College students’ rising mental health difficulties require smart early warning and response systems. Traditional screening systems rely on static questionnaires or single-source behavioral data, limiting timely identification and intervention. Deep learning models using diverse data streams often face challenges such as poor feature generalization, high inference latency, and limited interpretability, which hinder their use in dynamic academic settings. This study presents the Adaptive Multimodal Psychological Assessment Network (AMPAN), a deep learning architecture designed to integrate educational behavior vectors and multimodal social-emotional embeddings through a dual-stream feature fusion encoder. The encoder captures sequential patterns in survey and academic behavior data using temporal-aware attention blocks, while a cognitive state estimation layer infers latent stress and emotional imbalance. Temporal modelling and transparent reasoning make the framework effective in identifying mental health risks and analysing intervention on an individual basis. AMPAN was evaluated on a static multimodal dataset of 3500 college students, with sequential survey and academic behavior features across four semesters. The model achieved 95.3% accuracy in mental health risk classification and a 93.6% F1-score, outperforming BiLSTM and Multimodal Transformer by 7.8% and 9.1%, respectively. Temporal modeling was applied on ordered data sequences; real-time physiological or LMS/social media streams were not included in this evaluation. AMPAN’s attention visualizations increased counselors’ interpretability confidence by 46%, supporting clear psychological reasoning and enabling future deployment for real-time monitoring.
Xia Hou (Mon,) studied this question.