Characterizing and Evaluating Mental Health Misinformation on Social Media: A Qualitative and Deep Learning-Based Study

Social media plays a powerful role in accelerating the spread of misinformation, especially in the mental health domain, where misleading content may even cause disasters. Given the extensive coverage and complexity of social media data, manually moderating online misinformation is infeasible. Therefore, the present study proposes an integrated framework that combines qualitative analysis and deep learning to automatically detect and evaluate mental health misinformation. Guided by expert interviews and grounded theory, in the present study, a 21-level, fine-grained credibility assessment framework covering seven dimensions was developed. Using the framework, in this study, 814 Chinese social media posts were manually annotated, and a high-quality dataset was constructed. On this dataset, we trained and evaluated three deep learning models, that is, Gated Recurrent Unit (GRU), Bidirectional Encoder Representations from Transformers (BERT), and Robustly Optimized BERT Approach (RoBERTa), to automatically assess the credibility of mental health content. The results show that BERT, GRU, and RoBERTa models are effective at leveraging a range of clear sentiment-related cues and surface-level patterns to evaluate mental health misinformation on social media, particularly on dimensions such as Inflammatory Expression and One-sidedness of Expression. However, all three models face challenges in evaluating evidence quality and detecting context-dependent misinformation. When dealing with these challenges, BERT and GRU outperform RoBERTa, particularly in dimensions such as Logical Rigor. This study provides a robust, scalable, and expert-informed approach to improve the credibility of mental health information online.