Synergistic effect of insulin resistance and glycemic variability on mortality in ICU patients with gastrointestinal bleeding: impact of the TyG-GVI index and development of an explainable web-based clinical calculator.

OBJECTIVE

To investigate the association between the composite Triglyceride-Glucose Index combined with Glycemic Variability (TyG-GVI) and early-death risk in ICU patients with gastrointestinal bleeding, and evaluate whether integrating this biomarker into ensemble-learning models improves prognostic performance.

METHODS

This retrospective study analyzed adult ICU patients (≥18 years) with gastrointestinal hemorrhage as the principal diagnosis from the MIMIC-IV v3.1 database. TyG-GVI was calculated as the product of the triglyceride-glucose index and glucose coefficient of variation using first-24-hour laboratory data. Patients were stratified by the median TyG-GVI value. Primary outcomes were ICU and in-hospital mortality. We applied Kaplan-Meier survival analysis, multivariable Cox regression, restricted cubic spline (RCS), and ensemble-learning models (Random Forest, XGBoost, LightGBM) with SHAP interpretation.

RESULTS

In total, 998 eligible patients were split into low- (n=499) and high-TyG-GVI (n=499) groups. The high-TyG-GVI group showed more severe physiological disturbance. After adjustment, high TyG-GVI independently predicted ICU mortality (adjusted HR 1.87, 95%CI 1.36-2.55, P < 0.001). RCS revealed a non-linear relationship with an inflection threshold at TyG-GVI = 1.535. Subgroup analyses supported result robustness. The XGBoost model incorporating TyG-GVI yielded better discrimination (AUC = 0.827) than the conventional OASIS score (AUC = 0.759).

CONCLUSION

TyG-GVI is an independent non-linear prognostic marker for early mortality in ICU patients with gastrointestinal bleeding. Combining TyG-GVI with explainable artificial intelligence enhances risk-stratification for acute-care bedside prognostication.