Mapping heterogeneous predictors of problematic smartphone use in Chinese adolescents: A machine learning approach...

Problematic smartphone use (PSU) has become a significant global public health issue among children and adolescents. While previous research has identified various risk factors, studies utilizing nationally representative samples and advanced predictive modeling to decipher the complex, multifactorial nature of PSU remain limited. We aim to develop and validate a high-performance model for predicting PSU severity by analyzing a large-scale national sample of Chinese students (grades 4-12) and systematically evaluating six machine learning algorithms. The K-nearest neighbors (KNN) algorithm demonstrated superior performance without hyperparameter tuning, outperforming other models. Self-inconsistency emerged as the most influential predictor, substantially exceeding other factors in importance. The analysis revealed a clear hierarchy of predictors, with psychological maladjustment, gender, and school integration forming a secondary tier. Furthermore, through clustering analysis, we identified three distinct high-risk PSU subtypes (the contextualized compensatory use subgroup, the identity rigidity and maladaptive escape subgroup, and the extreme emotional dysregulation subgroup), each characterized by unique configurations of risk factors. This study provides robust evidence that machine learning models, particularly KNN, can achieve high predictive accuracy for PSU severity. The identification of a core predictor and distinct risk subtypes offer valuable insights for developing targeted, personalized intervention strategies, moving beyond one-size-fits-all approaches to effectively address adolescent PSU.