Personalized Prediction and Intervention for Adolescent Mental Health: Multimodal Temporal Modeling Using Transformer

Guiyuan Zhang^1*Shuang Li²

• ¹Guangxi Vocational College of Water Resources and Electric Power, Nanning, China
• ²Institute of Semiconductors, Chinese Academy of Sciences (CAS), Beijing, Beijing Municipality, China

The mental health problems of adolescents are becoming increasingly serious, and early prediction and personalized intervention have become important research topics. The existing methods have certain limitations in dealing with complex emotional fluctuations and multimodal data fusion. To address this issue, this paper proposes a novel model, MPHI Trans, which combines multimodal data and temporal modeling techniques to accurately capture the dynamic changes in adolescent mental health status. MPH Trans provides tailored intervention recommendations by predicting individuals' emotions and psychological states, promoting personalized mental health management. In the experiment, MPH Trans performed significantly better than various advanced models such as BERT, T5, and XLNet on the DAIC-WOZ and WESAD datasets. Specifically, the accuracy of MPHI Trans on the DAIC-WOZ dataset reached 89%, recall rate was 84%, precision rate was 85%, F1 score was 84%, and AUC-ROC was 92%; The accuracy on the WESAD dataset is 88%, recall rate is 81%, precision rate is 82%, F1 score is 81%, and AUC-ROC is 91%. In addition, the critical role of temporal modeling and multimodal fusion modules in model performance was verified through ablation experiments. After removing these modules, the performance of the model significantly decreased, demonstrating their indispensable role in capturing emotional fluctuations and information fusion