Motion Artifact Detection in PPG Signals Based on Gramian Angular Field and 2-D-CNN

Abstract

Due to the presence of motion artifacts (MAs), heart rate monitoring using PPG sensors in daily life and physical exercise is challenging, and there have been many studies on MA removal algorithms. However, most studies do not consider the quality evaluation of PPG signal before the MA removal. In this way, removing the MA directly regardless of whether there is motion artifact signal is not only a waste of computing resources, but also easy to introduce new noise. In this paper, the MA detection in PPG signal is performed by dividing the original signal into 6s signal segments and calculating the amplitude mean difference function (AMDF). Then the obtained AMDF is converted into a 2-D image through the Gramian Angular Field (GAF), and then classified by the Convolutional Neural Networks (CNN) classifier, so as to distinguish the contaminated signal and clean signal. In the subsequent processing, only the contaminated signal needs to remove the MAs, and the clean signal segment can be directly used for heart rate estimation. In this study, we achieve a classification accuracy of 0.966 in the local database, and a classification accuracy of 0.946 in the BIDMC PPG and Respiration Dataset published by PhysioNet. With the combination of feature extraction and SVM classifier, the proposed method has significantly improved the results.