Triangulating evidence for cardiometabolic Index: ROC cutoff, spline nonlinearity, and explainable machine learning...

The cardiometabolic index (CMI), a composite marker reflecting central adiposity and dyslipidemia, may offer a pragmatic tool for community cardiovascular risk screening.

We analyzed cross-sectional screening data from the ChinaHEART Luohe cohort. WHO chart-defined CVD high risk was a 10-year predicted risk ≥20%. CMI was evaluated per 1-unit increase and, for descriptive and stratified analyses, dichotomized using a rounded pragmatic analytic threshold of 0.7 derived from a crude ROC/Youden optimal threshold of 0.723. Discrimination (AUC with bootstrap 95% CIs), raincloud plots, restricted cubic splines (RCS), stepwise logistic regression, subgroup/interaction analyses, and an explainable ML pipeline (LASSO-random forest-SHAP) were applied.

Among 6,701 participants, 1,439 (21%) were classified as CVD high risk; prevalence was higher in the high-CMI group (≥0.7) than the low-CMI group (29% vs. 18%). CMI alone showed modest discrimination for WHO-defined CVD high-risk status (AUC: 0.571, 95% CI: 0.555-0.586), whereas multivariable models incorporating CMI showed higher discrimination (Model 3 AUC: 0.642, 95% CI: 0.626-0.659). The raincloud plot showed higher CMI in the high-risk group (P < 0.001), and RCS suggested nonlinearity (P for overall < 0.001; P for nonlinearity = 0.020). In Model 3, CMI was associated with higher odds of CVD high risk (OR: 1.31, 95% CI: 1.16-1.48 per 1-unit; OR: 1.50, 95% CI: 1.32-1.71 for ≥0.7 vs. <0.7). In ML, random forest achieved AUC 0.814; SHAP ranked CMI 3rd of 7 LASSO-selected features.

Higher CMI was associated with WHO-defined CVD high-risk status with a nonlinear pattern and consistent importance across conventional and explainable ML analyses, supporting its potential utility as an adjunct screening marker.