Machine-learning for photoplethysmography analysis: Benchmarking feature, image, and signal-based approaches

TL;DR

This paper benchmarks input representations for PPG-based machine learning across two prototypical tasks: cuffless blood pressure estimation (regression) and atrial fibrillation detection (classification). By comparing raw time-series CNNs, interpretable feature-based pipelines, and image-based approaches (CWT scalograms), the study finds that modern CNNs operating on raw PPG time series consistently provide the strongest performance, with image representations occasionally competitive and feature-based methods lagging in many settings. The results highlight a generalizable preference for end-to-end learning on raw data, while also emphasizing task- and dataset-dependent nuances between model complexity and generalization. The work informs researchers about input representation choices for PPG analytics and points to future directions in pretraining, uncertainty quantification, and out-of-distribution evaluation.

Abstract

Photoplethysmography (PPG) is a widely used non-invasive physiological sensing technique, suitable for various clinical applications. Such clinical applications are increasingly supported by machine learning methods, raising the question of the most appropriate input representation and model choice. Comprehensive comparisons, in particular across different input representations, are scarce. We address this gap in the research landscape by a comprehensive benchmarking study covering three kinds of input representations, interpretable features, image representations and raw waveforms, across prototypical regression and classification use cases: blood pressure and atrial fibrillation prediction. In both cases, the best results are achieved by deep neural networks operating on raw time series as input representations. Within this model class, best results are achieved by modern convolutional neural networks (CNNs). but depending on the task setup, shallow CNNs are often also very competitive. We envision that these results will be insightful for researchers to guide their choice on machine learning tasks for PPG data, even beyond the use cases presented in this work.

Figures (9)

• Figure 1: An exemplary PPG signal showing a pulse wave for each heartbeat. Pulse onsets, representing individual heartbeats, are shown as red circles. An inter-beat interval is labeled, corresponding to the time between consecutive heartbeats (adapted from ref7).
• Figure 2: The definition of PPG pulse wave morphology features. Higher-order statistical features extracted from PPG pulse waveforms include skewness and kurtosis, with the latter being the most statistically significant feature for blood pressure estimation.
• Figure 3: Structure of an MLP-based atrial fibrillation detection using features extracted from the PPG signal.
• Figure 4: Good-quality AF PPG segment (a) with extracted PP intervals (b) and quality index (c).
• Figure 5: Good-quality AF PPG segment (a) with extracted PP intervals (b) and quality index (c).