Prelisting predictions of early postoperative survival in infant heart transplantation using classification and regression tree analysis

Infants listed for heart transplantation experience high waitlist and early post-transplant mortality, and thus, optimal allocation of scarce donor organs is required. Unfortunately, the creation and validation of multivariable regression models to identify risk factors and generate individual-level predictions are challenging. We sought to explore the use of data mining methods to generate a prediction model. CART analysis was used to create a model which, at the time of listing, would predict which infants listed for heart transplantation would survive at least 3 months post-transplantation. A total of 48 infants were included; 13 died while waiting, and six died within 3 months of heart transplant. CART analysis identified RRT, blood urea nitrogen, and hematocrit as terminal nodes with alanine transaminase as an intermediate node predicting death. No patients listed on RRT (n = 10) survived and only three of 12 (25%) patients listed on ECLS survived >3 months post-transplant. CART analysis overall accuracy was 83%, with sensitivity of 95% and specificity 76%. This study shows that CART analysis can be used to generate accurate prediction models in small patient populations. Model validation will be necessary before incorporation into decision-making algorithms used to determine transplant candidacy.