Association of the TyG index and its body fat distribution composites with CHD and MACE risk in adults with OSAHS: incremental discrimination and exploratory mediation analysis

Figures

• 

  Figure 1. Variance inflation factors (VIFs) for candidate continuous covariates included in the multivariable models. All covariates retained in the primary models had VIF values<5, indicating an acceptable risk of multicollinearity.

• 

  Figure 2. Spearman correlation matrix of candidate covariates. This heatmap illustrates the correlation coefficients between candidate variables. The intensity of the color corresponds to the magnitude of the correlation, while the color scale denotes the direction of the relationship (positive or negative); numerical values within the cells represent the respective correlation coefficients. This figure facilitates the assessment of the correlation structure and supplements the VIF results in identifying potential multicollinearity. *P < 0.05, **P < 0.01, and ***P < 0.001.

• 

  Figure 3. Dose-response relationships were evaluated in the fully adjusted logistic regression model (Model 3) for (A) ZTyG, (B) ZTyG-WWI, (C) ZTyG-CVAI, and (D) ZTyG-BRI. Splines were fitted with three knots at the 10th, 50th, and 90th percentiles, using Z = 0 as the reference (OR = 1). Solid lines represent the adjusted odds ratios (ORs), and shaded areas denote the 95% confidence intervals (CIs). Histograms illustrate the exposure distribution. To minimize instability from sparse extreme values, curves are displayed within the 5th–95th percentile range, with tail sample sizes (x<P5 and x>P95) reported. Poverall represents the joint test for the overall association, and Pnon−linear tests for deviations from linearity.

• 

  Figure 4. Receiver operating characteristic (ROC) curves compare the discriminative capacity of nested models for CHD. Model 1:Traditional risk factors (age, sex, hypertension, diabetes, 24hSBP, smoking, LDL-C, and WBC count); Model 2: Model 1 + raw-scale AHI; Models 3a–3d: Model 2 + Z-standardized TyG, TyG-WWI, TyG-CVAI, and TyG-BRI, respectively. The Area Under the Curve (AUC) was used to quantify discriminative performance.

• 

  Figure 5. Supportive validation results for the predictive models. (A) Calibration Curves: Evaluates model calibration by comparing predicted probabilities with actual observed risks; the 45° dashed line represents ideal calibration. (B) Decision Curve Analysis (DCA) Evaluates the clinical net benefit of different models across various threshold probabilities. “Treat all” and “Treat none” serve as reference strategies. A higher curve indicates greater potential clinical utility within the corresponding threshold range.

• 

  Figure 6. Kaplan–Meier curves illustrate the cumulative incidence of MACE across quartiles (Q1–Q4) of (A) TyG, (B) TyG-WWI, (C) TyG-CVAI, and (D) TyG-BRI. Differences between survival curves were evaluated using the log-rank test. A greater degree of curve separation indicates a more robust capacity for MACE risk stratification by the respective index.

• 

  Figure 7. Time-dependent ROC curves were constructed based on the linear predictors of the Cox models using the inverse probability of censoring weighting (IPCW) method to evaluate the dynamic discriminative capacity for MACE at predefined follow-up intervals (12, 18, and 24 months). The AUCs of each candidate model at these time points are presented to illustrate the temporal stability of their predictive performance.

• 

  Figure 8. Exploratory statistical decomposition of the association between Z-standardized AHI (ZAHI) and incident MACE. (A) ZTyG-CVAI, (B) ZTyG-WWI, (C) ZTyG-BRI, and (D) ZTyG were evaluated separately as candidate pathway variables to estimate the total effect (TE), natural direct effect (NDE), and natural indirect effect (NIE). The figure presents effect estimates for each candidate pathway variable to describe potential indirect pathway signals. Because ZAHI and TyG-related candidate pathway variables were measured concurrently at baseline, this analysis should be interpreted as exploratory statistical decomposition rather than definitive causal mediation evidence.