Triangulating evidence for cardiometabolic Index: ROC cutoff, spline nonlinearity, and explainable machine learning for CVD high risk

Figures

• 

  Figure 1. ROC curves and Youden-index cutoff for CMI. Receiver operating characteristic (ROC) curves for CMI alone in the crude model and for multivariable models incorporating CMI (Models 2 and 3). AUC (95% CI) for each model is reported. The Youden-index optimal cutoff for CMI derived from the crude model is indicated.

• 

  Figure 2. Raincloud plot of CMI by CVD high-risk status. Raincloud plot comparing CMI distributions between non-high-risk and high-risk groups. Points represent individual participants; boxplots summarize median and interquartile range; half-violin density displays distribution shape. P-value for between-group difference is shown.

• 

  Figure 3. Restricted cubic spline analysis of CMI and CVD high-risk status. Restricted cubic spline showing the association between cardiometabolic index (CMI) and odds of WHO-defined CVD high-risk status. Solid line indicates adjusted odds ratio and shaded area indicates 95% confidence interval; P for overall and P for nonlinearity are shown.

• 

  Figure 4. Subgroup analyses of CMI and CVD high-risk status. Exploratory unadjusted forest plots of the associations between CMI and CVD high-risk status across prespecified subgroups. Panel A: per 1-unit increase in CMI; Panel B: high vs. low CMI using the rounded analytic threshold of 0.7. Odds ratios (ORs) with 95% CIs are shown, along with P for interaction. Analyses were performed using the available complete cases for the exposure, outcome, and stratification variable within each subgroup; therefore, counts may differ from those reported in Table 1.

• 

  Figure 5. Interaction plots of CMI with subgroup factors. Predicted probability of CVD high-risk status across CMI levels stratified by (A) age quartiles, (B) alcohol use, (C) heart rate quartiles, (D) marital status, (E) antihypertensive drugs, (F) antidiabetic drugs, and (G) statin use. Shaded bands indicate 95% confidence intervals; P for interaction is from likelihood ratio tests.

• 

  Figure 6. Explainable machine-learning validation (LASSO + RF + SHAP). Machine-learning pipeline for predicting CVD high-risk status. Left: LASSO-selected predictors (λ.1se) and coefficient direction. Middle: ROC curve of the random forest model (AUC shown). Bottom: SHAP beeswarm plot showing feature importance and direction of impact on predicted probability; color represents within-feature value (low to high).