Human biomonitoring of PFAS and metals in relation to diabetic kidney disease: Evidence from NHANES and machine learning interpretation

Abstract

Diabetic kidney disease (DKD) represents a growing health burden as the prevalence of diabetes continues to rise. Beyond traditional metabolic risk factors, environmental exposures may provide additional information for DKD characterization and prediction. We used data from NHANES 2005–2016 to examine the associations of heavy metals and per- and polyfluoroalkyl substances (PFAS) with DKD among adults with diabetes. Survey-weighted logistic regression was combined with Boruta feature selection, machine learning models, and Shapley Additive Explanations (SHAP) to evaluate both adjusted associations and predictive patterns. In unadjusted models, Pb and Cd were positively associated with DKD, but these associations did not persist after adjustment. In the fully adjusted model, PFOA and PFHxS remained significant and both showed inverse associations. These inverse associations require cautious interpretation and should not be read as evidence of protective effects because NHANES cannot establish temporal order. SHAP analysis showed that clinical and metabolic factors were the main source of predictive information. Environmental variables accounted for 24.18% of the total mean absolute SHAP importance in the RF model, with PFOA ranking highest among the exposure variables. This percentage represents the relative distribution of model-based importance and should not be interpreted as a causal contribution or attributable fraction. Other pollutants also retained model importance, reflecting a broader classification structure involving non-linear and correlated information across predictors. Environmental biomonitoring may provide complementary information for DKD risk characterization, but it should not replace established clinical and metabolic indicators.