Heart Rate Monitoring During Physical Exercise From Photoplethysmography Using Neural Network

Abstract

Photoplethysmography (PPG) signals have been widely used for heart rate (HR) monitoring. Compared to the electrocardiogram, PPG signals can be easily collected with wearable devices such as smart watches at a lower cost. However, the PPG signals are often contaminated by the motion artifact (MA) and noises, which greatly deteriorate the signal quality and pose significant challenges on HR monitoring. In this article, a new algorithm, using the spectral subtraction and the neural network (NN), is developed for accurate HR tracking in the presence of MA and noises. Specifically, the spectral component of MA is estimated from the acceleration (ACC) signals and then removed from the spectra of PPG. In addition, an NN model is developed based on new features extracted from ACC signals to identify the relationship between the ACC and HR variations in consecutive time windows. Such information is further used as a reference to select the spectral peak corresponding to the actual HR. A postprocessing algorithm is used to correct misidentified HR and to improve the accuracy. The NN-based algorithm is validated using the 2015 IEEE Signal Processing Cup Dataset. Our algorithm achieves an average absolute error of 1.03 beats per minutes (BPM) (standard deviation: 1.82 BPM), which outperforms previously reported works in the literature.