An interpretable machine learning model predicts the interactive and cumulative risks of different environmental chemical exposures on depression.

Humans are exposed to a multitude of environmental chemical mixtures (ECMs) in daily life that may influence depression risk. While prior studies have shown individual ECM exposures to depression, the cumulative and interactive effects of multiple co-exposures remain poorly characterized. This study aimed to develop an interpretable machine learning (ML) model to predict depression risk from ECMs and reveal their interactions mediated through endogenous metabolites and proteins. Using NHANES 2011-2016 data, we analyzed serum and urinary ECMs from 1333 adults, with depression assessed via PHQ-9 scores. Nine ML models were evaluated, with a random forest model showing the best performance (AUC: 0.967, and F1 score: 0.91) in predicting depression risk from ECM exposures. Shapley Additive Explanations (SHAP) identified serum cadmium and cesium, and urinary 2-hydroxyfluorene as the most influential predictors among 52 ECMs. An individualized depression risk assessment model was developed based on SHAP values for key ECMs. Mediation network analysis implicated oxidative stress and inflammation as crucial pathways relating ECMs to depression. This study presents an interpretable ML approach for elucidating cumulative environmental risks for depression, advancing our understanding of complex chemical-health interactions and potentially informing targeted interventions and prevention strategies for depression related to environmental exposures.