Prediction of 1-Year Mortality After Heart Transplantation in Adults with Congenital Heart Disease with Machine Learning Models

Abstract

PurposeMachine learning (ML) can be used assisting clinical decision-making. We developed a ML model for the prediction of 1-year mortality after heart transplantation (HT) in adults with congenital heart disease.MethodsThe United Network for Organ Sharing (UNOS) database was queried from 2000-2020 for adults with congenital heart disease who underwent isolated HT and had at least 1-year follow-up. The cohort was randomly split in derivation (70%) and validation (30%) dataset used to train and test a CatBoost decision tree model respectively. The primary outcome was 1-year mortality. Explainability analysis with Shapley Additive exPlanations (SHAP) was performed.ResultsA total of 1,032 recipients were included in the study (35±13 years, 61% males). At 1-year after HT, there were 205 deaths. After feature selection, area under the curve and predictive accuracy for the final ML model were 0.82 and 77% respectively. The impact of each model variable for each individual prediction in the validation dataset is represented by its SHAP value (Figure 1).ConclusionA ML model developed using data from the UNOS database showed satisfactory predictive accuracy for 1-year mortality after HT in adults with congenital heart disease. Explainability analysis helps interpret the results in a clinical manner. The nature of the selected features along with their effect on the prediction outcome can have significant implications in optimal patient selection, organ allocation and prognostication. Machine learning (ML) can be used assisting clinical decision-making. We developed a ML model for the prediction of 1-year mortality after heart transplantation (HT) in adults with congenital heart disease. The United Network for Organ Sharing (UNOS) database was queried from 2000-2020 for adults with congenital heart disease who underwent isolated HT and had at least 1-year follow-up. The cohort was randomly split in derivation (70%) and validation (30%) dataset used to train and test a CatBoost decision tree model respectively. The primary outcome was 1-year mortality. Explainability analysis with Shapley Additive exPlanations (SHAP) was performed. A total of 1,032 recipients were included in the study (35±13 years, 61% males). At 1-year after HT, there were 205 deaths. After feature selection, area under the curve and predictive accuracy for the final ML model were 0.82 and 77% respectively. The impact of each model variable for each individual prediction in the validation dataset is represented by its SHAP value (Figure 1). A ML model developed using data from the UNOS database showed satisfactory predictive accuracy for 1-year mortality after HT in adults with congenital heart disease. Explainability analysis helps interpret the results in a clinical manner. The nature of the selected features along with their effect on the prediction outcome can have significant implications in optimal patient selection, organ allocation and prognostication.