Machine Learning for Predicting Waitlist Mortality in Pediatric Heart Transplantation.

Background: Waitlist mortality remains a critical issue for pediatric heart transplant (HTx) candidates, particularly for candidates with congenital heart disease. Listing center organ offer acceptance practices have been identified as a factor influencing waitlist outcomes. We utilized machine learning (ML) to identify factors associated with waitlist mortality, combining variables associated with institutional offer acceptance practices as well as candidate-specific risk factors.

Methods: We analyzed the Organ Procurement and Transplantation Network database for pediatric HTx candidates listed between 2010 and 2020. Various statistical and ML models were employed to identify predictors of waitlist mortality or clinical deterioration leading to waitlist removal. The dataset was split into training (82%) and testing (18%), and the final model was selected based on predictive performance. SHAP values were used to assess variable importance.

Results: Among 5523 pediatric candidates, overall waitlist mortality was 9.8%. The CatBoost model achieved the highest predictive performance with an AUC-ROC score of 0.74 and a recall score of 0.75. Key predictors included candidate diagnosis, age/size, ventilator use, eGFR, serum albumin, ECMO, and institutional factors such as high offer refusal rates and low transplant volume.

Conclusions: Institutional organ offer acceptance practices influence waitlist outcomes for pediatric HTx candidates. Centers with higher organ refusal rates are associated with worse outcomes, independent of candidate-specific risk factors, underscoring the need for standardizing organ acceptance criteria across institutions to reduce variability in decision-making and improve waitlist survival. Additionally, addressing modifiable risk factors such as malnutrition and renal dysfunction could further optimize patient outcomes.