A Survey on ECG-Based Classification of Cardiac Arrhythmias using with Convolutional Neural Networks

Abstract

Recently, concern for cardiac health has increased, leading to the development of neural network models to diagnose arrhythmias. This study presents a systematic review of current approaches in the classification of cardiac arrhythmias with convolutional neural networks. The predominant databases, the most common arrhythmia types, preprocessing techniques, the most applied convolutional neural network models, and the most used evaluation metrics are addressed. The findings show a trend towards diagnoses such as normal sinus rhythm, left bundle branch block and right bundle branch block. In preprocessing, the use of filters to reduce noise in electrocardiogram signals, segmentation and balancing of the records is highlighted. The MIT-BIH arrhythmia database was identified as the most used in studies. Finally, the effectiveness of convolutional neural network models combined with long short-term memory networks and transformer-based attention modules is highlighted.