A Deep Learning approach for estimating Time-Domain Heart Rate Variability parameters from wrist photoplethysmography during daily activity

Heart rate variability (HRV) is a critical indicator of autonomic nervous system regulation and cardiovascular health, typically measured using electrocardiography (ECG). Wrist devices are gaining popularity as non-invasive alternatives to monitor heart rate (HR) and pulse rate variability (PRV) in unconstrained setting through photoplethysmography (PPG). However, movement artifacts severely deteriorate signal quality, making estimation reliability challenging. Deep learning approaches are establishing as a promising means to address this issue, although they focus on HR while overlooking the more complex task of reconstructing PRV and thus limiting the insights about autonomic dynamics. This study introduces a novel methodology based on an ensemble of convolutional denoising autoencoders (CNN-DAEs) to denoise PPG time series under high-intensity activity, followed by a PRV tracking algorithm designed to improve the reliability of PRV time-domain parameters (meanHR, stdRR, RMSSD). The CNN-DAEs are trained using a custom loss function that emphasizes both the fidelity of the reconstructed PPG signal and the physiological plausibility of the derived inter-beat intervals. We validated our approach on two public datasets (IEEESPC and PPG-Dalia). HR estimation errors were 1.74bpm for IEEESPC and 4.69bpm for PPG-Dalia. Additionally, estimation errors for stdRR and RMSSD were as low as 7.94ms and 5.01ms on the IEEESPC dataset, and 17.06ms and 12.69ms for PPG-Dalia, respectively. The proposed approach provided reliable HR and, to some extent, stdRR and RMSSD estimates during high-intensity activities. This methodology could be adopted for the monitoring of autonomic parameters during daily-life activities.