Heart rate estimation in PPG signals using Convolutional-Recurrent Regressor

Abstract

Heart rate monitoring using PPG signal has emerged as an attractive as well as an applied research problem which enjoys a renewed interest in the recent years. Spectral analysis of PPG for heart rate monitoring, though effective when the subject is at rest, suffers from performance degradation in case of motion artifacts which mask the peak related with the actual heart rate. Leveraging the recent advancements in deep (machine) learning and exploiting the signal, spectral and time-frequency perspectives, we introduce an effective method for heart rate estimation from PPG signals acquired from subjects performing different exercises. We extract a set of features characterizing the signal and feed these feature sequences to a hybrid convolutional-recurrent neural network (C-RNN) in a regression framework. Experimental study on the benchmark IEEE signal processing cup dataset reports low error rates reading 2.41 ± 2.90 bpm for subject-dependent and 3.8 ± 2.3 bpm for subject-independent protocol thus, validating the ideas put forward in this study.