Benchmarking of open-source algorithms for heart rate estimation from motion-corrupted photoplethysmography

Abstract

Photoplethysmography holds promise for continuous, non-intrusive heart rate monitoring through wearable devices. However, motion artifacts can impact the reliability of heart rate estimates. The integration of accelerometer data has been proven helpful in mitigating these artifacts. Although several algorithms that combine photoplethysmography and accelerometer data for heart rate estimation have been proposed, it remains unclear which performs best. We performed a systematic and comprehensive search and evaluation of all relevant published algorithms (N = 126) and benchmarked all available open-source methods (N = 11) using the same real-world dataset. A robust methodological framework was employed for assessing these algorithms, featuring a comprehensive set of performance metrics. Out of 126 retrieved articles, 11 provided open-source implementations and were included in the benchmarking. We found that deep learning algorithms consistently outperformed model-based algorithms and algorithms that did not correct for accelerometer data, particularly in dynamic conditions with substantial motion artifacts. The BeliefPPG algorithm performed best across all metrics, with an estimation bias of 0.7 ± 0.8 bpm, an estimation variability of 4.4 ± 2.0 bpm, and a Spearman's correlation of 0.73 ± 0.14 bpm with the heart rate ground truth. These findings underscore the potential of deep learning techniques to enhance the reliability of photoplethysmography-based heart rate monitoring through integration with accelerometer data in real-world conditions. This work provides valuable insights into the performance of these algorithms and highlights the importance of developing broader, more diverse datasets to enhance generalizability in future research.