Prognostic value of the hemoglobin glycation index in critically ill patients with hemorrhagic stroke: A retrospective cohort study using MIMIC-IV

Dysglycemia is common in neurocritical care; however, bedside indices that capture individual glycation biases are limited. This study aimed to test whether the hemoglobin glycation index (HGI), defined as the difference between observed glycated hemoglobin A1c (HbA1c) and HbA1c predicted from admission glucose, independently predicts mortality in intensive care unit (ICU) patients with hemorrhagic stroke. We performed a retrospective cohort study using the Medical Information Mart for Intensive Care IV (MIMIC-IV) v3.1 database, including 1175 ICU admissions for hemorrhagic stroke. The HGI was calculated from glucose and HbA1c levels measured within 24 hours (h) of ICU admission. The primary outcome was 30-day all-cause mortality; 90-day and 180-day mortality rates were secondary outcomes. Multivariate Cox models, restricted cubic splines, predefined subgroup analyses, and sensitivity analyses were performed. Discrimination and clinical utility were evaluated using cross-validated metrics and decision curve analysis, and model interpretability was assessed using SHapley Additive exPlanations (SHAP). The 30-day, 90-day, and 180-day mortality rates were 22.6%, 26.6%, and 29.8%, respectively. As a continuous variable, a higher HGI was associated with lower 30-day mortality (hazard ratio [HR] 0.83, 95% confidence interval [CI]: 0.73–0.95) and lower 90-day mortality (HR 0.86, 95% CI: 0.76–0.98). With the lowest HGI tertile (T1) as the reference, the second (T2) and third (T3) tertiles had lower adjusted risks of 30-, 90-, and 180-day mortality. Restricted cubic spline analysis indicated a nonlinear (U-shaped) relationship, with increased risk at low HGI levels. Early HGI captures glycation bias beyond measured glucose levels and stratifies the short-term mortality risk after hemorrhagic stroke. These findings support HGI as a practical biomarker that merits external and prospective validation.