Comprehensive Summary
This study examined whether machine learning models can accurately predict the risk of depression in U.S. adults using demographic, clinical, sleep, and dietary variables. To test this, data were drawn from 7,108 participants in the NHANES 2011–2016 cycles, and 11 machine learning models were trained and evaluated using ROC curves, calibration plots, and decision curve analysis. Random Forest, Lasso, XGBoost, and LightGBM demonstrated the strongest overall performance, with Random Forest achieving the highest AUC but showing signs of overfitting. Feature importance analyses and SHAP interpretation consistently identified eight key predictors of depression risk: body mass index, education level, marital status, annual family income, family income–to–poverty ratio, trouble sleeping, dietary inflammatory index, and composite dietary antioxidant index. Increased depression risk was associated with higher BMI, sleep disturbances, pro-inflammatory diets, and lower socioeconomic indicators. Higher education, income, and antioxidant intake appeared protective. The researchers concluded that integrating these readily available factors into predictive models improves interpretability and offers insight into the multifactorial nature of depression.
Outcomes and Implications
This study demonstrates that data already gathered in standard clinical and public health settings can be used to estimate depression risk. The strong influence of sleep disturbances and diet-related indices highlights modifiable risk factors. Through screening, lifestyle counseling, and preventive measures, these risk factors can be addressed at an early stage. Clinically, these results imply that machine learning-based risk models could help primary care physicians identify people who are more likely to experience depression and prioritize mental health assessment before symptom severity worsens. This work supports earlier, prevention-focused approaches to depression rather than relying solely on symptom-based diagnosis.