The mental health risks of rural adolescents have the characteristics of strong concealment, fast evolution speed and obvious scene differences. Traditional questionnaire screening and manual interview are difficult to meet the needs of early identification. This paper proposes an advanced machine learning detection model for rural adolescents in China, which integrates learning behavior, sleep rhythm, social support, psychological assessment and weekly diary text into a unified computing framework, and realizes low risk, attention risk and high risk identification through multi-source feature construction, emotional risk embedding and two-stage hierarchical classification. Experimental results show that the model achieves 91.40% Accuracy, 0.891 Macro-F1 and 0.931 AUROC under 5-fold cross validation, and the high risk recall reaches 0.860. The research shows that the combination of multimodal fusion and interpretable analysis can more effectively capture the early psychological risk signals of rural adolescents, and provide a computable basis for school grading intervention.
