Abstract 14239: Risk Factors for One-year Mortality and Allograft Loss in Pediatric Heart Transplant Patients Using Machine Learning: Analysis of the Pediatric Heart Transplant Society Database

Abstract

Background: Pediatric heart transplant (TX) patients are at greatest risk of graft loss (GL) in the first year, which may increase due to accumulated risks in the pre-TX period. Machine learning (ML) may identify clinically relevant outcome characteristics without prior statistical assumptions. We assessed whether ML could improve risk assessment for one-year patient or graft survival in pediatric heart TX. Methods: Patients in the PHTS database from January 2010 to December 2016 and January 2017 to June 2019 were included in training (TRN) and testing (TST) sets, respectively. Primary outcome was time until death, GL, or censoring. Using TRN, we selected the 12 most important recipient predictors based on contributions to model fit by applying boosted decision trees. Proportional hazards, boosting, and oblique random survival forest models were developed using these predictors. Discrimination and calibration were compared at 1 year post-TX using TST. The contribution of each predictor to patient’s predicted log-hazard was estimated using SHapley Additive exPansion (SHAP) values. Results: TRN (N=2,686) included 1,368 (51%) patients with cardiomyopathy (CM), 1,247 (46%) with congenital heart disease (CHD), and 71 (2%) other. TST (N=1,101) included 509 (46%) with CM, 569 (52%) with CHD, and 23 (2.1%) other. Incidence of GL or death (95% CI) in TRN and TST was 8.13 (7.07, 9.25) and 9.04 (6.93, 11.3) events/100 person years. Among the three models fitted to TRN, boosting obtained the highest concordance index at 1-year post-TX in TST, 0.752 (95% CI 0.695-0.810), and exhibited no evidence of miscalibration. Primary diagnosis of non-CM was the strongest predictor according to the boosting model (mean SHAP 0.398), followed by cardiopulmonary bypass time (mean SHAP 0.308);( Figure 1 ). Conclusions: ML can identify salient predictor variables and develop generalizable risk prediction equations using data from pediatric TX patients and may improve patient selection.