Project 465728

An accurate, equitable estimation of pre-transplant mortality is an essential component of fairly prioritizing patients for liver transplant. For the past 20 years, transplant centers in the United States and Canada have relied on the Model for End-Stage Liver Disease (MELD) and MELD-Na as the scores by which to prioritize patients for transplant. These scores have drawn criticism for yielding sex-based inequities in rates of pre-transplant mortality, failing to robustly adapt to the changing demographics of the transplant waitlist, and inaccurately modelling the disease trajectories of patients with nonstandard progressions of disease. In this work, we apply modern methods from machine learning to build a model that is more accurate and equitable than the MELD-Na. We hypothesize that, unlike the MELD-Na which is a linear function of patient biomarkers evaluated at a single point in time, neural network models that incorporate longitudinal changes in patient biomarkers will better be able to model each patient's risk of pre-transplant mortality. Additionally, where related work proposes to incorporate sex into the MELD-Na score to reduce sex-based disparities, we will leverage larger groups of features to equitably predict mortality without making assumptions about which groups are adversely affected.