Predicting Liver Utilization Rate and Post- Transplant Outcomes from Donor Text Narratives with Natural Language Processing

Abstract

Liver transplantation is a critical, life-saving treatment option for patients with terminal liver disease. Despite an organ shortage, many donated livers are discarded for reasons such as poor organ condition and physical incompatibility with a recipient. Current clinical models for liver risk assessment only utilize tabular data and result in poor precision and recall. Critical information relevant to this decision-making is likely included in the free-text clinical notes from donor evaluations that contain pertinent medical and social history of the donor that is currently unavailable in tabular data sources. This article describes the development of a model using these free-text clinical notes using a variety of Natural Language Processing (NLP) and machine learning (ML) techniques to predict the outcomes of three key metrics: 1) liver utilization rate, 2) 30-day mortality rate, and 3) 1-year mortality rate. The free-text narratives were useful for predicting liver utilization, with an associated area under the curve (AUC) score of 0.81, but were not useful for predicting both mortality outcomes, with associated AUC scores of 0.53 and 0.52, for 30-day and 1-year mortality, respectively. Using a locally interpretable model-agnostic explanations (LIME) algorithm, key phrases, like “dcd” and “alcohol” were found to be associated with unutilized livers, while “brain” and “heroin” were associated with utilized livers. Based on these findings, modeling donor text narratives may substantially contribute to improved decision-making and outcomes of liver transplantation.