219. Construction of an ai-based behavioral analysis-based early warning and intervention model for college students' mental health

Abstract Background In recent years, mental health problems among college students have shown multi-factor characteristics, including concealment and dynamics, with a continuous rise in risks such as anxiety, depression, and academic burnout. Traditional monitoring methods relying on self-assessment scales and regular interviews are insufficient in terms of timeliness and objectivity, making it difficult to identify high-risk individuals in a timely manner. However, with the development of artificial intelligence and behavioral computing technologies, psychological state analysis based on multi-source behavioral data has gradually become a research hotspot. Existing research shows that learning, social interaction, daily routines, and internet usage behaviors are closely related to mental health, but these studies mostly remain at the level of single data sources or static assessments, lacking a practical integrated early warning and intervention model. Therefore, this study proposes an AI-based behavioral analysis-based early warning and tiered intervention model for college students' mental health. Through multi-source behavioral data fusion and intelligent modeling, it aims to achieve early identification and precise intervention of psychological risks. Methods This study used 1268 students from a comprehensive university as subjects. For 12 consecutive months, logs from the learning management system, campus card consumption and access records, dormitory daily routines, and anonymized internet usage characteristics were collected. Standard psychological scales were also obtained as a reference. Behavioral feature vectors were constructed using feature engineering and temporal windows. A multimodal neural network incorporating attention mechanisms was introduced to predict psychological risk levels, and the contribution of key behavioral features was analyzed using the SHapley Additive exPlanations (SHAP). Based on this, a three-tiered early warning system of "low-medium-high risk" was established, and a differentiated intervention process was designed. Results The results showed that the research model performed stably in psychological risk identification, with a prediction accuracy of 86.7% and an AUC of 0.912, outperforming logistic regression and single-feature models (p<.01). The recall rate for high-risk individuals was 0.84, and the false positive rate was 12.5%. Feature analysis showed that disrupted nighttime sleep patterns, fluctuating learning behavior, and decreased social interaction were the most discriminative indicators. Longitudinal results showed that the proportion of high-risk students whose scores deteriorated within 3 months was 41.3%, significantly higher than the 8.6% in the low-risk group. Intervention assessment results indicated that stratified intervention reduced the average score of medium- and high-risk students by 21.4% within 6 months, and the intervention effectiveness increased by approximately 18%. Discussion In summary, the AI-based behavioral analysis-based early warning model can achieve early identification and dynamic tracking of psychological risks without increasing the burden on students, demonstrating high practical value. The tiered intervention model helps improve the efficiency of psychological service resource allocation and provides data-driven support for university mental health management. In the future, wearable physiological signals and multi-center data can be combined to further enhance the model's generalization ability, and human-machine collaborative intelligent intervention models can be explored within the framework of privacy and ethics.