Prediction of severe sepsis-associated acute kidney injury incorporating immune-inflammatory profiles: development and validation of a machine learning model in a multicenter prospective cohort study

Abstract

Introduction Severe sepsis-associated acute kidney injury (SA-AKI) is a prevalent and life-threatening complication in critically ill patients, leading to increased mortality and a heightened risk of chronic kidney dysfunction. Current prediction models for severe SA-AKI have largely overlooked the inclusion of immune and inflammatory indicators, which more accurately represent the underlying pathophysiology of sepsis—the dysregulated host response to infection. Methods Using a multicenter prospective cohort of 1,715 septic patients from five independent ICUs, we developed and validated a machine learning model integrating immune-inflammatory profiles to predict progression to severe SA-AKI, defined as KDIGO stage 2 or 3 per the Acute Disease Quality Initiative consensus criteria (occurring in 670 patients [39.1%]). Immune-inflammatory variables and routine clinical data were collected within 24 hours of sepsis diagnosis. Six machine learning algorithms were trained and evaluated using area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, calibration, and decision curve analysis (DCA). Cross-institutional stability was evaluated by leave-one-center-out cross-validation (LOCO-CV) sensitivity analysis. Model interpretability was assessed using Shapley Additive Explanations (SHAP). Results Among the six models, random forest achieved the highest sensitivity (0.881) while maintaining strong discriminative ability (AUC 0.912, 95% CI 0.879-0.941, specificity 0.794) in the validation set. The final model incorporated nine clinically accessible variables: SOFA score, CD38+CD8+ T-cell count, tumor necrosis factor-alpha, interleukin-6, immunoglobulin G, central venous oxygen saturation, bilirubin, and histories of chronic kidney disease and chronic cardiac insufficiency. The model exhibited adequate calibration (Brier score 0.115, calibration slope 1.207), positive net benefit on DCA, and good interpretability. LOCO-CV confirmed consistent performance across the five centers, with a mean AUC of 0.900 (range 0.863–0.938), demonstrating cross-institutional robustness. Conclusion An interpretable random forest model incorporating immune-inflammatory profiles accurately predicted progression to severe SA-AKI in critically ill patients with sepsis. Following external validation and further clinical implementation, this immune-inflammatory profile-based model may offer an effective and practical strategy for early risk stratification. Clinical trial registration http://www.chictr.org.cn, identifier ChiCTR2300074175.