Impact of the Interaction Between Screen Time and Activity Interests on Adolescent Depression Risk: Construction of a Predictive Model Based on Machine Learning

Abstract

Background: Adolescent depression is an increasing public health concern, with excessive screen time elevating depression risk and activity interests providing protection. However, most studies examine these behaviors separately and rely on limited analytical methods. This study used machine learning (ML) to develop a predictive model and evaluate the combined influence of screen time and activity interests on adolescent mental health.

Methods: A multi-center survey was conducted among adolescents aged 10–14 years in Chongchuan District, Nantong. Depression-related domains were assessed using the Child and Adolescent Mental Health Screening Questionnaire, integrating seven validated scales. A twostage feature-selection strategy identified 11 key predictors. Three ML models (logistic regression [LR], extreme gradient boosting [XGBoost], and categorical boosting [CatBoost]) were trained with an 80:20 stratified split. Class imbalance was addressed using synthetic minority oversampling technique and class-weighting. Model performance and interpretability were evaluated using receiver operating characteristic (ROC) and calibration curves, partial dependence plots, and shapley additive explanations (SHAP) analyses.

Results: A total of 2202 valid questionnaires were analyzed. The distribution of depression severity was as follows: safe 59%, mild 17%, moderate 11%, and severe 13%. The integrated questionnaire demonstrated strong reliability (Cronbach’s α = 0.910) and good construct validity (Kaiser–Meyer–Olkin [KMO] = 0.91; root mean square error of approximation [RMSEA] = 0.049; and comparative fit index [CFI] = 0.859). ROC-Youden analysis confirmed expert-defined cutoffs (29, 32, and 35). Feature selection identified 11 key predictors, with activity interest and psychological functioning consistently ranking highest in importance. Across the three ML models, LR exhibited the best generalizability, XGBoost showed overfitting, and CatBoost achieved balanced performance. SHAP and partial dependence analyses revealed nonlinear screen-time effects and dose-dependent protective effects of activity interest, including the moderation of high screen exposure in severe-risk groups.

Conclusions: This study suggests that ML models can be used to screen adolescents at risk of depression by capturing the combined influence of screen time and activity interests. The model is intended for screening rather than diagnosis and may support school-based early identification, and further validation in clinical contexts is needed.