Abstract
This study investigates the trajectory of anxiety severity and prevalence among 1,172 non-graduating undergraduate students in China during the transition to online learning in early 2020. Using the Zung Self-Rating Anxiety Scale (SAS), researchers compared student mental health before and one month after the start of the new semester. The study also evaluates the efficacy of an XGBoost machine learning model in predicting psychological shifts. Results indicate a significant increase in anxiety once the online semester commenced, with the XGBoost model achieving over 80% prediction accuracy, substantially outperforming traditional statistical methods.
Introduction
While early COVID-19 research established widespread psychological distress, few studies examined the longitudinal trajectory of that distress, particularly during the sudden shift to isolated online learning. This study addresses two primary objectives: tracking changes in student anxiety levels between the peak of the initial outbreak and the start of the academic semester, and determining if modern machine learning techniques can provide more accurate early-warning predictions than traditional linear regression models.
Methodology
The study employed a longitudinal survey design involving 1,172 non-graduating undergraduate students (aged 18–22) from 34 provincial-level units across China. Participants were surveyed twice: in mid-February 2020 (during the winter break) and mid-March 2020 (one month into the new online semester).
Psychometric and Technical Approach
- Assessment: The 20-item Zung Self-Rating Anxiety Scale (SAS) was used to screen for anxiety.
- Statistical Analysis: T-tests and single-factor analyses identified demographic trends and significant score changes.
- Machine Learning: Researchers implemented XGBoost (Extreme Gradient Boosting) to build two predictive models. Model 1 utilized 23 features (20 SAS items plus gender, major, and grade), while Model 2 used only the 20 SAS items from the initial test to predict outcomes in the second. These models were benchmarked against traditional multiple stepwise regression.
Results and Discussion
The transition to the new semester coincided with a notable surge in psychological distress. The prevalence of students screening positive for anxiety rose from 15.7% to 18.86%. Statistical analysis confirmed that mean SAS scores were significantly higher in the second test across nearly all demographic subgroups, including gender and academic major.
Machine Learning Performance
The XGBoost models demonstrated high predictive power, successfully capturing non-linear relationships that traditional regression models (which only explained ~27-32% of the variance) failed to account for. The models achieved impressive accuracy in predicting both absolute anxiety levels and the direction of change (increase, decrease, or no change).
Key Insights
- Demographic Vulnerability: Freshmen and sophomores experienced significantly higher anxiety increases compared to juniors, suggesting that older students may possess greater psychological resilience and social adaptive abilities.
- Geographic Sensitivity: Students from Hubei province, the initial epicenter, exhibited higher anxiety levels, indicating a psychological "long tail" even as local infection rates stabilized.
- Feature Importance: The model identified specific physical symptoms—such as tremors ("My arms and legs shake") and fatigue—as the strongest predictors of future anxiety severity.
Conclusion
The findings suggest that the start of a new semester under online learning conditions significantly increased anxiety among Chinese undergraduates. The superior performance of the XGBoost model highlights machine learning as a robust tool for clinical screening and early intervention in university settings. Future research should focus on refining these models with larger datasets and exploring abbreviated screening tools based on the high-importance items identified by the model.
Limitations
The study acknowledges potential self-selection bias due to participant attrition and the reliance on self-reported measures rather than clinical diagnoses. Additionally, the lack of data regarding family infection status remains a confounding variable for future exploration.
Summary updated