Multimodal Deep Learning for Early Mental Health Risk Screening in Adolescents Using Text and Structured Data

This work presents an extended overview of a multimodal deep learning framework for the early screening of mental health risk among adolescents. Text-derived emotional cues are combined with structured demographic and clinical indicators, as well as metrics on academic performance. A dual-stream neural network architecture is used to grasp both sequential patterns in text and interactions among structured variables. This extended version provides deeper motivation, enhanced feature engineering, extended model evaluation, and a more comprehensive discussion of limitations and future directions. The results show improved stability and accuracy over their single-modality counterpart models.