Atherogenic index of plasma and cardiovascular high-risk status in the ChinaHEART luohe cohort: multivariable association modeling with nonlinear dose-response and effect heterogeneity

Background Community screening programs increasingly use World Health Organization (WHO) cardiovascular disease (CVD) risk charts to identify individuals at high predicted 10-year risk. The atherogenic index of plasma (AIP), derived from triglycerides (TG) and high-density lipoprotein cholesterol (HDL-C), may capture atherogenic dyslipidemia and support pragmatic risk stratification. Methods We conducted a cross-sectional analysis of baseline data from the China Health Evaluation And risk Reduction through nationwide Teamwork (ChinaHEART) community screening program in Luohe, China. Among 6,860 screened participants, 6,702 with complete data for AIP computation, WHO risk classification, and prespecified covariates were included. The outcome was the WHO CVD risk chart-defined predicted 10-year CVD high-risk category (high risk: ≥20%), rather than adjudicated or incident CVD events. AIP was calculated as log10(TG [mmol/L]/HDL-C [mmol/L]) and modeled as both a continuous and categorical exposure; spline models tested nonlinearity, and ROC analyses evaluated discrimination and derived a Youden-index cutoff. In addition, we performed an explainable machine-learning pipeline for CVD high-risk prediction using LASSO logistic regression for feature selection (AIP forced-in), followed by a random forest classifier and SHAP-based interpretation. Results Of 6,860 screened participants, 6,702 were included in the analytic sample (median age 58 years; 38% men). The WHO CVD risk chart-defined predicted 10-year CVD high-risk category was present in 1,440 (21%) participants and was more frequent in the high-AIP group than in the low-AIP group. Higher AIP was associated with higher odds of CVD high-risk status. Restricted cubic splines supported a non-linear association. Discrimination was modest for AIP alone (AUC 0.557) and improved in adjusted models (AUC 0.650). In the machine-learning pipeline (LASSO + random forest), the random forest model achieved an AUC of 0.792, and SHAP analyses ranked LDL-C and history of hypertension as the strongest contributors, with AIP remaining among the top predictive features. Conclusion In this community-based ChinaHEART population, higher AIP was non-linearly associated with the WHO CVD risk chart-defined predicted 10-year CVD high-risk category. Although AIP alone had limited discrimination, it may serve as a simple adjunct marker to triage individuals for intensified risk assessment in primary-care screening settings.