Utilizing Deep Learning and RDF to Predict Heart Transplantation Survival

Dennis Medved, Johan Nilsson, Pierre Nugues

Abstract: In this paper, we describe the conversion of three different heart transplantation data sets to a Resource Description Framework (RDF) representation and how it can be utilized to train deep learning models. These models were used to predict the outcome of patients both pre- and post-transplant and to calculate their survival time. The International Society for Heart & Lung Transplantation (ISHLT) maintains a registry of heart transplantations that it gathers from grafts performed worldwide. The American organization United Network for Organ Sharing (UNOS) and the Scandinavian Scandiatransplant are contributors to this registry, although they use different data models. We designed a unified graph representation covering these three data sets and we converted the databases into RDF triples. We used the resulting triplestore as input to several machine learning models trained to predict different aspects of heart transplantation patients. Recipient and donor properties are essential to predict the outcome of heart transplantation patients. In contrast with the manual techniques we used to extract data from the tabulated files, the RDF triplestore together with SPARQL, enables us to experiment quickly and automatically with different combinations of features sets, to predict the survival, and simulate the effectiveness of organ allocation policies.