Development and validation of a multidimensional machine learning model for predicting delayed graft function in...

Delayed graft function (DGF) remains a significant complication following deceased donor kidney transplantation. This study aimed to develop and validate a multidimensional machine learning model for predicting DGF by integrating clinical data, machine perfusion parameters, donor scores, and histopathological scores.

A retrospective analysis was conducted on 961 deceased donor kidney transplant recipients from January 2019 to December 2021. The dataset was stratified by the target variable and randomly divided into training (80%) and independent testing (20%) cohorts. Fifteen data combinations across four dimensions were evaluated using six machine learning algorithms. Model performance was assessed using AUC, calibration curves, and decision curve analysis. SHAP analysis was employed for feature interpretation.

The optimal model combining clinical supplementary data with donor scores and histopathological scores using LightGBM achieved an AUC of 0.895 (95% CI 0.806-0.985) in the testing cohort, with accuracy of 89.1%, sensitivity of 59.1%, and specificity of 93.0%. Clinical data served as the foundational predictive dimension (AUC = 0.874), while histopathological scores provided significant incremental value (ΔAUC = 0.016). Among the top five ranked combinations, four included histopathological scores. Gain-based feature importance identified donor score, renal tubular necrosis, and total pathological score as the leading predictors, with pathological indicators occupying four of the top five positions; SHAP analysis further revealed blood urea nitrogen and donor score as the most influential individual features, with renal tubular necrosis being the top-ranked histopathological predictor.

This multidimensional machine learning model demonstrates excellent predictive performance for DGF. The "scenario-model-threshold" clinical decision framework provides a practical tool for guiding donor kidney assessment and immunosuppressive therapy decisions in diverse clinical contexts.