A systematic review of machine learning and explainable artificial intelligence for electrocardiogram based cardiovascular disease prediction

Abstract

Cardiovascular disease (CVD) are a leading cause of global morbidity and mortality, making early and accurate detection essential for improving patient outcomes. The electrocardiogram (ECG) is one of the most important diagnostic tools for identifying various cardiac abnormalities. Recent advances in machine learning (ML) and deep learning (DL) have shown strong potential for improving ECG-based CVD detection. This review critically analyzes recent ECG-based ML/DL models for CVD prediction, with particular attention to model performance, dataset characteristics, preprocessing strategies, and the integration of explainable artificial intelligence (XAI). The review covers a broad range of ECG-related applications, including arrhythmia, myocardial infarction, atrial fibrillation, heart failure, pediatric and fetal ECG analysis, wearable ECG monitoring, ECG image analysis, and ECG-centered multimodal approaches. Across the 75 included studies, reported diagnostic performance was generally high, with accuracy commonly ranging from approximately 90 to 99%, sensitivity from approximately 81 to 99%, specificity from approximately 84 to 99%, AUROC values reaching up to approximately 0.98–0.99, and F1-scores exceeding 0.90 in several benchmark datasets. However, these results varied substantially according to dataset characteristics, disease category, preprocessing strategy, model architecture, validation protocol, and evaluation metrics. Although many studies reported excellent internal validation performance, relatively few performed external or prospective clinical validation, highlighting an important gap in the translation of AI-assisted ECG models into routine clinical practice. The most frequently used datasets included MIT-BIH, PTB-XL, PhysioNet, CPSC, and UK Biobank, while dominant model families included CNN, LSTM, Transformer, SVM, RF, and hybrid CNN-LSTM architectures. Although ML/DL models have reported promising diagnostic performance, several challenges remain, including class imbalance, data quality variation, limited generalizability, model interpretability, and the limited availability of external and prospective clinical validation, which remain major barriers to the widespread clinical adoption of AI-assisted ECG diagnostic systems. In addition, ethical, legal, and social issues, such as data bias, privacy, transparency, and clinical trust, must be addressed before these models can be reliably and responsibly deployed in clinical practice. This review provides an integrated overview of ECG-based AI models for CVD prediction and highlights future research directions for developing more interpretable, generalizable, and clinically trustworthy AI-assisted ECG diagnostic systems.