Comparative Analysis of Simplified 1D Convolutional Neural Network Performance on Balanced ECG Datasets for Cardiac Arrhythmia Detection

Abstract

Globally, cardiovascular illnesses continue to be the leading cause of morbidity and mortality. Among them, arrhythmia is a prevalent cardiovascular disorder that affects a significant portion of the population. Timely detection and precise classification of cardiac arrhythmias are essential for actual intervention and patient care. Convolutional Neural Networks (CNNs), a cutting-edge advancement in deep learning, have shown remarkable success in automating the identification of cardiac arrhythmias from electrocardiogram (ECG) data. However, the performance of these models can vary significantly depending on the class distribution of the dataset, with imbalanced datasets presenting a significant challenge. This study presents a comparative analysis of one-dimensional convolutional neural network (1D CNN) performance when applied to multiple balanced ECG datasets. It sheds light on the impact of data distribution on model effectiveness. In the balanced dataset, the 1D CNN achieved improved accuracy and F1-score to detect abnormal ECG patterns.
The comparison was conducted on different batch sizes of 16, 32, 64, and 128. Among batch sizes 16 and 32, the BorderlineSMOTE data balancing on MIT-BIH 1D CNN mode yielded superior performance to other balancing methods based on the accuracy, F1-score and ROC-AUC Score. Similarly, in the case of an increase in batch sizes to 64 and 128, ADASYN oversampling applying on the specified model provided improved results compared to others. The accuracy and F1-score vary from a range of 0.9752 to 0.9804 and 0.8622 to 0.8994 respectively. These findings emphasize the need for careful consideration of medical dataset-balancing techniques while applying deep learning models.