Cardiac disease classification using total variation denoising and morlet continuous wavelet transformation of ECG signals

Abstract

Cardiac diseases (CDs) are a leading cause of death, and electrocardiogram (ECG) is the most important noninvasive method for the diagnosis of CDs. The objective of this study was to diagnose CDs using Morlet continuous wavelet transform from 1-channel ECG. Four CDs were studied: heart failure (HF), sudden cardiac death (SCD), ventricular fibrillation (VF), and atrial fibrillation (AF). One hundred ECG signals were obtained from the open source MIT-BIH cardiac arrhythmia database. The distribution of patients' diseases was 20 HF, 20 SCD, 20 VF, 20 AF, and 20 normal. Total variation denoising (TVD) was used to filter ECG signals without smoothing sharp edges, and then Morlet continuous wavelet coefficient matrices were calculated. Five features were calculated from each row in the coefficient matrix (5 features × 9 rows), based on statistical parameters. The classification of CDs versus normal was based on binary logistic regression, and the classification of specific CDs was by multinomial logistic regression. The accuracy of detecting HF, SCD, VF, and AF with respect to normal was 95%, 100%, 100%, and 100%, respectively. The overall accuracy of identifying the correct CD was 98%. In conclusion, features extracted from wavelet transform of a TVD filtered ECG signal successfully distinguished CD patients from normal cardiac patients, and identified the type of CD. This method could be implemented in ECG machines for auto-detection of CDs.

Keywords: Cardiac Disease Detection, Electrocardiogram, Logistic Regression, Morlet Continuous Wavelet, Total Variation Denoising