Explainable Artificial Intelligence for Predictive Modeling in Nephrology: A Scoping Review

Abstract Globally, kidney-related illnesses affect over 850 million people, with acute kidney injury impacting 13.3 million individuals annually and chronic kidney disease contributing significantly to morbidity and mortality. While machine learning models show promise for predicting kidney disease outcomes, their "black box" nature limits clinical adoption. Explainable artificial intelligence (XAI) aims to address this gap by making model decisions interpretable to clinicians. This scoping review synthesizes evidence on XAI for predicting kidney disease outcomes, assessing clinical applicability and methodological quality. Following PRISMA guidelines, searches were conducted in Scopus, Web of Science, and EBSCOHost (January 2019–December 2025). Studies were included if they used formal XAI techniques (SHAP, LIME, or equivalent) for predictive modeling of kidney outcomes (AKI, CKD, mortality, or kidney failure) in adult populations. Quality appraisal used the Effective Public Health Practice Project (EPHPP) tool. Of 395 identified records, 18 studies met inclusion criteria. Quality appraisal found 12 studies (67%) to be high quality. XGBoost was the dominant algorithm (67%), achieving AUC values of 0.78–0.95. SHAP was the primary explainability method (83%), consistently identifying serum creatinine, estimated glomerular filtration rate, age, and diabetes as key predictors. However, 94% of studies were retrospective, 67% were single-center, and 40% remained at proof-of-concept stage without external validation. No studies reported clinical workflow integration. Explainable AI models demonstrate good predictive accuracy for kidney disease outcomes but face substantial translational barriers. Prospective, multi-center studies and strategies to move proof-of-concept models into clinical practice are urgently needed.