Improving the Recognition Performance of Apnea from Electrocardiogram Signal Based on Convolutional Neural Network (CNN)

Abstract

Due to the consequences of apnea heart disease, include high blood pressure, heart failure, and heart attacks, early diagnosis is important and has now become a subject of interest for researchers in the world. So far, many researches have been done in this field. Among these, we can mention the use of various pre-processing in the field of time and frequency, along with classifications based on machine learning. These classifications work well in most cases. But in cases where the number of features extracted from the input data increases or the hidden patterns in the data are complex, they do not perform well. In addition to classification, the poor performance of these studies can be caused by pre-processing or feature extraction itself. In this article, in order to solve the problems of previous studies, pre-processing of the ECG signal has been done using a Butterworth pass filter, Golay Savizki filter and wavelet transform. Pre-processing makes the R peaks in the input to be well distinguished. Finally, by extracting the exact position of the R-R interval in the ECG with minimal error, efficient features that enable the proper description of different classes can be extracted. These features include the wavelet, can provide good results due to their time-frequency nature. The simulation results also show these features well. In terms of choosing the classification, a convolutional neural network (CNN) has been used to diagnose the disease. The accuracy percentage of 97.91% obtained from the simulation of the proposed method on the sleep apnea database of University College Dublin is a proof of good performance of this method.