Abnormal Cardiac Condition Classification of ECG Using 3DCNN - A Novel Approach

The automated ECG data analysis identifies patients with cardiac problems, thereby ensuring accuracy, saving time and medical resources. This article presents a novel approach to the abnormal cardiac condition classification of ECG using 3DCNN. The classification is primarily based on three cardiac conditions: Sinus Rhythm, Abnormal Arrhythmia, and Congestive Heart Failure from the MIT-BIH arrhythmia database. The ECG signals are first converted to scalogram images using wavelet transform, and the scalogram images are then segmented and stacked to form a three-dimensional image. The scalogram conversion leverages the conventional filtering and feature extraction steps. The patch-wise approach focuses on the local patterns and extracts the subtle features relevant to the diseased and non-diseased conditions. The 3DCNN filters pose an advantage on 2DCNN by simultaneously learning representation from a few patches, thereby exploring the temporal properties of ECG. The extracted features from the filters are then classified using various classifiers, among which the Naive Bayes classifier performed the best, yielding the highest accuracy of 90.4%.