Machine learning approaches for the prediction of 1-year mortality after heart transplantation: a systematic review and meta-analysis
D Mutahar, A ZakaAbstract
Background
Accurate prediction of 1-year all-cause mortality following heart transplantation is critical for donor-recipient matching, equitable organ allocation, and post-transplant management. Conventional risk scores, including the Donor Risk Index and Independent Mortality Prediction After Cardiac Transplantation (IMPACT) score rely on predefined predictor variables and demonstrate modest discrimination. Machine learning (ML) approaches have emerged as registry-based risk-prediction strategies that may offer improved predictive performance compared with conventional risk scores.
Purpose
To compare the discrimination between ML-based models and conventional risk scores for predicting 1-year all-cause mortality after heart transplantation.
Methods
This study was conducted in accordance with PRISMA and TRIPOD-AI guidelines. PubMed, EMBASE, Web of Science, and Cochrane databases were searched until February 2026 for studies directly comparing ML models with conventional risk models in adult heart transplant recipients. The primary outcome was comparative discrimination measured by C-statistics with 95% confidence intervals (CIs) for predicting the risk of 1-year all-cause mortality. Random-effects meta-analysis was performed to pool discrimination estimates and assess between-group differences. Feature importance was synthesised from studies reporting SHAP (Shapley additive explanation) analyses. Prespecified subgroup analysis was conducted in patients based on validation status and ML-subtype.
Results
Eight studies including 14 distinct models and 305,186 transplant recipients were included. Summary C-statistic for 1-year all-cause mortality was higher for ML models than conventional risk scores, C-statistic 0.68 [95% CI 0.66-0.70] vs 0.62 [0.60-0.64]; absolute difference 0.06 [95% CI 0.03-0.08], p<0.001). Random-forest ensemble models were the top-performing ML model, with summary C-statistic 0.67 [0.65-0.69]). External validation was performed in 37.5% of included models, and calibration was inconsistently reported. In subgroup analysis, the difference in C-statistic was 0.12 (95% CI 0.09-0.14; p<0.001) in internally validated studies and 0.01 (95% CI 0.02-0.04; p=0.40) in externally validated studies. Across SHAP analyses, recipient/donor age, creatinine and bilirubin were the strongest predictors of all-cause mortality. PROBAST-AI overall risk of bias was high.
Conclusion
ML models marginally outperformed conventional risk scores in the discrimination of 1-year all-cause mortality after heart transplantation. Immediate clinical applicability remains uncertain, due to limited external validation, inconsistent calibration, and methodological heterogeneity.ML vs Conventional for 1-Year Mortality