Predicting depression risk with machine learning models: identifying familial, personal, and dietary determinants

The pathogenesis of depression is highly complex, therefore, the development of predictive models using readily available clinical parameters to identify individuals at risk of adverse depressive outcomes holds significant clinical value. 7108 participants from the United States National Health and Nutrition Examination Survey were collected. A total of 11 machine learning models were employed, including CatBoost, Decision Tree, Gradient Boosting Tree, LightGBM (LGB), Logistic Regression (LR), Lasso, Naive Bayes, Neural Network, Random Forest (RF), Support Vector Machine, and XGBoost, with comparisons made against the generalized linear regression model. Model performance was rigorously assessed using receiver operating characteristic (ROCs), calibration curves, and decision curves analysis. Feature importance was interpreted through Shapley Additive exPlanations to identify key influencing factors at the whole level and interpret individual heterogeneity through instance-level analysis. Significant differences in overall characteristics were observed between depressed patients and healthy controls. The RF model demonstrated superior performance, followed by Lasso, XGBoost, and LGB models, which also showed relatively high predictive accuracy. The training set AUC values for the RF, Lasso, XGBoost, and LGB models were 0.998, 0.713, 0.723, and 0.804, respectively, while their corresponding test set AUC values were 0.705, 0.719, 0.714, and 0.687. Based on variable importance ranking from RF, Lasso, XGBoost, and LGB models, we identified eight key predictors: body mass index, education level, marital status, annual family income, family income to poverty ratio, trouble sleeping, composite dietary antioxidant index, and dietary inflammatory index. These variables were integrated to develop a comprehensive statistical model for predicting depression risk. We developed a robust predictive model for assessing depression risk, incorporating eight clinically accessible predictors. This model demonstrates reliable predictive performance for depression onset and provides valuable reference for clinical decision-making. Clinical trial number is not applicable.