Risk prediction of sepsis-associated acute kidney injury: development, validation of a machine learning model with multicenter data.

Abstract

Background: Sepsis-associated acute kidney injury (SA-AKI) is a common and severe complication in intensive care unit (ICU) patients, with high incidence and mortality. Conventional diagnostic indicators, such as serum creatinine and urine output, have limitations in early identification. Existing machine learning (ML)-based prediction models are mostly single-center and lack interpretability, highlighting the clinical need for reliable and applicable early prediction tools.

Methods: Data were collected from the Medical Information Mart for Intensive Care IV (MIMIC-IV), eICU Collaborative Research Database (eICU-CRD), and Bobai County People's Hospital ICU (BB-ICU). Eligible patients were those meeting the Sepsis-3.0 criteria, with SA-AKI as the primary outcome. Ten core variables were selected from 83 indicators measured within 24 h of ICU admission through a four-step screening process. Six ML algorithms were optimized using grid search and five-fold cross-validation. Model performance was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). The optimal model was interpreted via SHAP, externally validated, and translated into a web-based calculator using the Shiny framework.

Results: The incidence of SA-AKI was 45.1% in MIMIC-IV, 28.1% in eICU-CRD, and 34.1% in BB-ICU. The XGBoost model performed best, with an internal area under the curve (AUC) of 0.72 and external validation AUCs of 0.82 (eICU-CRD) and 0.84 (BB-ICU). The XGBoost model also showed balanced classification metrics (sensitivity 0.73, specificity 0.60), a calibration curve close to the diagonal, and the highest net benefit in DCA. SHAP analysis identified partial thromboplastin time (PTT), creatinine, 24-hour urine output, and SAPSII as the key predictive factors.

Conclusions: The ML model developed in this study demonstrates good discriminative ability, calibration, and clinical applicability. SHAP improves the model's interpretability, and the web-based tool supports early SA-AKI risk stratification. This study provides a technical basis for addressing the clinical challenge of early SA-AKI prediction.