1-Dimentional Convolutional Neural Network based Heart Rate Estimation Using Photoplethysmogram Signals

Abstract

Recently, as the importance of healthcare has increased, researches are being conducted to measure health status in real time. Heart Rate (HR) measurement is one of the important health conditions that measure heart beat rates. HR measurement can be performed using Photoplethysmogram (PPG) or Electrocardiogram (ECG) signals. Since the PPG or ECG signals are different from people to people, conventional HR estimator occasionally results in large errors. To develop a reliable HR estimator, an HR estimation technique using PPG is proposed in this paper, based on a deep learning technique. The proposed HR estimation technique has the following key features. We develop a new artificial neural network which is 1-Dimensional Convolutional Neural Network (1D-CNN) composed of ten convolutional layers and two fully connected layers. To assess the estimation performance, cross validation is used. The training and verification of the proposed 1D-CNN technique are performed on Python 3.7.5 with Keras 2.0 library. The proposed HR estimation technique performs training and verification using field PPG data. Overfitting is prevented by increasing the limited training data by data augmentation. In training, the loss function is the Mean Square Error (MSE), which is commonly used in regression problems. In the verification, the error between the predicted HR and the actual HR is compared using Mean Absolute Error (MAE). As a result of the final performance verification through cross validation, the proposed technique shows an MAE of 1.23 Beats Per Minute (BPM). This results indicate that the proposed technique enables quick and accurate HR estimation with only PPG signals. Therefore, if this technique is applied to medical and wearable devices, the proposed technique can replace the existing HR monitors.