Risk Stratification for Graft Failure After Pediatric Heart Transplantation Using Random Survival Forests

Prediction in pediatric heart transplantation is challenging due to clinical heterogeneity. This presents the dilemma of modeling non-linear relationships and complex variable structures. It may be overcome by survival-based machine learning techniques designed to accommodate time to event outcomes, contrary to traditional approaches.

In this study, we developed and evaluated a Random Survival Forest model to predict graft failure among pediatric heart transplant recipients using United Network for Organ Sharing registry data (post-July 2016). The primary outcome was graft failure.

Data was divided randomly into training (n = 3121) and testing (n = 1041) sets in a 75:25 ratio, and the scikit-survival implementation of the Random Survival Forest algorithm was chosen for this analysis. On the held-out test set, the final model achieved a Harrell's C-index of 0.60, and recorded 1- and 2-year Time-dependent Area Under the Curve (AUC) of 0.636 and 0.630 respectively; the 1- and 2-year brier scores were 0.078 and 0.097 respectively. Risk stratification analysis showed that the model-derived risk separated pediatric patients into low- and high-risk groups, with significant divergence in Kaplan-Meier survival curves (log-rank p = 0.003) and 1-year estimated survival of 92.8% and 88.1%. Clinical utility of the model was further evaluated using decision curve analysis at 1-year post transplant, the model suggested greater net benefit.

These results suggest that, even in the absence of high discrimination, machine learning approaches like Random Survival Forests can uncover clinically relevant patterns and may support future development of a framework for risk-adapted monitoring in pediatric heart transplantation.