Interpretable Machine Learning for Mortality Prediction in S-AKI Patients Undergoing Hemodialysis

Abstract

This study developed a machine learning model to predict in-hospital mortality risk among ICU patients with sepsis-associated acute kidney injury (S-AKI) undergoing hemodialysis. A retrospective analysis of 1,467 patients from the MIMIC-IV database and 226 from the eICU-CRD database was conducted, with models evaluated internally and externally. The RF model achieved excellent performance, with AUROCs of 0.798 (95% CI: 0.754–0.843) and 0.790 (95% CI: 0.723–0.857) in internal and external validations, respectively. Decision curve analysis indicated a net benefit of ~0.2 at a 10% mortality threshold, demonstrating good clinical applicability. SHAP analysis identified prothrombin time, APS III, systolic blood pressure, mean arterial pressure, and respiration rate as key predictors, with increased mortality risk associated with prothrombin time >10s, APS III >80, Nibp_systolic <110 mmHg, and Nibp_mean <70 mmHg. This model offers potential for supporting prognosis management and individualized treatment in clinical practice.