Utilizing Machine Learning to Identify Predictors of Corticosteroid Discontinuation 1-Year After Adult Heart Transplant.

Abstract

Objective: To use machine learning methods to identify factors associated with corticosteroid (CS) discontinuation 1 year after adult heart transplantation (HT).

Design: Retrospective, observational, cohort study.

Data source: This study used data from the United Network for Organ Sharing (UNOS) database.

Patients: We included adults (age ≥ 18 years) who underwent their first HT between January 2000 and December 2023 in the UNOS database with follow-up through December 2024, who were discharged on a CS.

Measurements: We divided the cohort into those with or without CS at 1-year post-transplant follow-up. We used the eXtreme Gradient Boosting (XGBoost) algorithm to build a model predicting CS discontinuation at 1 year. Relevant recipient, donor, and transplant variables were included to train the model, with Shapley Additive Explanations (SHAP) used to identify and interpret the most important and predictive features.

Results: We identified 72,730 HT recipients; 13,017 (17.9%) had CS discontinued within the first year. Compared with the CS cessation group, those who continued CS were more likely to have had a lower BMI, lower ischemic etiology of cardiomyopathy, lower intra-aortic balloon pump (IABP) and left ventricular assist device (LVAD) use before HT, better renal function, and sustained longer donor ischemic time. Model performance was strong, with an area under the curve of 0.854 (95% confidence interval: 0.848-0.861). Lower average transplant center volume (number of transplants per year), shorter donor ischemic time, and LVAD use at transplant predicted CS discontinuation.

Conclusions: In a large national database, utilizing novel ML modeling techniques, we identified annual transplant center volume, donor ischemia time, and LVAD use at the time of HT as the best predictors of CS discontinuation 1 year after HT.