Triglyceride–glucose index and hyperhomocysteinemia within a cardiovascular–kidney–metabolic biomarker network: insights from interpretable machine learning and epidemiological modeling

Abstract

Background Although the triglyceride–glucose (TyG) index is an established surrogate marker of insulin resistance and metabolic burden, its relationship with hyperhomocysteinemia (HHcy) within the cardiovascular–kidney–metabolic (CKM) framework remains insufficiently understood. This study aimed to investigate the association between TyG and HHcy and to prioritize CKM-related biomarkers using interpretable machine learning and epidemiological modeling. Methods This retrospective cross-sectional study included 31,116 adults undergoing health examinations. HHcy was defined as serum homocysteine >15 μmol/L. The relationships among CKM-related biomarkers were characterized using smooth curve fitting, Spearman correlation heatmaps, and network visualization. An extreme gradient boosting model coupled with SHapley Additive exPlanations was used to evaluate feature importance for HHcy identification. Sequential multivariable logistic regression and restricted cubic spline analyses were performed to assess the linear and nonlinear associations between TyG and HHcy. Subgroup and sensitivity analyses were conducted to assess the robustness of the observed associations, including alternative adiposity adjustment, creatinine-based estimated glomerular filtration rate substitution, and restriction of extreme TyG values. Results Network analysis showed that TyG clustered predominantly with adiposity and metabolic indices, whereas homocysteine showed closer connectivity with renal-function-related markers. Within the prespecified CKM-related candidate feature set, creatinine ranked highest for HHcy identification, whereas TyG provided limited incremental discriminatory information beyond basic clinical characteristics. In sequential logistic regression, the TyG–HHcy association changed substantially across adjustment models: a crude positive association was attenuated and became inverse after adjustment for CKM-related covariates. In Model 4, which additionally adjusted for log(C-reactive protein + 1) and serum creatinine, each one-unit increase in TyG was associated with lower odds of HHcy (OR = 0.819, 95% CI: 0.765–0.877). Restricted cubic spline analysis indicated a significant nonlinear association (P-nonlinear < 0.001). Conclusion TyG appears to be embedded as a metabolic component within a broader CKM biomarker network, whereas renal-function-related markers, particularly creatinine, are more closely linked to the HHcy phenotype. These findings support a CKM network-based interpretation of the TyG–HHcy relationship and highlight the relevance of metabolic–renal interrelationships in cardiometabolic risk assessment.