Machine learning prediction of exposure to acrylamide based on modelling of association between dietary exposure and internal biomarkers.

Abstract

The ubiquitous occurrence of acrylamide in various thermal processing food products poses a potential health risk for the public. An accurate exposure assessment is crucial to the risk evaluation of acrylamide. Machine learning emerging as a powerful computational tool for prediction was employed to establish the association between internal exposure and dietary exposure to acrylamide among a Chinese cohort of middle-aged and elderly population (n=1,272). Five machine learning regression models were constructed and compared to predict the daily dietary acrylamide exposure based on urinary biomarkers including N-acetyl-S-(2-carbamoylethyl)-L-cysteine (AAMA), N-acetyl-S-(2-carbamoylethyl)-L-cysteine-sulfoxide (AAMA-sul), N-acetyl-S-(2-carbamoyl-2-hydroxyethyl)-L-cysteine (GAMA), and N-acetyl-S-(1-carbamoyl-2-hydroxyethyl)-L-cysteine (iso-GAMA). Other important covariates such as age, gender, physical activities, and total energy intake were also considered as predictors in the models. Average dietary intake of acrylamide among Chinese elderly participants was 8.9μg/day, while average urinary contents of AAMA, AAMA-sul, GAMA, and iso-GAMA were 52.2, 19.1, 4.4, and 1.7nmol/g Ucr (urine creatinine), respectively. Support vector regression (SVR) model showed the best prediction performance with a R of 0.415, followed by light gradient boosting machine (LightGBM) model (R=0.396), adjusted multiple linear regression (MLR) model (R=0.378), neural networks (NN) model (R=0.365), MLR model (R=0.363), and extreme gradient boosting (XGBoost) model (R=0.337). The present study firstly correlated dietary exposure with internal exposure to acrylamide among Chinese elderly population, providing an innovative perspective for the exposure assessment of acrylamide.