Machine learning-based prognostic model for trauma and surgical ICU patients: the role of peak glucose and glycemic variability

Stress-induced hyperglycemia and glycemic variability are common in critically ill patients, especially those with trauma or postoperative complications. Using the MIMIC-IV database, we retrospectively analyzed 21,875 patients admitted to the Trauma Surgical Intensive Care Unit (TSICU) or Surgical Intensive Care Unit (SICU). Patients were stratified by peak glucose and glycemic coefficient of variation, with 28-day, 180-day, and 1-year mortality as outcomes. Kaplan–Meier and restricted cubic spline analyses showed that severe hyperglycemia and high variability were consistently associated with higher mortality ( P < 0.001). Subgroup analyses supported these findings. Machine learning models based on LASSO-selected features were further constructed, among which the random forest demonstrated the best predictive performance. These results indicate that elevated glycemic variability and severe hyperglycemia are independent predictors of mortality in TSICU and SICU patients, and that random forest models integrating glucose indices with clinical features may enhance individualized risk stratification and management.