Liver transplant donor-recipient matching with offline reinforcement learning

Abstract Despite advances in liver transplantation (LT), deciding when to transplant a patient within the context of high waitlist mortality, organ scarcity, and risk of graft failure, remains an ongoing challenge. Existing approaches focus on the static prediction of successful LT donor-recipient pairs without weighing the competing interests such as the risk of graft failure against the risk of waitlist mortality and how these risks change over time. Instead, we used an offline reinforcement learning (RL) approach to represent the problem as the optimization of the series of decisions to wait, delist, or transplant a candidate at different timepoints. Using waitlist trajectories for LT candidates from the national Scientific Registry of Transplant Recipients (SRTR) database, we trained a model resulting in the avoidance of 73% of donor-recipient pairs that led to graft failure or death, preservation of 93% of successful transplants, and potentially suitable donors were found for 47% of those patients that died on the waitlist. Notably, the analysis of decisions and post-transplant survival revealed that our model learned features suggestive of successful donor-recipient pairs. Overall, we demonstrate how RL-based approaches better portray real-world LT donor-recipient matching decisions, illustrating their potential as useful clinical tools.