A multimodal deep learning approach for mental health classification of university students: an intelligent early warning system.

Introduction: Mental health issues among university students are becoming increasingly prominent, making an accurate and efficient mental state monitoring system a critical challenge in higher education management. Existing intelligent screening approaches mostly rely on single-modality data and are not tailored to counselor-student dialogue records or structured background information. Consequently, current systems struggle to provide reliable early warnings for high-risk students, especially under highly imbalanced risk-level distributions. Methods: This study develops a multimodal deep learning model, BCL (BERT-CNN-LASSO), which integrates counselor-student dialogue text with static background features (e.g., cumulative GPA, number of failed courses, financial aid status) to classify students into four ordered mental health risk levels for early warning. The model was evaluated on a curated real-world dataset of 600 university students collected from a counseling management system. BCL employs a BERT-based encoder for text representation, a dual-path CNN architecture for multi-scale feature extraction, and LASSO-based selection of static variables to enhance robustness under highly imbalanced risk distributions. Results: In the four-level mental health risk classification task, BCL achieved an overall accuracy of 0.752. At the category level, the model attained an accuracy of 0.897 for the highest-risk group (Level 1) and 0.967 for the lowest-risk group (Level 4), while maintaining a recall of 0.731 for Level 1, thereby reducing missed high-risk cases and limiting false alarms for low-risk students. Discussion: Comparative experiments demonstrate that each core component of BCL contributes to performance gains. These findings indicate that the proposed multimodal approach is a promising technical solution for intelligent early warning of university students' mental health problems.