Automated ECG-Based Heart Disease Prediction Using Self-Supervised Learning

Abstract

Remarkably being the primary cause of death worldwide, cardiovascular diseases (CVDs) emphasize the need of accurate and quick diagnosis. An indispensable tool for the diagnosis of heart problems is electrocardiogram (ECG); yet, the scarcity of labelled ECG data poses a major challenge for the creation of efficient machine learning models. This article presents a novel system based on self-supervised learning approaches for the automated prediction of heart disease via ECG data. Using self-supervised learning, the proposed method generates generalizable representations from unlabelled ECG data, hence reducing need on large annotated datasets. Two separate phases comprise the framework: (1) Self-Supervised Pre-training, in which a Transformer-based encoder or 1D convolutional neural network (CNN) is trained via contrastive learning to derive significant features from raw ECG signals; and (2) Supervised Fine-tuning, in which the pre-trained encoder undergoes fine-tuning on labelled ECG data especially for the classification of heart disease. Using publically available datasets such as PTB-XL and MIT-BIH, the model is evaluated and shows improved capacity in forecasting a spectrum of cardiac diseases. By combining the benefits of SSL with deep learning, this system offers a scalable and effective method for automated heart disease prediction—which might be applied in real-time ECG monitoring and early detection. The proposed approach addresses the problems of restricted data and the necessity of more general application, therefore enabling better and exact cardiovascular healthcare choices.