Estimating Heart Rate Variability from Wrist-Worn Photoplethysmography Devices in Daily Activities: A Preliminary Convolutional Denoising Autoencoder Approach

Wrist photoplethysmography (PPG) devices are gaining popularity as a non-invasive means to monitor pulse rate variability (PRV) in daily-life settings. Yet, movement artifacts make reliable estimation of PRV challenging during physical activities even if not very intense. Various approaches based on spectral analysis and deep learning (DL) have provided mean HR over time with low estimation errors. However, mean HR dynamics cannot be adopted to derive detailed information about autonomic activity, for which PRV time series is necessary. In this preliminary work, we propose a novel approach combining a convolutional denoising autoencoder (CNN-DAE) with a physiologically-constrained custom loss function, which leverages synchronous electrocardiographic (ECG) recordings and inter-beat interval (IBI) information to reconstruct the PPG signal, free from artifacts, and obtain reliable PRV. The reconstructed PRV has been averaged across time windows to estimate the mean HR and compare it against those obtained from standard bandpass filtering procedures of PPG and ECG's HRV, which was used as the gold standard reference. Our preliminary results suggest that our method can accurately estimate PRV, providing mean HR unbiased estimates with significantly lower error rates than conventional approaches. This suggests that the proposed methodology could be adopted to denoise PPG time series in uncontrolled environments.