Simultaneous assessment of stress hyperglycemia ratio and glucose variability to predict all-cause mortality in sepsis patients across different glucose metabolic states: an observational cohort study with interpretable machine learning approach

Background: Stress hyperglycemia ratio (SHR) and glycemic variability (GV) reflect acute glucose elevation and fluctuation, which are associated with adverse outcomes in patients with some diseases. However, the relationship between combined assessment of SHR and GV and mortality risk in sepsis remains unclear. This study aims to investigate the associations of SHR, GV, and their combination with sepsis mortality among individuals with different glucose metabolic states, and to develop a mortality prediction model using machine learning (ML) models. Methods: Patients with sepsis were screened in the MIMIC-IV database, stratified into normal glucose regulation (NGR), prediabetes mellitus (Pre-DM), and diabetes mellitus (DM) groups based on glucose metabolic status. Associations with mortality were analyzed using Kaplan–Meier (KM) curves, Cox proportional hazards model, restricted cubic splines (RCS), and landmark analyses. Five ML algorithms were employed for prediction, with SHapley Additive explanations (SHAP) interpreting key predictors. Results: A total of 4838 patients were enrolled, with a median age of 68 years. Overall, 641 patients (13.2%) died in the ICU, and 936 patients (19.3%) died within 28 days after admission to the ICU. In NGR patients, combined high SHR (>1.23; highest tertile) and high GV (>28.56; highest tertile) – determined based on tertile distribution – conferred the highest 28-day mortality risk (HR = 2.06, 95% CI: 1.40–3.04). Pre-DM patients with low SHR/high GV (SHR < 1.23, GV > 28.56) showed the greatest 28-day mortality risk (HR = 2.45, 95% CI: 1.73–3.48). DM patients with high SHR/low GV (SHR > 1.23, GV < 28.56) had the highest 28-day mortality risk (HR = 1.46, 95% CI: 1.06–2.01). Machine learning models – particularly XGBoost (AUC: 0.746), Random Forest (AUC: 0.776), and Logistic Regression (AUC: 0.776) – demonstrated the strongest predictive performance for these endpoints. Conclusion: The combined assessment of SHR and GV may provide useful information for predicting mortality in sepsis patients – particularly among individuals with NGR and Pre-DM. This integrated approach highlights the potential need for personalized glycemic management strategies, which warrants further investigation in prospective studies.