Glycemic variability and postoperative mortality following cardiac surgery: evidence from a real-world ICU cohort

Glycemic variability (GV) may capture peri-operative metabolic instability more sensitively than absolute glucose levels, but its prognostic value after major cardiac surgery remains uncertain.

We retrospectively analyzed 2,943 ICU patients from the MIMIC-IV database who underwent coronary artery bypass grafting or valve surgery. GV was defined as the coefficient of variation of glucose during ICU stay. Primary outcome was 30-day all-cause mortality, with 90- and 360-day mortality as secondary endpoints. Missing covariates (< 20%) were multiply imputed with CART. Multivariable Cox regression, restricted cubic splines, and inverse probability of treatment weighting (IPTW) were applied. Subgroup analyses with interaction testing were conducted. Discriminatory performance was assessed by ROC curves, and nine machine learning models were benchmarked with LASSO-based feature selection and cross-validation.

Higher GV was consistently associated with elevated mortality risk. In fully adjusted analyses, patients in the highest vs. lowest quartile had hazard ratios of 6.72 (95%CI 1.89-23.93) for 30-day, 2.46 (1.31-4.59) for 90-day, and 2.19 (1.44-3.32) for 360-day mortality. Associations appeared linear, remained robust after IPTW, and were confirmed in sensitivity analyses. No significant interactions were detected across clinical subgroups. GV alone showed fair discrimination for 30-day mortality (AUC 0.686), whereas the XGBoost model incorporating perioperative variables achieved the highest discrimination (AUC 0.845). SHAP analysis indicated GV contributed incremental prognostic information but was not the leading predictor.

Among ICU patients undergoing cardiac surgery, greater GV was independently associated with higher short- and long-term mortality. GV may serve as a complementary marker to established clinical predictors for postoperative risk stratification.