Improved Deep Learning Based ECG Classification through Automated Feature Selection and Weighted Loss Function

Abstract

This paper proposes two approaches to significantly improve the detection rate of abnormal heartbeats in an Electrocardiogram (ECG) based deep learning heartbeat classifier. We introduce an automated feature selection procedure using the Kendall rank correlation coefficient to improve the performance of already existing classifier models. Further, we propose a methodology to cope with the class imbalance present in many ECG and other medical datasets by using a weighted loss function. The proposed methods demonstrate a significant improvement in the detection of Supraventricular Ectopic Beat (SVEB) and Ventricular Ectopic Beat (VEB) type heartbeats. Boasting an impressive 20% increase in terms of recall for the SVEB class when compared to state-of-the-art classifiers. This advancement could lead to more reliable and efficient tools for early arrhythmia detection, particularly beneficial in places where professional medical care is not easily accessible.