The Screen-Time Trap: An Explainable Machine Learning Framework for Pediatric Screen-Time Overuse Prediction

Abstract

Excessive screen time in children and adolescents is associated with adverse outcomes such as poor sleep, eye strain, anxiety, and obesity risk. Existing studies have often focused on smartphone addiction or descriptive association studies rather than explainable predictive modeling of screentime overuse in pediatric populations. In this study, we develop and evaluate an explainable machine learning framework to predict whether children exceed the recommended daily screen-time limit using a real-world dataset of 9,668 Indian children aged 8–18 years after preprocessing (4,928 male and 4,740 female). Screen-time overuse was observed in 86.7% of male respondents and 84.8% of female respondents. Our pipeline includes preprocessing, duplicate removal, encoding, imputation, feature scaling, SMOTE-based class balancing on the training set, comparative evaluation of multiple classifiers, hyperparameter optimization, ensemble learning, and SHAP-based explanation. Among the evaluated models, XGBoost with hyperparameter tuning achieved the best performance, with 85% accuracy, 94% precision, 89% recall, 91% F1-score, and 91% ROC-AUC. Explainability analysis showed that poor sleep and the educational-to-recreational screen-time ratio were the strongest predictors, while eye strain and anxiety also emerged as important factors. These findings show that explainable machine learning can support early identification of children at risk of screen-time overuse and inform targeted interventions.