Donated livers are assigned to eligible matches among patients on the transplant list according to a sickest-first policy, which ranks patients by their score via the Model for End-stage Liver Disease (MELD). While the MELD score is indeed predictive of mortality on the transplant list, the score was fit with just three features, for a different task (outcomes from a shunt insertion procedure), and on a potentially un-representative cohort. These facts motivate us to investigate the MELD score, assessing its predictive performance compared to modern ML techniques and the fairness of the allocations vis-a-vis demographics such as gender and race. We demonstrate that assessing the quality of the MELD score is not straightforward: waitlist mortality is only observed for those patients who remain on the list (and don't receive transplants). Interestingly, we find that MELD performs comparably to a linear model fit on the same features and optimized directly to predict same-day mortality. Using a wider set of available covariates, gradient-boosted decision trees achieve .926 AUC (compared to .867 for MELD-Na). However, some of the additional covariates might be problematic, either from a standpoint of procedural fairness, or because they might expose the process to possible gaming due to manipulability by doctors.
展开▼
