[Paper Review] Online Heart Rate Prediction using Acceleration from a Wrist Worn Wearable

Ryan McConville, Gareth Archer+5 more|Bristol Research (University of Bristol)|Jun 25, 2018

Non-Invasive Vital Sign Monitoring6 references4 citations

TL;DR

This paper proposes an online active learning approach to predict heart rate from wrist-worn accelerometer data, reducing reliance on energy-intensive PPG sensors. By dynamically querying PPG only when prediction uncertainty is high, the method achieves a mean absolute error of 2.89 bpm while using the PPG sensor only 20.25% of the time over four weeks of free-living data from three patients recovering from heart valve interventions.

ABSTRACT

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).

Motivation & Objective

• To investigate whether accelerometer data from a wrist-worn wearable can reliably predict heart rate over extended periods in real-world conditions.
• To address concept drift in heart rate-acceleration relationships due to changes in fitness, aging, or recovery stages after cardiac intervention.
• To minimize energy consumption by reducing reliance on power-hungry PPG sensors through intelligent, on-demand querying.
• To develop and evaluate an online active learning framework that adapts to changing physiological patterns over time.
• To demonstrate the feasibility of long-term, low-power heart rate monitoring using only accelerometer data with periodic PPG validation.

Proposed method

• An ensemble of regression models is trained on one-minute windows of accelerometer features extracted from three axes, including statistical, spectral, and zero-crossing metrics.
• The system uses uncertainty-based active learning: PPG is queried only when prediction variance exceeds a threshold $ O $, indicating high uncertainty.
• Models are retrained when prediction error exceeds a threshold $ T $, or when they exceed a time-to-live (TTL) age, ensuring adaptation to concept drift.
• The algorithm runs online on edge gateways, processing streaming data and updating models in real time with minimal latency.
• The method balances prediction accuracy and sensor usage by prioritizing PPG queries during high-uncertainty intervals, reducing overall PPG sampling frequency.
• The system is evaluated on 4 weeks of free-living data from three patients across two recovery phases, with features aligned to one-minute PPG heart rate measurements.

Experimental results

Research questions

• RQ1 Can accelerometer data alone predict heart rate with acceptable accuracy over extended periods in free-living conditions?
• RQ2 How does performance degrade over time when using a static, offline model trained on initial data?
• RQ3 Can an online active learning approach effectively adapt to concept drift in the heart rate-acceleration relationship during patient recovery?
• RQ4 What is the trade-off between prediction accuracy and PPG sensor usage when using uncertainty-based querying?
• RQ5 To what extent can energy savings be achieved by minimizing PPG usage while maintaining clinically acceptable heart rate prediction accuracy?

Key findings

• The proposed online active learning method achieved a mean absolute error (MAE) of 2.89 beats per minute on patient 549, with only 20.25% of samples requiring PPG measurement.
• Performance remained strong across all three patients, with MAE values between 2.5 and 5.0 bpm, significantly outperforming the offline baseline model.
• Increasing the uncertainty threshold $ O $ reduced PPG usage from ~50% to 15–20% of samples, with only a marginal increase in MAE.
• The online approach successfully tracked heart rate trends, including peaks and troughs, more accurately than the offline baseline, as visually confirmed in prediction vs. true heart rate plots.
• The method demonstrated robustness to concept drift, maintaining accuracy over months despite changes in patient physiology during recovery.
• The system achieved substantial energy savings by reducing PPG sensor usage by up to 80%, while maintaining clinically relevant prediction accuracy.