Abstract
Photoplethysmography (PPG) is a non-invasive physiological sensing method used in many clinical applications, increasingly supported by machine learning. However, systematic comparisons of input representations and models remain limited.
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 (BP) estimation and atrial fibrillation (AF) 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, the strongest performance is observed for deeper convolutional neural networks (CNNs). However, depending on the task, smaller or lower-capacity CNNs can also achieve competitive performance, as confirmed by Bland–Altman analyses and statistical significance analyses based on bootstrapping.
By providing a controlled, like-for-like comparison across signal, feature, and image-based representations, this study offers practical guidance for selecting robust machine-learning approaches for real-world PPG applications.