Machine learning models for predicting patients survival after liver transplantation

Abstract

In our work, we have built models predicting whether a patient will lose an organ after a liver transplant within a specified time horizon. We have used the observations of bilirubin and creatinine in the whole first year after the transplantation to derive predictors, capturing not only their static value but also their variability. Our models indeed have a predictive power that proves the value of incorporating variability of biochemical measurements, and it is the first contribution of our paper. As the second contribution we have identified that full-complexity models such as random forests and gradient boosting lack sufficient interpretability despite having the best predictive power, which is important in medicine. We have found that generalized additive models (GAM) provide the desired interpretability, and their predictive power is closer to the predictions of full-complexity models than to the predictions of simple linear models.