A study on a student mental health intervention model combining artificial intelligence and smartphone sensors

In recent years, deep learning-based multimodal emotion recognition methods have been widely applied in student mental health monitoring. However, main-stream models still exhibit limitations in feature coupling and temporal sequence modeling. To address these challenges, this study proposes an enhanced residual spatiotemporal attention network that integrates multi-dimensional perceptual features with residual connection mechanisms, focusing on improving the spatial recognition of key emotional cues and preserving temporal dependencies. An intervention experiment was conducted among a real-world cohort of university students. The results show that the proposed model achieved recognition accuracy above 83% for depression, anxiety, and stress categories. Participants in the intervention group also demonstrated significant improvements in health aware-ness and behavioral performance, with Cohen's d effect sizes generally exceeding 1.2. The findings suggest that the model is well-suited for dynamic emotion recognition and cognitive guidance tasks in university mental health intervention scenarios. This study contributes a viable technical approach and model frame-work for enabling personalized and adaptive mental health interventions in higher education settings.