Bridging machine learning and clinical practice: a multicentre nomogram for 90-day graft failure risk stratification in heart transplantation

Background Early graft failure within 90 postoperative days is the leading cause of mortality after heart transplantation. Existing risk scores, based on linear regression, often struggle to capture the complex, multifactorial biological interactions necessary for personalised donor–recipient matching. This study utilised explainable machine learning (ML) to identify robust predictors of 90-day graft failure and developed a clinically interpretable, ML-informed nomogram designed specifically for cross-population generalisability. Methods Using the UNOS registry (2008–2020; n=25 200), XGBoost/Random Forest models identified 90-day graft failure predictors from 32 donor–recipient variables. Explainable AI (SHapley Additive exPlanations) analysis revealed key predictors and their non-linear interactions, which were translated into a clinically applicable nomogram. External validation was performed on a large, single-centre Chinese cohort (Wuhan Union Hospital ; 2018–2023; n=563), assessing performance via area under the curve (AUC), calibration and decision curve analysis (DCA). Findings The final model incorporated eight predictors: recipient factors (prior cardiac surgery, age, bilirubin, body mass index (BMI)), donor factors (age, gender, BMI) and cold ischaemia time. The XGBoost-derived nomogram demonstrated consistent discrimination (AUC 0.67, 95% CI 0.64 to 0.70) and calibration. Patients stratified into the high-risk group (top quantile by nomogram score) had a 2.4-fold increased hazard of graft failure (HR 2.42, 95% CI 2.11 to 2.78). DCA confirmed the model’s clinical utility across a wide range of risk thresholds (0.0–0.4). External validation in the Chinese cohort affirmed its generalisability (AUC 0.67). Conclusion This study introduces an ML-informed nomogram for 90-day graft failure, validated across USA and Chinese populations. By translating ML insights into a clinically interpretable tool using routinely available pretransplant variables, it bridges a key translational gap in transplant risk prediction. This tool can aid in optimising donor–recipient matching and personalising post-transplant management, with the potential to help address geographic disparities in heart transplant outcomes.