Background: Center variances in usage of technical variant grafts for pediatric liver transplant may be contributing to the variance in waitlist mortality. Previous work demonstrates an underutilization of deceased donor technical variant grafts in the USA and that many children die after receiving and declining acceptable grafts for transplant. We aim to model the acceptance and graft selection of deceased donor livers for pediatric transplantation.

Methods: We obtained data from 192,814 deceased donor liver offers received by pediatric liver transplant candidates listed from 2007 – 2020 from the OPTN registry. We selected clinically significant variables that determined offer acceptance and graft type selection and were available at the time of the organ offer. We performed time-based training, validation, and held-out testing folds to measure and compare the performance of multiple statistical and machine learning predictive models.

Results: The predictive performance for offer acceptance models ranged from the Random Forests at c-stat 0.888 (very good) to Nearest Neighbor at 0.710 (good). The predictive performance for graft type models ranged from XGBoost at c-stat 0.971 (excellent) to Decision Tree at 0.951 (excellent). Historical center experience variables showing the volumes and rates of technical variant graft usage were listed among the top 25 most influential variables in the models. Candidates that died waiting had previously declined acceptable offers and failed to utilize technical variant grafts.

Discussion: Historical center-based organ offer acceptance and graft selection practices are significantly predictive of their future practices. Candidates that died waiting while declining acceptable offers and failing to utilize technical variant grafts were often listed at centers lacking experience with technical variant grafts.