Prediction of Moderate-to-Severe Sepsis-Associated Acute Kidney Injury Using a Dual-Timepoint Machine Learning Model: Development, Multiregional Validation, and Clinical Deployment Study.

Abstract

Background: Sepsis-associated acute kidney injury (SA-AKI) is a frequent and life-threatening complication in patients in the intensive care unit (ICU), significantly increasing both mortality rates and the risk of chronic kidney dysfunction. However, existing prediction models have often focused on overall risk and lack severity-based stratification, which limits their clinical applicability.

Objective: This study aimed to identify critical time points in SA-AKI progression development and validate dynamic, stratified machine learning prediction models for moderate-to-severe (Kidney Disease: Improving Global Outcomes guideline stages 2-3) SA-AKI through multicenter, multiregional external validation, ultimately deploying them as publicly accessible, interpretable clinical decision support tools.

Methods: This study used three independent ICU databases: Medical Information Mart for Intensive Care-IV v3.0 (n=12,842; model development and internal validation), electronic ICU collaborative research database v2.0 (n=15,767; North American multicenter external validation), and the First Affiliated Hospital of Hainan Medical University ICU (n=210; Chinese single-center external validation). We identified 48 hours (acute phase) and 7 days (subacute phase) as critical time points. Based on clinical data from the first 24 hours of ICU admission, we used a two-stage feature selection process combining light gradient boosting machine (LightGBM) and Shapley additive explanation (SHAP) cross-validation analysis with clinical expert review, followed by modeling using 8 machine learning algorithms. The optimal model was selected based on the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis. Internal validation used 5-fold cross-validation, while external validation and subgroup analyses assessed generalizability across different regions and populations. SHAP values and partial dependence plots were used to interpret the influence of key features on predictions.

Results: Our dual-timepoint LightGBM model demonstrated robust predictive performance. For the 48-hour prediction task, the model achieved an AUC of 0.839 (95% CI 0.824-0.854) in the internal test set, with AUCs of 0.770 (95% CI 0.762-0.779) and 0.793 (95% CI 0.726-0.856) in the external validation cohorts, respectively. For the 7-day prediction task, the corresponding AUCs across the three cohorts were 0.834 (95% CI 0.818-0.850), 0.720 (95% CI 0.711-0.729), and 0.773 (95% CI 0.687-0.851), respectively. Subgroup analyses confirmed robust model performance across different age, gender, and comorbidity subgroups. SHAP analysis identified urine output, mechanical ventilation, Sequential Organ Failure Assessment score, creatinine, Glasgow Coma Scale score, and nephrotoxic drug use as core predictive features. Decision curve analysis confirmed that LightGBM provided consistent clinical benefit across different threshold ranges. The optimal LightGBM model was deployed as a publicly accessible web-based prediction app with integrated SHAP interpretability.

Conclusions: This study developed and validated a dynamic, stratified prediction system that provides stage-specific risk assessment for moderate-to-severe SA-AKI. The system underwent rigorous multiregional, multicenter validation and was translated into an interpretable clinical decision support tool, providing a scientific foundation for precision management.