Coronary Heart Disease Prediction Technique Based on Self-Attention Mechanism and Contrastive Learning

Cardiovascular disease remains one of the leading causes of mortality worldwide, with coronary heart disease(CHD) being a predominant form of cardiac pathology. Electrocardiogram (ECG) is widely employed for cardiac disease detection; however, accurate interpretation of ECG signals requires substantial clinical expertise. To address this challenge, this paper proposes a novel approach that combines Butterworth filtering and​continuous wavelet transform (CWT) to extract time-frequency domain features from ECG signals. Subsequently, a self-attention mechanism and contrastive learning framework are utilized to pre-train an ECG time-frequency feature extractor, ultimately enabling the development of a dedicated prediction network for CHD diagnosis. The proposed method is rigorously evaluated using ECG data from the publicly available MIMIC-III dataset. Experimental results demonstrate exceptional performance, achieving 90% accuracy and 90% F1-score in CHD prediction, outperforming existing models such as convolutional neural networks (CNNs).