Health Misinformation Detection on Social Networks

Abstract

The spread of health misinformation on social networks has become a critical public health concern, undermining trust in medical advice, promoting harmful behaviors, and weakening responses to global crises like the COVID-19 pandemic. Unlike general misin- formation, health-related falsehoods often exploit scientific language, emotional appeal, and evolving medical knowledge, making them harder to detect and regulate. Social plat- forms, with their rapid content sharing and echo chambers, further amplify misleading narratives before fact-checkers can respond. Manual moderation and traditional detection systems struggle to keep pace with the volume and complexity of content, especially when misinformation is subtle, context- dependent, or shared with misleading intent. The dynamic and informal nature of so- cial media language presents additional challenges for rule-based or keyword-driven ap- proaches. To address these issues, this thesis explores a deep learning-based solution using a Mod- ernBERT architecture. By fine-tuning the model on labeled health misinformation data from social platforms, we demonstrate its ability to capture contextual and linguistic patterns that distinguish false claims from factual ones, providing a foundation for more scalable, automated detection tools. Keywords: health misinformation, social media, BERT, ModernBERT, deep learning, natural language processing, misinformation detection, public health, transformer models, contextual embeddings.