The role of artificial intelligence in predicting cardiovascular outcomes: a systematic review and meta-analysis

This systematic review and meta-analysis evaluated the performance, validation strategies, and clinical applicability of AI-based models for cardiovascular outcome prediction. Following PRISMA 2020 guidelines, seven databases were searched for studies published between 2015 and 2025. Studies reporting predictive performance of AI models were included. A bivariate random-effects model was used to pool sensitivity, specificity, and area under the curve (AUC), while heterogeneity, publication bias, and risk of bias (PROBAST) were assessed. Seventeen studies met the inclusion criteria. AI models, including DL, ensemble learning, and traditional ML methods, showed strong performance, with a pooled AUC of 0.86, sensitivity of 83.5%, and specificity of 81.7%. DL models achieved the highest performance with an AUC of 0.89. Models trained on large datasets and those externally validated demonstrated better generalizability. Moderate heterogeneity (I2 = 48%) reflected differences in study design, populations, and methodologies. Nearly half of the studies showed high risk of bias, mainly due to overfitting and limited calibration. Funnel plot asymmetry suggested possible publication bias. Explainability tools improved interpretability. Overall, AI models show strong potential but require more rigorous validation and transparent reporting before widespread clinical implementation.Abbreviations: AI: artificial intelligence; AUC: area under the curve; CNN: convolutional neural network; CVD: cardiovascular disease; DL: deep learning; EHR: electronic health record; F1-score: harmonic mean of precision and recall; Grad-CAM: gradient-weighted class activation mapping; K-fold: K-fold cross-validation; LIME: local interpretable model-agnostic explanations; MACE: major adverse cardiovascular events; ML: machine learning; MLP: multi-layer perceptron; PROBAST: prediction model risk of bias assessment tool; PRISMA: preferred reporting items for systematic reviews and meta-analyses; SHAP: Shapley additive explanations; SROC: summary receiver operating characteristic; TRIPOD-AI: transparent reporting of a multivariable prediction model for individual prognosis or diagnosis – artificial intelligence; CONSORT-AI: consolidated standards of reporting trials – artificial intelligence; XGBoost: extreme gradient boosting