Combating mental health misinformation on social media: A knowledge-guided multimodal framework

The proliferation of mental health misinformation on social media poses severe public health risks, necessitating automated detection systems. However, existing misinformation detection research remains limited in its ability to perform stance-aware evidence reasoning and robust multimodal analysis. In this study, we propose a Knowledge-Guided Mental Health Multimodal Misinformation Detection Network ((Formula presented) ). The framework contains three key modules: the Dual-Channel Attention Verification module differentiates supporting and contradictory evidence from knowledge graphs; the Statistical-Guided Attention Prior module uses statistical cues to guide attention over noisy multimodal representations; and the Directed Co-Attention Integration module hierarchically integrates verified knowledge with multimodal content. We conduct experiments on the MentalMisinfo dataset, including 2135 YouTube videos and 754 BitChute videos. (Formula presented) achieves the best overall performance on both datasets, improving F1-score by 0.0142 on YouTube and 0.0238 on BitChute over the strongest baselines, with statistically significant gains across 10 random seeds. Ablation analyses further confirm the contribution of each component. Our work advances knowledge-enhanced multimodal misinformation detection and provides practical implications for social media platforms and public-health stakeholders.