Textual Annotation in the Prediction of Heart and Lung Transplantation Outcomes from Donor Data.

In this paper, we analyze a register of donated organs from Northern Europe and we build models to predict the future of donated hearts and lungs. More precisely, we trained models to predict if a heart or lungs will be used in a transplantation and, if transplanted, the one- and five-year survival of the recipient. We report our experiments, where we only used donor data consisting of categorical, numeric, and textual variables, and the results we obtained. Overall, for the organ transplantation and the five-year survival, our models reached macro F1 scores of respectively 0.77 for hearts and 0.68 for lungs, while remaining around 0.5 for the one-year survival. With the exception of one-year survival, this shows the contribution of the donor data to the organ transplantation and recipient survival. The textual annotations, while having in themselves a predictive capacity, do not improve the scores obtained with the other features. We finally carried out an analysis of the significant textual features that could explain the prediction or improve its interpretability.Clinical relevance-Organ donation is documented in registers in the form of numeric, categorical, and textual data. In this paper, we show that donor data from a Northern Europe register can predict if donated hearts or lungs will be transplanted with macro F1s of respectively 0.77 and 0.68.