P075 - PROGNOSTIC VALUE OF THE TYG INDEX AND ITS DERIVATIVES FOR ALL-CAUSE MORTALITY IN T2DM-CHD PATIENTS: A SHAP-INTERPRETABLE MACHINE LEARNING MODEL USING NHANES DATA

Background: Patients with type 2 diabetes mellitus and coronary heart disease (T2DM-CHD) face elevated mortality risks, yet effective risk stratification remains limited. The triglyceride-glucose (TyG) index and its derivatives (TyG-BMI, TyG-WC, TyG-WHtR) are promising insulin resistance surrogates. This study aimed to assess the prognostic value of TyG-related indices and to develop an interpretable machine learning model for individualized mortality prediction.

METHODS AND RESULTS: We analyzed 1,253 adults with T2DM-CHD from NHANES (1999-2018), with 625 deaths recorded. TyG indices and 35 clinical variables were examined. Key features were selected using Boruta and SHAP-based recursive feature elimination. Six machine learning models were trained and evaluated. A random forest model incorporating age, CRP, eGFR, TyG-BMI, and five additional features yielded the best performance (AUC: 0.822 external validation). SHAP analysis highlighted age, inflammation, renal function, and TyG-BMI as dominant predictors. TyG-BMI demonstrated a robust, nonlinear inverse association with all-cause mortality in multivariable Cox and restricted cubic spline analyses (HR for Q4 vs Q1: 0.75, 95% CI: 0.58-0.97, P = 0.027). The TyG index and other derivatives were not independently predictive after adjustment. A web-based calculator was developed to provide real-time, interpretable risk estimates.

Conclusion: TyG-BMI is a stable and independent protective predictor of all-cause mortality among T2DM-CHD patients, outperforming the original TyG index. The SHAP-interpretable random forest model offers strong generalizability and supports personalized decision-making through an accessible online tool. These findings support the integration of TyG-BMI in future cardiometabolic risk assessment strategies.