Cardiac Arrhythmia Diagnosis based on Features Extraction and Convolutional Neural Network from ECG Signals

Abstract

Conventionally, arrhythmias are diagnosed and classified by using manual inspection of electrocardiogram (ECG) signal. However, the diagnosis needs experts’ interpretation and is time-consuming. To make the diagnosis efficient, the subject’s ECG recordings from the MIT-BIH database have been segmented in short ECG segments of 60 seconds, and the investigation has been conducted in two categories: features extraction and pattern recognition. Here, the features (i.e. NN50, Mean RRI, kurtosis, and skewness) have been extracted in terms of time-domain and statistical analysis. From these analyses, it is evident that the value of NN50, Mean RRI, kurtosis, and skewness are respectively in the range of 30 – 50, 600–1000 ms, 1–3, and -1 to 1 for the ECG segments of the healthy group, while arrhythmia segments show values beyond those ranges. Besides, a pre-trained AlexNet convolutional neural network (CNN) has been used to facilitate the diagnosis. Here, the scalogram images of ECG segments have been obtained from time-frequency analysis using continuous wavelet transform (CWT), and these images have been fed into the CNN classifier for pattern recognition and classification. In this approach, the sensitivity, specificity, and overall accuracy are 98.7%, 100%, and 99.30%, respectively. So, the proposed analyses have delivered concrete results, and it can be an effective way to differentiate the healthy and arrhythmic groups.