Leveraging social media footprints for predicting college student anxiety: a machine learning approach.
TL;DR
Public social-media signals were associated with SAS anxiety scores in this consenting, active Weibo-user sample and supported cross-sectional score prediction within the study context, demonstrating the potential for integrating computational approaches with psychological assessment for research on anxiety-related behaviors and digital phenotyping among college students.
Abstract
Background
Anxiety is one of the most prevalent mental health concerns among college students worldwide, yet traditional assessment methods relying on self-report questionnaires are time-consuming, susceptible to response bias, and difficult to scale. Social media platforms, which students use extensively, generate rich behavioral and linguistic data that may reflect underlying psychological states. This study investigates whether passively collected social media footprints are associated with anxiety scores among college students and may support the prediction of anxiety-related patterns.
Methods
In this cross-sectional study, we surveyed 3,211 students across 11 universities in four provinces of China using the Self-Rating Anxiety Scale (SAS), yielding 2,368 valid responses (α = 0.918, KMO = 0.838). With informed consent, we collected 83,968 Weibo posts from participants' public social media accounts. After rigorous preprocessing, 56,060 posts from 1,316 users were retained. We extracted multi-dimensional features spanning linguistic, emotional, cognitive, behavioral, and temporal dimensions. Four machine learning models-Random Forest, XGBoost, LightGBM, and SVR-were trained and compared. SHAP (SHapley Additive exPlanations) values were used for model interpretation.
Results
The Random Forest model achieved the best predictive performance on the test set (R² = 0.77, RMSE = 13.41, MAPE = 25.94%, MSLE = 0.087). SHAP-based interpretation identified grade level, professionalism, emotional expression, risk-related language, curiosity index, emotional tone, and word count as major features contributing to the model. These features align with established psychological dimensions of anxiety symptomatology, including emotional, cognitive, and behavioral indicators.
Conclusions
Public social-media signals were associated with SAS anxiety scores in this consenting, active Weibo-user sample and supported cross-sectional score prediction within the study context. The proposed pipeline offers a transparent, reproducible, and privacy-conscious approach to estimate cross-sectional anxiety-related patterns from publicly available social media data. These findings demonstrate the potential for integrating computational approaches with psychological assessment for research on anxiety-related behaviors and digital phenotyping among college students.