Artificial Intelligence Applied to Electrocardiogram-Based Cardiovascular Risk Assessment: A Systematic Review

Abstract

Abstract Cardiovascular diseases (CVDs) are the leading cause of mortality worldwide, underscoring the need for effective risk prediction and early detection. Although the electrocardiogram (ECG) is a widely available and low-cost diagnostic tool, its traditional interpretation is limited by subjectivity. Artificial intelligence (AI) has emerged as a promising approach, capable of extracting hidden prognostic information from ECG signals. This systematic review aimed to assess original studies applying AI techniques to ECGs for cardiovascular risk prediction and mortality. Original studies that used ECG signals as the sole input variable for AI models, focusing on cardiovascular risk outcomes, were included. A systematic search was conducted in different databases, and data were synthesized narratively. Eleven studies were included, predominantly retrospective cohorts applying convolutional neural networks (CNNs) to predict cardiovascular risk or mortality. The sample primarily consisted of adult populations in high-income countries. Primary outcomes included all-cause mortality, cardiovascular death, and major adverse cardiovascular events (MACE). Reported AUROC values ranged from 0.63 to 0.961 in training sets, with some models outperforming traditional risk scores. AI-ECG models demonstrated the potential to detect subclinical disease, enabling early risk stratification even in normal ECGs. However, challenges remain regarding population diversity, model interpretability, and prospective validation. The application of AI to ECG analysis represents a promising advancement in personalized cardiovascular risk assessment. Nonetheless, further research is needed to ensure the safety, effectiveness, and equitable clinical integration of these technologies.