Per- and polyfluoroalkyl substances and advanced cardiovascular-kidney-metabolic syndrome: mixture and machine learning methods

Cardiovascular-kidney-metabolic (CKM) syndrome denotes coexisting cardiovascular, renal and metabolic dysfunctions with substantial public-health impact. Per- and Polyfluoroalkyl substances ( PFAS) are persistent environmental chemicals implicated in cardiometabolic and renal pathways, but the roles of individual PFAS and PFAS mixtures in CKM remain unclear. We analyzed pooled NHANES 2015–2020 data (n = 4,217) to evaluate associations between serum PFAS and advanced CKM (stages 3–4 vs. 0–2). We used multivariable logistic regression (two adjustment sets), restricted cubic splines, weighted quantile sum (WQS) regression and quantile g-computation (Qgcomp) to assess individual and mixture effects. For prediction, eight supervised classifiers were trained/validated (70%/30% split) with cross-validated hyperparameter tuning; the top performing model (XGBoost) was interpreted using SHAP values. Higher PFHxS, n-PFOA and Sm-PFOS were inversely associated with advanced CKM (PFHxS: OR, 0.75, 95% CI 0.60–0.95, P = 0.017; n-PFOA: OR, 0.73, 95% CI 0.57–0.94, P = 0.014; Sm-PFOS: OR, 0.77, 95% CI 0.61–0.96, P = 0.020). Splines indicated significant overall associations for PFHxS, n-PFOA, n-PFOS and Sm-PFOS with largely linear dose–response shapes. The WQS negative index was inversely associated with advanced CKM (OR, 0.84, 95% CI 0.73–0.98, P = 0.02); WQS and Qgcomp consistently identified n-PFOA, PFHxS and Sm-PFOS as principal contributors. Classifier discrimination ranged from AUC 0.695 (k-NN) to 0.841 (XGBoost); SHAP interpretation of XGBoost showed predominantly negative PFAS contributions to predicted risk, with n-PFOA, Sm-PFOS and PFHxS ranking highest, aligning with mixture analyses. Higher serum PFHxS, n-PFOA, and Sm-PFOS were associated with lower odds of advanced CKM and were consistently identified as important contributors in both mixture and machine-learning analyses. These associations were directionally consistent across methods, but because of the cross-sectional design, they should be interpreted as hypothesis-generating and require prospective validation.