Interpretable machine learning for chronic kidney disease prediction: Insights from SHAP and LIME analyses

Abstract

Chronic kidney disease (CKD) is a progressive condition requiring early detection for optimal patient outcomes. This study developed an interpretable machine learning framework using XGBoost with SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) for transparent CKD prediction. We evaluated the approach on two datasets: UAE Tawam Hospital data (n = 491) and UCI CKD data (n = 400). XGBoost with SMOTE optimization achieved 88.4% accuracy (AUC = 0.904) on the hospital dataset and 94.6% accuracy (AUC = 0.948) on the UCI dataset after Rigorous overfitting prevention through conservative hyperparameter ranges and performance monitoring ensured clinical credibility. SHAP analysis identified clinically relevant predictors: eGFRBaseline, HbA1c, and CholesterolBaseline for the hospital cohort, and specific gravity, hemoglobin, and serum creatinine for the UCI cohort. LIME provided complementary patient-level explanations that validated global SHAP patterns. The convergence between global and local interpretability methods confirms model reliability across diverse clinical contexts. This framework addresses the transparency barrier to machine learning adoption in healthcare while maintaining clinically realistic performance levels. The approach provides a foundation for integrating interpretable artificial intelligence into CKD screening and management workflows.