Evaluating the relationship between environmental chemicals and obesity: Evidence from a machine learning perspective

Environmental chemicals are increasingly recognized as important contributors to obesity, yet the number of studies evaluating this relationship remains insufficient. This study aimed to investigate these associations using interpretable machine learning techniques. Data from 1183 participants in the 2011-2012 National Health and Nutrition Examination Survey were analyzed. Several machine learning models, including Support Vector Machines, Random Forest, k-Nearest Neighbors, Naive Bayes, AdaBoost, and XGBoost, were employed to predict generalized and abdominal obesity using environmental chemical exposures and demographic information. The XGBoost model was further explored for its ability to interpret variable contributions, utilizing SHapley Additive exPlanations (SHAP) to identify key predictors. Logistic regression models revealed that 4-OH-PHEN, 2-OH-NAP, and 2-OH-PHEN were positively associated with generalized obesity, whereas UMo and 3-OH-FLUO exhibited negative associations. Similarly, 4-OH-PHEN demonstrated a positive association with abdominal obesity, whereas 3-OH-FLUO, USr, and BPb were negatively associated. To further examine these relationships, dose-response associations between environmental chemicals and obesity were analyzed using restricted cubic spline plots. A nonlinear relationship was identified between UMo and obesity (P-nonlinear=0.016). Mediation analysis revealed that blood lipids partially mediated the relationship between certain environmental chemicals and obesity. This study underscores the importance of interpretable machine learning in understanding the complex associations between environmental chemicals and obesity. It identified specific chemicals associated with generalized and abdominal obesity and shed light on the mediating role of blood lipids. These findings contribute to the growing body of evidence on the role of environmental exposures in obesity and provide potential pathways for future research and interventions.