Integrating clinical anxiety scales with pre-trained language models for anxiety recognition on social media

Comprehensive Summary

This unicenter, retrospective study investigates whether integrating clinical anxiety scales with pre-trained language models can improve anxiety recognition from social media data, particularly before formal diagnosis. Researchers constructed a dataset consisting of 3170 users with 936400 posts on the Sina Weibo platform, comparing traditional machine-learning and transformer-based models, as well as two clinical anxiety scales - the Self-Rating Anxiety Scale (SAS) and the Hamilton Anxiety Scale (HAMA) - against each user’s anxiety outcomes. The study incorporates 4 modules: The pre-trained embedding module (PEM) using BERT to convert posts and anxiety scale items into mathematical values; an anxiety score calculation module that assigns anxiety scores to each post; the TopK posts selection module that then retains the most anxiety-relevant posts while discarding neutral content; and an anxiety classification module that combines support vector machine-based temporal modeling of anxiety fluctuations of each user over time and TextCNN-based detection of anxiety-related linguistic patterns. Among the multiple comparators, the SAS-based model achieved the best performance (AUROC = 0.9259, F1-Score = 0.9279). However, no temporal or external validation was performed, as the study was not replicated across different time windows, platforms, languages, or cultures. Furthermore, anxiety diagnoses were inferred from self-disclosed diagnostic or treatment-related statements in posts rather than clinician-confirmed diagnoses, limiting generalizability.

Outcomes and Implications

This study implies the importance of social media data as a potential resource for anxiety recognition. Social media data analyzed by models may result in more personalized, context-aware care in addition to traditional care led by clinicians susceptible to biases. Furthermore, the use of social media posts can allow medical professionals to reach underserved populations and marginalized groups long before they are able to enter formal care, which can ultimately lead to more representative population-level mental health insights.