Deep Learning Based Heart Rate Estimation for Photoplethysmogram Signals

Abstract

Heart Rate (HR) is a basic and importantbiomarker that measures heart beat rates. This paper proposes HR estimation technique based on deep learning technique using Photoplethysmogram (PPG) signals. For reliable and accurate HR estimation, we propose a 1-Dimensional Convolutional Neural Network (1D-CNN) model which consists of 10 convolutional layers and 2affine layers. To examine the HR estimation accuracy, cross validation for all possible combination of training and test data sets is performed. The programming tools are Python 3.7.5 andKeras 2.0. To avoid overfitting due to the small data set, data augmentation technique is incorporated. The loss function for training is the Mean Square Error (MSE), one of the famous errors for a regression problem. For verification performance, we measure Mean Absolute Error (MAE). According to the verification results, the proposed estimator shows an MAE of 1.23 Beats Per Minute (BPM). This results indicate that the proposed technique can be used as an alternative of the existing techniques. If the technique is applied to wearable devices, reliable day-and-night HR measuring can be possible.