Artificial Intelligence for Cardiovascular Risk Prediction: An Umbrella Review of Applications and Translational Challenges

Comprehensive Summary

This umbrella review examined how artificial intelligence is being used for cardiovascular risk prediction and how well those models translate into clinical practice. Traditional risk scores often miss important biomarkers and environmental stressors and can perform poorly in low- and middle-income countries and among non-Caucasian populations. In contrast, machine learning and deep learning models can capture nonlinear relationships across diverse risk factors, improving discrimination and reclassification. The review searched major databases using medical subject headings and keywords, included studies with quantitative metrics and robust designs, and excluded non-English and non-original reports. Of 3,500 records retrieved, 48 studies met inclusion criteria, including randomized trials, cohort studies, and systematic reviews or meta-analyses. Study quality was generally moderate to good, with randomized trials showing low risk of bias and cohort studies scoring 7–9 stars on the Newcastle-Ottawa Scale. Because the included studies were heterogeneous in populations, data sources, validation methods, and outcome definitions, the authors used narrative synthesis rather than meta-analysis. Overall, AI-based models showed stronger predictive performance than conventional tools, with imaging-based models performing best, often exceeding AUC 0.90, while EHR-based models generally ranged from 0.82 to 0.88. Across studies, AI models achieved 80–95% accuracy in identifying high-risk patients. Other applications also showed benefit, including improved MI and heart failure prediction with natural language processing, better calibration with personalized models, stronger coronary event prediction using polygenic risk scores, and high-sensitivity atrial fibrillation detection with wearable AI. However, external validation, calibration, fairness evaluation, and multicenter testing remained limited, which weakens confidence in broad clinical adoption and more research is required.

Outcomes and Implications

These findings suggest that AI could improve early cardiovascular risk identification, especially when imaging, EHR, genomic, and wearable data are combined. In practice, the strongest near-term value appears to be in diagnostic support and workflow efficiency, while long-term prevention tools still need stronger validation, fairness testing, and prospective multicenter studies before routine adoption.