Identifying Students in Need of Psychological Support: An AI-Driven Approach Using Public Behavioral Data

Background: Mental health challenges are increasingly prevalent among Chinese university students, as reflected in national reports and policy statements. How- ever, traditional psychological assessments,such as self-report scales, which often serve as the initial step in the psychological service process,are limited by issues of ecological validity, recall bias, and operational complexity, particularly when applied to large populations. Recent research suggests that public behavioral data may carry indirect signals of individuals’ psychological states, offering new possibilities for mental health monitoring beyond traditional tools. In this study, public behavioral data refers specifically to non-private, institutionally recorded behavioral records routinely collected by student governance departments,such as hygiene task participation, academic attendance, and disciplinary deductions,for administrative or educational management purposes. These records provide a non-intrusive and scalable means of assessing students’ psychological well-being in real-world settings. This study explores whether machine learning models can effectively identify students who may require psychological support using such publicly recorded behavioral data, providing a cost-effective and ethically feasible alternative to conventional screening methods. Methods: Behavioral and mental health data from 1,255 students at a Chi- nese university were collected with informed consent. Mental health assessments, based on the Chinese College Students’ Mental Health Scale, served as supervised labels, while public behavioral data,covering academic achievement and academic behavioral performance and compliance-related behaviors,were used as indepen- dent variables. Three typical supervised AI learning models (Logistic Regression, K-Nearest Neighbors, and LightGBM) were tested for classification performance. Notably, mental health scale data were only used during model training for label- ing purposes. Once the model was established, predictions could be made solely based on public behavioral data, without requiring any additional psychological assessments.In addition to the main classification task, two exploratory analyses were con- ducted. First, to identify the optimal behavioral observation window for predict- ing psychological support needs, we incrementally accumulated weekly behavioral data over a 17-week period and evaluated model performance at each time point using Accuracy, F1-score, and AUC. Second, to explore intra-individual vari- ability among students potentially in need of psychological support, we tracked their predicted probability scores over time based on weekly updates of dynamic features. Both analyses were implemented using the previously established Light- GBM model, with static features (e.g., demographic and academic performance) held constant while dynamic features (e.g., cumulative disciplinary deductions) were updated weekly. Results: LightGBM outperformed traditional classification algorithms (Logistic Regression and K-Nearest Neighbors) in identifying students potentially in need of psychological support, achieving an AUC of 0.972 and an F1-score of 0.875. In addition to robust classification performance, we found that model metrics peaked when behavioral data were cumulatively observed over a 15-week period. This suggests that approximately 15 weeks of behavioral accumulation may serve as an optimal observation window for identifying students requiring psychologi- cal support. Furthermore, prediction trajectories of individual students revealed notable fluctuations over time, highlighting the dynamic nature of psychologi- cal states and the potential for more personalized, time-sensitive mental health interventions. Conclusion: This study confirms that supervised machine learning can serve as a real-time alternative to traditional psychological assessment methods, enhancing the effciency of university counseling service departments . Additionally, two key findings emerged: (1) determining the minimum data collection period required for effective assessment of mental health status, optimizing university mental health service resource management; and (2) there are significant dynamic change patterns in the psychological states of students who may need psychological sup- port, providing crucial insights for more precise personalized interventions. Future research should explore the generalizability of these findings across different university settings and student populations.