Association between stress hyperglycemia ratio and all-cause mortality in patients with acute kidney injury: a retrospective analysis of the MIMIC-IV database using machine learning

The stress hyperglycemia ratio (SHR) reflects relative hyperglycemia levels. Prior studies indicate that SHR is related to the prognosis of certain diseases. This study investigated the predictive accuracy of SHR for in-hospital and intensive care unit (ICU) mortality in acute kidney injury (AKI) patients. We analyzed 4,418 AKI patients from the Medical Information Mart for Intensive Care (MIMIC-IV) database. SHR was calculated as admission blood glucose (mg/dl) divided by (28.7 × HbA1c(%) − 46.7). Patients were stratified into four groups based on SHR levels. Restricted cubic splines (RCS) and Cox proportional hazards models were employed to assess the association between SHR and mortality, and Kaplan-Meier survival analysis was conducted to compare mortality rates stratified by SHR groups. Subgroup analyses evaluated the consistency of these associations across clinically relevant populations. Machine learning algorithms (Boruta feature selection, logistic regression, random forest) were employed to build predictive models. Among 4,418 patients (36.03% female), in-hospital and ICU mortality rates were 6.88% and 4.48%, respectively. A U-shaped relationship was observed between SHR and mortality (nonlinear p = 0.015 for in-hospital, p = 0.048 for ICU mortality). Kaplan-Meier analysis revealed significant survival differences across SHR groups (log-rank p < 0.01). Cox regression showed that each unit increase in SHR was associated with 23% higher in-hospital mortality (HR = 1.23, 95% CI: 1.09–1.40) and 47% higher ICU mortality (HR = 1.47, 95% CI: 1.25–1.74). Machine learning models achieved AUCs of 0.76–0.77 for mortality prediction. In critically ill AKI patients, higher SHR is associated with increased mortality after adjustment for key covariates. However, its modest standalone discriminative capacity indicates that SHR is best combined with other clinical variables rather than used alone.