Multimodal Temporal Data Fusion and Dynamic Attention Based Psychological Crisis Prediction for University Students

Nowadays, the psychological problems faced by college students are becoming more and more complicated. It is very important to identify psychological risks as early as possible and intervene well. Our study proposed a psychological crisis prediction model. The core idea is to integrate multiple time series data and add a dynamic attention mechanism to more comprehensively judge the psychological state of students. This model integrates students' behavioral data, social data, and emotional data, and can more comprehensively and meticulously portray the process of their psychological changes. We used a layer of Transformer-based temporal feature fusion module in it, which is specifically used to capture those long-term correlations. At the same time, the dynamic attention module can automatically focus on more critical features, especially patterns such as negative emotion accumulation and social isolation, which the model will pay special attention to. Our study used data from 530 students from three universities to conduct experiments, and the results were quite good. The F1 value of the model reached 0.88 and the AUC reached 0.93, which is significantly better than the traditional method. In addition, we also conducted feature contribution analysis and found that the model is very interpretable, especially sensitive to those core psychological risk factors. Overall, this model has verified its potential in predicting psychological crises. In the future, it can be used as an intelligent and scalable tool to help schools manage mental health more accurately and provide more targeted support.