Machine Learning–Based Prediction Models for Early Detection and Risk Stratification of Cardiovascular Diseases: A Systematic Review

Abstract

Background: Machine-learning (ML) models can use high-dimensional electronic health record, imaging, waveform and biomarker data to estimate cardiovascular risk, but apparent gains in discrimination may be offset by poor calibration, overfitting and limited external validation. Objectives: To compare ML-based cardiovascular risk prediction with conventional risk scores, evaluate model quality and external validity, and quantify discrimination in studies reporting directly comparable AUC estimates. Methods: A PRISMA-structured search framework covered PubMed, Scopus, Embase and Cochrane through September 2026. Cohort studies developing or validating ML models for incident cardiovascular disease or ASCVD risk were eligible when they reported clinically interpretable discrimination and a conventional comparator. PROBAST was used as the primary quality framework. Random-effects models separately pooled reported AUCs for ML models and conventional scores in three large comparative cohorts. Results: The three directly comparable cohorts included hundreds of thousands of participants. Pooled ML AUC was 0.791 (95% CI 0.758–0.823; I²=98.6%), whereas pooled conventional-score AUC was 0.738 (95% CI 0.723–0.752; I²=92.2%). The descriptive difference was about 0.053, but paired inference was not performed because within-study covariance was unavailable. Larger systematic reviews confirm modest average discrimination gains but identify high risk of bias, inconsistent calibration reporting and scarce independent external validation. Conclusion: ML can improve discrimination and broaden risk estimation beyond conventional variable sets, but current evidence does not support uncritical replacement of established scores. Independent external validation, calibration, decision-curve analysis, fairness assessment and prospective impact studies are required.