A Component-Isolated Analysis that Confirmed the Absence of An Incremental Signal of the Triglyceride–Glucose Index in CVD Mortality.

Abstract

Background: The triglyceride–glucose (TyG) index is widely proposed as a low-cost insulin-resistance surrogate associated with cardiovascular disease (CVD) outcomes. However, virtually all published TyG studies report univariate or partially adjusted hazard ratios, without isolating TyG's contribution against models that also contain its mathematical components — fasting triglycerides and fasting glucose. Whether TyG adds incremental predictive value when those components are already present remains unresolved.
Methods: We developed logistic regression (LR) and Cox proportional-hazards models for CVD mortality on NHANES 2003–2018 (n = 16,322; 408 CVD deaths; median follow-up 8.2 years). Models used 23 features and were trained on a 70/15/15 stratified split. We performed component-controlled bootstrap ablation (2,000 iterations) comparing 23-feature models against 21-feature models with TyG and TyG-BMI removed but with triglycerides and fasting glucose retained. The frozen LR model was then applied to a temporally distinct NHANES cohort (1999–2002, n = 4,318) for transportability assessment with time-matched (10-year) and unrestricted analyses. Secondary analyses compared TyG with HOMA-IR and performed competing-risk regression.
Results: TyG showed no incremental discriminative value when its components were already in the model (LR ΔAUROC −0.0002, 95% CI −0.004 to +0.003, p = 0.43; Cox TyG HR 1.069, p = 0.27). In the development cohort, the LR model achieved an AUROC of 0.853 (95% CI 0.814–0.890) and a Cox C-index of 0.888, with a clear TyG quartile gradient (Q1 0.91% to Q4 3.61% CVD mortality). The univariate TyG dose–response gradient was preserved across cohorts (7.4% → 13.0% across strata, p < 0.001), but model discrimination dropped substantially under temporal transport (AUROC 0.853 → 0.652 at 10 years; 0.597 unrestricted). TyG and HOMA-IR were equivalent univariately (AUROC 0.590 vs 0.586, r = 0.562). Cause-specific Cox models suggested a possible CVD-specific pattern for TyG (HR 1.203, p = 0.089), with no association for non-CVD mortality (HR 1.004, p = 0.95). Exploratory Platt recalibration improved calibration metrics but was fitted and evaluated on the same temporal cohort and requires independent confirmation.
Conclusions: TyG is a useful univariate cardiometabolic biomarker but provides no incremental value as an additional feature in multivariable models that already contain triglycerides and fasting glucose. Temporal transportability of the prediction model is limited; calibration adjustment requires independent validation. These findings refine the appropriate role of TyG in CVD risk modeling.