Abstract

Depression is one of the most common mental health concerns worldwide and is associated with considerable personal, social and economic consequences. In recent years, the growing availability of population health data has encouraged researchers to apply machine learning to depression prediction. However, there is still limited evidence that interpretable machine learning can provide transparent identification of important predictors while maintaining predictive performance using recent nationally representative data. The present study therefore investigated predictors of depressive symptoms using the NHANES 2021–2023 dataset and evaluated the usefulness of explainable machine learning for depression risk assessment. Using a dataset of 4,227 adults, depressive symptoms were assessed using the Patient Health Questionnaire-9 (PHQ-9). To identify factors associated with depressive symptoms, exploratory statistical analyses, traditional machine learning algorithms, deep learning models and explainability techniques were applied. The results showed that significant differences were observed between depressed and non-depressed participants in age, income-poverty ratio, body mass index, sleep duration, sedentary behaviour, smoking status and diabetes status. Income-poverty ratio consistently emerged as one of the most influential predictors across statistical analyses, machine learning models and explainability analyses. Among the predictive models, Random Forest achieved the highest discrimination performance (ROC-AUC= 0.692), whereas Logistic Regression achieved the highest recall (60.2%). Deep learning models produced performance comparable to traditional machine learning approaches and offered no meaningful predictive advantage. The findings demonstrate that interpretable machine learning can identify meaningful socioeconomic, demographic, lifestyle and health-related predictors of depressive symptoms while maintaining transparent model interpretation. The study further suggests that, for structured population health data, traditional machine learning models may provide practical and interpretable alternatives to more complex deep learning approaches, thereby contributing to future mental health research and evidence-informed clinical practice.

Keywords:Depression; Depressive symptoms; Mental health; Interpretable machine learning; Explainable artificial intelligence; Risk assessment; Public health; NHANES

Abbreviations: AI: Artificial Intelligence; ML: Machine Learning; DL: Deep Learning; XAI: Explainable Artificial Intelligence; NHANES: National Health and Nutrition Examination Survey; PHQ-9: Patient Health Questionnaire-9; BMI: Body Mass Index; SVM: Support Vector Machine; MLP : Multilayer Perception; SMOTE: Synthetic Minority Oversampling Technique; ROC-AUC: Area Under the Receiver Operating Characteristic Curve; SHAP: Shapley Additive Explanations; LIME: Local Interpretable Model- Agnostic Explanations