Abstract
Due to the limited availability of donor hearts, precise and transparent prediction of post-transplant outcomes is critical for optimizing recipient selection and ensuring long-term survival. In this study, we propose a novel machine learning framework named the Generalizable Interpretable Neural Network (GINN), designed to achieve both high predictive accuracy and full model interpretability for survival prognosis following heart transplantation. GINN operates on structured clinical features using an additive representation approach, enabling explicit attribution of risk contributions from each clinical factor. We developed the GINN model based on comprehensive heart transplant data to predict one-year mortality and externally validated it on four independent cohorts. Using the same development data for training, the GINN model demonstrated robust predictive performance across three large-scale international transplant databases: United Network for Organ Sharing (UNOS 1994-2024, n = 144, 979 ), Eurotransplant registry ( n = 3, 061 ) and Scandiatransplant registry (1997-2018, n = 1, 546 ), achieving AUROC scores of 0.827, 0.789, and 0.776 respectively. These results indicate strong generalizability and cross-population transferability. Moreover, in a small dataset from the Department of Cardiac Surgery, Jiangxi Provincial People's Hospital (External-CN, n = 14 ), GINN maintained high risk identification capability with an AUROC of 0.821. The model constructed risk response functions based on nine key clinical variables, elucidating the marginal effects of donor and recipient age, donor function, preoperative support measures, and diagnostic types on postoperative risk. The findings suggest that GINN offers excellent generalization across geographic and sample-scale domains while maintaining predictive accuracy and providing stable and traceable risk explanations on structured clinical tabular data.