Abstract

Human exposure to per- and polyfluoroalkyl substances (PFASs) has received considerable attention, particularly in pregnant women because of their dramatic changes in physiological status and dietary patterns. Predicting internal PFAS exposure in pregnant women, based on external and relevant parameters, has not been investigated. Here, machine learning (ML) models were developed to predict the serum concentrations of PFOA and PFOS in a large population of 588 pregnant participants. Dietary exposure characteristics, demographic parameters, and in particular, serum fatty acid (FA) data were used for the model development. The fitting results showed that the inclusion of FAs as covariates significantly improved the performance of the ML models, with the random forest (RF) model having the best predictive performance for PFOA (R2 = 0.33, MAE = 1.51 ng/mL, and RMSE = 1.89 ng/mL) and PFOS (R2 = 0.12, MAE = 2.65 ng/mL, and RMSE = 3.37 ng/mL). The feature importance analysis revealed that serum FAs greatly affected PFOA concentration in the pregnant women, with saturated FAs being associated with decreased PFOA levels and unsaturated FAs with increased levels. Comparison with one-compartment pharmacokinetic model further demonstrated the advantage of the ML models in predicting PFAS exposure in pregnant women. Our models correlate for the first time blood chemical concentrations with human FA status using ML, introducing a novel perspective on predicting PFAS levels in pregnant women. This study provides valuable insights concerning internal exposure of PFASs generated from external exposure, and contributes to risk assessment and management in pregnant populations.