An interpretable XGBoost model for risk prediction of progression from sepsis-associated acute kidney injury to chronic kidney disease

To develop an interpretable machine learning (ML) model for predicting the risk of progression from sepsis-associated acute kidney injury (SA-AKI) to chronic kidney disease (CKD) guiding early stratified interventions. Using data from 1,315 SA-AKI patients [Medical Information Mart for Intensive Care IV (MIMIC-IV) database], we constructed an extreme gradient boosting (XGBoost) model with SHapley Additive exPlanations (SHAP) interpretability. Performance was evaluated by discrimination, calibration, and clinical utility [decision curve analysis (DCA)]. CKD incidence was 36.7% (median onset: 7.6 months). The XGBoost model achieved: superior discrimination [training area under the curve (AUC) 0.920; validation AUC 0.951 versus Sequential Organ Failure Assessment (SOFA) renal 0.616 and logistic regression (LR) 0.822], robust calibration, and clinical applicability. SHAP identified actionable thresholds (age >65, maximum serum creatinine >0.9 mg/dl) for early intervention. Feature stability analysis revealed a stage-dependent coefficient drift for serum creatinine (Δβ = +0.84), reflecting dynamic pathophysiology. Crucially, the model provides clinically interpretable outputs without requiring SHAP expertise, enabling seamless integration into workflows. Our model delivers personalized, interpretable CKD risk alerts for SA-AKI patients, empowering clinicians to stratify follow-up care. External validation is warranted to confirm generalizability.