Development of interpretable machine learning models for early diagnosis of sepsis-associated acute kidney injury

Abstract

Sepsis-associated acute kidney injury (S-AKI) is associated with high mortality in critically ill ICU patients, and early diagnosis enables timely targeted intervention. This retrospective dual-center study included 687 septic ICU patients for model development and 125 independent cases for external validation. Following missing value imputation via MICE and class balancing using SMOTE, L1-regularized least absolute shrinkage and selection operator (LASSO) regression (λ_min = 0.0066) was used to select 19 routine clinical predictors. Ten machine learning algorithms were developed and systematically compared. XGBoost delivered balanced overall predictive performance with good calibration, yielding an AUC of 0.91 and an internal validation sensitivity of 0.86. Consistent performance on the external dataset confirmed model stability, and decision curve analysis demonstrated sustained net clinical benefit. SHAP analysis identified blood urea nitrogen as the primary risk biomarker and revealed nonlinear and synergistic relationships among core indicators. Multiple sensitivity analyses further confirmed model robustness. Developed using real-world ICU data from Northwest China, this interpretable XGBoost model partially alleviates the black-box disadvantage of traditional machine learning pipelines. It may serve as an auxiliary tool for early S-AKI risk stratification among regional sepsis patients, which may provide risk clues for clinicians to implement targeted intensive monitoring.