The interpretable machine learning model for depression associated with heavy metals via EMR mining method.

Area of Science:

• Environmental Health
• Computational Psychiatry
• Toxicology

Background:

• Research on the link between heavy metal exposure and depression is limited.
• Machine learning (ML) offers potential for identifying complex environmental health associations.
• Understanding these links is crucial for public health interventions.

Purpose of the Study:

• To develop an interpretable and efficient ML model for detecting depression associated with heavy metal exposure.
• To identify specific heavy metals and their exposure routes (blood, urine) linked to depression.
• To leverage advanced ML techniques for robust prediction and explanation.

Main Methods:

• Utilized data from the US National Health and Nutrition Examination Survey (NHANES) (2013-2020) with 19,368 participants.
• Developed and compared five ML models, optimizing the best model using a Genetic Algorithm (GA).
• Employed SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) for model interpretability.

Main Results:

• An eXtreme Gradient Boosting (XGB) model, optimized by GA, achieved high performance (AUC: 0.686, accuracy: 97.1%) in identifying depression using 16 heavy metal indicators.
• SHAP analysis indicated elevated blood cadmium positively influenced depression prediction.
• Negative influences on depression prediction were observed for urine concentrations of barium, thallium, tin, manganese, antimony, lead, and tungsten, and blood levels of lead, cadmium, mercury, selenium, and manganese.

Conclusions:

• An efficient and robust GA-XGB model successfully identified depression linked to heavy metal exposure.
• Blood cadmium showed a positive correlation with depression.
• Specific heavy metals in urine and blood exhibited negative correlations with depression, highlighting complex exposure-response relationships.