Prediction of Donor Heart Acceptance for Transplant and Its Clinical Implications: Results From The Donor Heart Study

Despite a shortage of potential donors for heart transplant in the United States, most potential donor hearts are discarded. We evaluated predictors of donor heart acceptance in the United States and applied machine learning methods to improve prediction.

We included a nationwide (2005-2020) cohort of potential heart donors in the United States (n=73 948) from the Scientific Registry of Transplant Recipients and a more recent (2015-2020) rigorously phenotyped cohort of potential donors from DHS (Donor Heart Study; n=4130). We identified predictors of acceptance for heart transplant in both cohorts using multivariate logistic regression, incorporating time-interaction terms to characterize their varying effects over time. We fit models predicting acceptance for transplant in a 50% training subset of DHS using logistic regression, least absolute shrinkage and selection operator, and random forest algorithms and compared their performance in the remaining 50% (test) of the subset.

Predictors of donor heart acceptance were similar in the nationwide and DHS cohorts. Among these, older age (P value for time interaction, 0.0001) has become increasingly predictive of discard over time while other factors, including those related to drug use, infection, and mild cardiac diagnostic abnormalities, have become less influential (P value for time interaction, <0.05 for all). A random forest model (area under the curve, 0.908; accuracy, 0.831) outperformed other prediction algorithms in the test subset and was used as the basis of a novel web-based prediction tool.

Predictors of donor heart acceptance for transplantation have changed significantly over the last 2 decades, likely reflecting evolving evidence regarding their impact on posttransplant outcomes. Real-time prediction of donor heart acceptance, using our web-based tool, may improve efficiency during donor management and heart allocation.