Predicting 1-year Graft Survival in Liver Transplantation: A Machine Learning Study Using Registry Data From Germany and the United States

Background. Liver transplantation is a life-saving procedure, but predicting posttransplant outcomes remains challenging due to the complexity of preoperative data. Traditional models often fail to capture multidimensional risk factors, which may be better integrated through machine learning (ML). This study aimed to develop and evaluate ML models using preoperative data from national transplant registries in Germany and the United States. Methods. ML algorithms were trained and tested on data from 9007 adult liver transplantation recipients in Germany and 105 075 recipients from the US cohort. The analysis incorporated 19 common preoperative variables, including recipient demographics, donor characteristics, and laboratory values. The models were trained for 1-y graft survival, using multiple ML methods and 5-fold cross-validation. Results. The Random Forest Classifier achieved the highest predictive performance when trained on the German cohort alone (area under the curve = 0.69), with Model for End-Stage Liver Disease score and cold ischemia time as the strongest predictors. Validation of the model on the US cohort after merging the 2 data sets showed poor performance (area under the curve = 0.54), possibly due to differences in donor selection and recipient characteristics, including younger donors, higher drug use, and shorter ischemia times. Conclusions. Although ML demonstrated good predictive capabilities in the German cohort, it was not possible to generalize its performance on the external cohort from the United States, highlighting cohort-specific differences that hinder model transferability. With increasing development and utilization of ML models, validation in region-specific cohorts is encouraged before clinical implementation.