Machine Learning-Based Ternary Risk Stratification of Adolescent Social Media Addiction Using an Adapted Bergen Framework, PHQ-A Indicators, and Objective Screen-Time Data

Abstract

Social media addiction among adolescents is a complex psychological and behavioral construct, which cannot be estimated simply by the screen-time duration. In this study, the development and validation of a machine learning model to quantify and categorize the level of social media addiction among adolescents by combining various factors, including psychometric addiction indicators, depressive symptoms, demographic and behavioral measures, and empirical smartphone screen-time data. A local cross-sectional data set was obtained from 808 adolescents aged 13-18 years attending intermediate and preparatory schools in Baghdad, Iraq. The severity of social media addiction was quantified using an extended Bergen-based scale in Arabic language with 18 items grouped into six addiction dimensions: salience, mood modification, tolerance, withdrawal, conflict, and relapse. Following the Bergen Social Media Addiction Scale scoring protocol, the three items per addiction dimension were averaged, and the averages were summed up to obtain the overall score between 6 and 30. The classification criteria were established based on two different data sources: a local Iraqi study utilizing the same Bergen score threshold value of 24 to diagnose social media addiction and an international item response theory study endorsing the use of Bergen scores as severity scales especially values around 14 and 19. The outcome variable was defined as a Bergen-anchored composite risk score, not as the Bergen score alone or the PHQ-A score alone. The adapted 18-item Bergen-based score was used as the primary addiction component, while the PHQ-A score was aligned to the same 6–30 range and incorporated as a supporting psychological severity component, not as an independent depression diagnosis. Participants were then classified into three risk categories: low, moderate, and high. Objective smartphone screen-time indicators collected over ten consecutive days were used only as independent predictive features and were not included in target construction.  Six machine-learning models were evaluated in their baseline configurations—Random Forest, XGBoost, LightGBM, CatBoost, SVM with RBF kernel, and Logistic Regression (L2)—along with optimized, hybrid, and soft-voting variants. The modeling pipeline used a stratified 80/20 train/test split, five-fold stratified cross-validation within the training data, and SMOTE applied only to the training set. The original ternary target distribution was imbalanced, with 313 low-risk, 474 moderate-risk, and 21 high-risk cases; after the train/test split, SMOTE balanced the training classes from 250/379/17 to 379/379/379 cases. In ternary classification, the best performance was achieved by the PSO-optimized CatBoost+XGBoost hybrid model, with test accuracy of 91.77%, weighted F1-score of 0.9173, and MCC of 0.8452. This model slightly outperformed the best simple baseline model, Random Forest, which achieved test accuracy of 91.14%, weighted F1-score of 0.9131, and MCC of 0.8347. These findings suggest that combining adapted psychometric responses, PHQ-A-supported psychological severity information, self-reported behavioral variables, and objective screen-time indicators may support early risk stratification of adolescent social media addiction. However, the findings should be interpreted as internal, cross-sectional decision-support evidence, because objective screen-time data were available only for a subsample and the model was not designed to replace clinical assessment