Neural Cox Model for Liver Transplant

Liver transplantation is a crucial for patients with end-stage liver disease and predicting post-transplant survival is a complex challenge. Deciding which patient receives a transplant or not is an important and hard medical decision that takes into account many factors. The aim of this thesis is to investigate a model capable of forecasting survival trends in patients subjected to liver transplant. In particular, we exploit a hybrid neural Cox proportional hazards model. To the best of our knowledge, this approach is novel and provides promising results. Indeed, this model is designed in collaboration with Sant'Orsola hospital of Bologna to be integrated within the decision process of surgeons. Our approach, which provides a data driven estimation of the the survivability function associated to patients who do not receive a transplantation, is compared with the more classical Cox survival model. More specifically, we compare our approach to two widely used libraries, namely, scikit-survival and lifelines. The results obtained on a synthetic dataset prove that the neural cox model effectively compares the classical model. We believe that the integration of deep learning with classical statical approaches can surpass the limitation of both approaches: on the one hand, classical cox model can be simplistic in describing complex relationships that neural networks can instead capture; while on the other hand, the use of pre existing survival framework allows to obtain a more transparent process within the realm of deep learning.