Exploring early life exposome related to low birth weight through machine learning methods

Abstract

BACKGROUND AND AIM: Regarding the multifactorial etiology of low birth weight (LBW), the contribution of environmental hazards is increasing. Exposome refer to totality of a life’s exposure including multiple pollutants, lifestyle and sociodemographic factors should provide a holistic view on LBW. This study aims to explore early life exposome related to LBW and identify important factors through machine learning methods. METHODS We recruited 486 mother-infant pairs from Taiwan Birth Panel Study, with LBW defined as birth body weight below 2500 gm. Early life factors included 27 environmental pollutants measured in cord blood or maternal urine, estimated outdoor air pollutants (particulate matter 2.5 (PM2.5), nitrogen dioxide (NO2) and greenness (Normalized Difference Vegetation Index (NDVI)) and questionnaires on socioeconomic levels, health behaviors and diet during pregnancy. The association between exposome and LBW were examined by five machine learning methods including Naïve bayes, Support vector machine (SVM), Random forest, XGBoost (eXtreme gradient boosting) and Ensemble learning. Model performance was evaluated by area under Receiver operator characteristic (ROC) curve. And, features selection was used to identify important features. RESULTS SVM method and Random forest provided the best prediction based on AUC= 0.73 and 0.74, respectively. The top five important features selected from best model prediction of Random forest include: perfluorooctyl sulfonate (PFOS) in cord blood, uranium and zinc in maternal urine, maternal pre-pregnancy body mass index, maternal gestational weight gain. CONCLUSIONS Machine learning approaches could be useful tools for modeling early life exposome associated with LBW. Environmental pollutants and maternal weight status are critical for LBW risk. KEYWORDS exposome, machine learning, environmental pollutants, low birth weight