Association of Glycemic Variability, Stress Hyperglycemia Ratio, and Hemoglobin Glycation Index With 28-Day Mortality in Sepsis: A Multicenter Retrospective Study With Mediation and Machine-Learning Analyses

Background Glycemic dysregulation is a hallmark of sepsis, but the comparative prognostic value of different glycemic metrics remains unclear. We aimed to systematically compare the associations of Glycemic Variability (GV), Stress Hyperglycemia Ratio (SHR), and Hemoglobin Glycation Index (HGI) with mortality in sepsis and to develop a machine learning-based prediction model for mortality in sepsis. Methods This multicenter retrospective study included 7,260 adult patients with sepsis from the MIMIC-IV and NWICU databases. GV was calculated as the coefficient of variation of blood glucose; SHR and HGI were computed using established formulas. The primary outcome was 28-day all-cause mortality. Associations between glycemic indices and mortality were assessed using multivariable Cox regression, restricted cubic splines, mediation analysis, and subgroup analyses. An XGBoost model incorporating glycemic and clinical variables was developed, and variable importance was evaluated using SHAP (SHapley Additive exPlanations) values. Results After full adjustment, patients in the highest quintile of GV (HR = 1.62, 95% CI 1.33–1.97) and SHR (HR = 1.51, 95% CI 1.25–1.82) had significantly higher 28-day mortality, whereas those in the highest HGI quintile had lower mortality (HR = 0.74, 95% CI 0.61–0.91). Lactate partially mediated these associations. The XGBoost model demonstrated excellent performance (AUC = 0.798), and SHAP analysis identified GV, SHR, and HGI among the top predictors of mortality. These findings remained robust in sensitivity analysis using Random Forest imputation. Conclusion GV, SHR, and HGI are independent and complementary prognostic markers in sepsis. A multidimensional evaluation of glycemic dysregulation enhances risk stratification, and integrating these indices into machine learning models substantially improves predictive accuracy.