Transfer learning for improved electrocardiogram diagnosis of cardiac disease: exploring the potential of pre-trained models

Abstract

Predicting the onset of cardiovascular disease (CVD) has been a hot topic for researchers for years, and recently, the concept of transfer learning has been gaining traction in this field. Transfer learning (TL) is a process that involves transferring information gained from one task or domain to another related task or domain. This paper comprehensively reviews recent advancements in pre-trained TL models for CVD, focusing on electrocardiogram (ECG) signals. Forty-three articles were chosen from Scopus and Google Scholar sources and reviewed, focusing on the type of CVD detected, the database used, the ECG input format, and the pre-training model used for transfer learning. The results show that more than 80% of the studies utilize 2-dimensional (2D) ECG input from the two most utilized available ECG datasets: MIT-BIH arrhythmia (ARR) and MIT-BIH normal sinus rhythm. alexnet, visual geometry group (VGG), and residual network (ResNet) are among the pre-trained TL models with the highest number used among reviewed articles. Additionally, the development of pre-trained TL models over time has made it possible to detect CVD with ECG signals. It can also address limited data problems, promote the development of more dependable and resilient detection systems, and aid medical professionals in diagnosing CVD and other diseases.