Machine Learning for Predicting Long-Term Cardiovascular Outcomes in Kidney Transplant Recipients

Cardiovascular disease is the leading cause of morbidity and mortality in kidney transplant recipients (KTRs). Conventional risk assessment tools often underperform in this population due to their unique cardiovascular milieu shaped by end-stage kidney disease and immunosuppression.

This study aimed to evaluate the performance of machine learning (ML) models in predicting major adverse cardiovascular events in KTRs.

We analyzed a data set of 518 adult KTRs from a single center (2015-2024) incorporating 1,014 demographic, clinical, laboratory, and imaging parameters in addition to ischemic work-up and revascularization details. After preprocessing, 44 features were used to train and validate ML models using AutoGluon with 5-fold cross-validation and permutation-based feature selection as well as domain knowledge. SHAP was used for model interpretation. The best performing ML model was compared to multivariable logistic regression in addition to a historical model from prior literature, that is, Soveri risk score.

Over a mean follow-up of 5.3 ± 2.3 years, 102 patients experienced major adverse cardiovascular events (mean age 52.9 ± 13.2 years; 56.9% male). The CatBoost model achieved the highest area under the receiver-operating characteristic curve (AUROC) (0.72; 95% CI: 0.607-0.811) and area under the precision-recall curve (0.36; 95% CI: 0.230-0.540), greater than the logistic regression (AUROC = 0.67; 95% CI: 0.559-0.772, area under the precision-recall curve = 0.31; 95% CI: 0.228-0.553). The differences were not statistically significant. Both models had significantly higher AUROC than the Soveri risk score. Top predictive variables included coronary artery disease, left atrial reservoir strain, dialysis duration, and left ventricular end-diastolic volume index.

ML models are feasible for predicting cardiovascular events in KTRs.