Mental health support for schools with wearable biosensor monitoring using deep learning

With the development of wearable technology, it has become an urgent problem for the current campus health support system to realize the intelligent perception and recognition of students' mental health based on micro-electro-mechanical systems sensors. This study focused on multimodal behavioral-physiological signal fusion and proposed a multimodal attention-based psychological state network (MAP-Net). The study constructed a standardized micro-electro-mechanical systems data processing flow to extract behavioral-physiological markers including spectral energy and sample entropy. Moreover, a deep feature extraction structure was designed. By introducing the attention mechanism, the dynamic weighted modeling of multi-source signals was realized, so as to capture the subtle changes of mental states more accurately. The experiment was based on wearable data from 36 high school students over a period of 30 days to compare the performance of MAP-Net with the rest of the models. The results indicated that MAP-Net achieved 96.88%, 94.21%, and 94.91% accuracy, recall, and F1-score under the three-modal inputs, which were higher than the rest of the models. It also performed optimally in anomaly detection rate (91.43%) and warning accuracy (93.12%). Notably, MAP-Net's accuracy decreased by only 3.2% in a long-term robustness evaluation spanning 1,500 min, substantially outperforming competing literature models such as MulT (7.9%) and MDNN-Stress (9.5%) in terms of degradation. These results strongly demonstrated the efficiency and reliability of the model in campus mental health monitoring. It provided a solid technical basis and model support for the construction of a mental health monitoring and early warning system that can be widely promoted and intervened in a timely manner in the campus environment.