods than a single clinical measurement.Source: §1 · p. 1
- PPG heart-rate sensors can consume up to 5000 times the power of wearable accelerometers, impeding the long battery life desired for wearables.The paper identifies this energy gap as motivation for predicting heart rate from acceleration.Source: §1 · p. 1
- Heart-rate prediction from acceleration is difficult because some influences on heart rate are not captured by acceleration, and their relationship may change over time.The paper gives aging and changes in fitness as examples, and argues that a non-adaptive approach may predict poorly.Source: §1 · p. 1
2 SCENARIO
The scenario pairs home-based wrist acceleration data with PPG heart-rate measurements across patients recovering from heart-valve interventions. The proposed gateway approach compares offline regression with online active learning that can request PPG measurements.
2 SCENARIO
- The EurValve project recruited 42 patients for phased deployment of a Smart Home in a Box that receives wrist accelerometer data through four gateways.The gateways may also support local data processing and machine learning.Source: §2 · p. 1
- Patients wear an accelerometer-only wearable with about three weeks of battery life and a PPG-capable health watch with up to four days.Wearing both devices is intended to provide up to six weeks of free-living data per patient.Source: §2 · p. 2
2.1 Desiderata
- The baseline trains a regression model on aligned acceleration and heart-rate features, while the proposed gateway-based online approach can request true PPG heart rate when needed.Source: §2.1 · p. 2
3 EXPERIMENT
The experiments compare offline and online heart-rate prediction from wrist acceleration using data from three patients monitored in two approximately 14-day phases. The online approach improved on the offline baseline while reducing heart-rate measurements at higher uncertainty thresholds.
3 EXPERIMENT
- Three patients wore the device for approximately two weeks after intervention and another two weeks around two months later; acceleration features were aligned to heart rate at one-minute intervals.Features included statistics and spectral measures from one-second windows across three acceleration axes.Source: §3 · p. 2
3.1 Offline Baseline
- Offline MAE rose from 4–6 within a phase to 6–8 when training and testing on phases about two months apart, and predictions for Patient 0130 rarely deviated far from the mean.Within-phase performance was up to 2× better than the mean-heart-rate dummy predictor; the reported MSE for Patient 0130 was 94.8.Source: §3.1 · p. 2, Figure 2 · p. 2, Figure 1 · p. 2
3.2 Active Learning
- Because labels cannot be queried retrospectively and P(y|X) may change over time, the online setting decides at each time point whether to measure heart rate or use a prediction.The method frames continuous acceleration with on-demand heart-rate labels as an active learning problem.Source: §3.2 · p. 3
3.3 Algorithm
- PPAW uses an ensemble’s prediction uncertainty and recent history to decide when to query heart rate, update learners, and retrain models.It also retrains learners when their error exceeds T or their time-to-live expires; N controls the recent labeled-data history.Source: §3.3 · p. 3
3.4 Results
- The evaluation measured MAE, MSE, and the fraction of samples requiring a true heart-rate label across all three patients; it fixed N=5 and TTL=10 while varying O.MAE is interpreted as average beats-per-minute error.Source: §3.4 · pp. 3–4
3.5 Varying Uncertainty (O)
- MAE stayed between 3 and 4 across tested uncertainty settings; higher O reduced heart-rate queries, with only a small performance decrease and significant improvement over the offline baseline.The query rate was around 50% at O=1 and around 15–20% at a higher O setting.Source: Figure 5 · p. 4, Figure 6 · p. 4, §3.5 · p. 4
4 CONCLUSION
The paper investigates whether acceleration can predict heart rate over time, and presents an online approach evaluated on three patients. It reports improved prediction with infrequent heart-rate measurements, including a patient example with MAE 2.89 and 20.25% sensor use.
- The approach investigates predicting heart rate from acceleration and tracking heart-rate changes over time.Source: §4 · p. 4
- For patient 549, predictions followed true heart-rate trends with MAE=2.89 and MSE=21.07 while measuring heart rate 20.25% of the time.Source: Figure 7 · p. 4
- MAEs of 2.5–5 beats per minute were achieved across three patients while querying the heart-rate sensor relatively little.For patient 549, the example reports MAE=2.89 while requesting heart-rate measurements 20.25% of the time.Source: §4 · p. 4