Predicting short-term mortality in sepsis-associated acute kidney injury: an explainable machine learning model with a web-based prototype

Abstract Geriatric Nutritional Risk Index (GNRI) is associated with prognosis in sepsis-associated acute kidney injury (S-AKI). However, its nonlinear relationship with mortality and its optimal integration into multivariable prediction models remain unclear. Machine learning (ML) methodologies, capable of capturing nonlinear relationships among variables, hold promise for enhancing the accuracy of short-term mortality risk prediction. This study included 773 critically ill S-AKI patients receiving continuous renal replacement therapy (CRRT). Key predictors were selected from 25 candidate variables using LASSO regression. A panel of candidate machine learning algorithms was evaluated for predicting 28-day all-cause mortality after LASSO-based feature selection; the final model was selected on test-set performance. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), calibration (Hosmer–Lemeshow test, Brier score), and decision curve analysis (DCA). The final model was interpreted using SHapley Additive exPlanations (SHAP) and deployed as a web-based prototype. The logistic regression model was selected as the final model and achieved the highest test-set AUC (0.765, 95% CI: 0.706–0.824) among all candidate algorithms, retaining nine predictors identified by LASSO regression. SHAP analysis identified SOFA_CRRT and serum creatinine as the most important predictors, with GNRI ranking seventh and showing a pronounced nonlinear association with mortality; risk increased sharply when GNRI was below 75, whereas a protective association was observed above 90. Decision curve analysis indicated net clinical benefit across a threshold probability range of 0.1 to 0.6. Calibration was suboptimal (Hosmer–Lemeshow test p = 0.004), which we acknowledge as a limitation. The model was deployed as a publicly accessible web-based prototype ( https://freergq.shinyapps.io/GNRISHAP_Model/ ) for research-use risk assessment. The interpretable model predicts 28-day mortality and identifies a GNRI threshold (< 75) associated with elevated risk. The web-based prototype supports research-use risk assessment; prospective external validation is required before any clinical application.