Behavioral, Life Style and Psychological Predictors of Adolescent Depression: A Machine Learning Approach

Adolescent depression has one of the major public health concerns due to its negative impacts on psychological well-being, academic success, and social functioning. Proactive detection of adolescent at risk is essential for timely support and assistance; however, conventional statistical methods may not adequately capture the complex nonlinear relationships among behavioral, lifestyle and psychological factors. The present study investigated the performance of five supervised machine learning approaches for predicting adolescent depression utilizing a publicly available Teen Mental Health Dataset. To improve model performance, class imbalance was addressed using Adaptive Synthetic Sampling and Bayesian optimization was applied for hyperparameter tuning. Model performance was assessed using different metrices and was further examined through learning curve analysis, repeated stratified cross validation and non-parametric statistical tests. Among evaluated models XGBoost and CatBoost achieved the best predictive performance, outperforming the remaining classifiers across all evaluation metrics. Friedman and Wilcoxon signed rank test confirmed significant differences in model performance. Model interpretation using SHAP and LIME consistently identified sleep duration, stress level, anxiety level, and daily social media use as the most important factors associated with adolescent depression. These results shows that predictive machine learning models can help to the early identification of adolescents at risk of depression while providing valuable insights into the behavioral factors associated with mental health. Such information may assist clinicians, educators and public health practitioners in developing targeted screening and early intervention strategies.