Harnessing Large Language Models for Early Risk Assessment of Anorexia on Social Media

Anorexia is a serious eating disorder characterized by an intense fear of gaining weight and a distorted body image. This often leads to severe self-imposed food restrictions, which can significantly harm both physical and mental health. Early detection is crucial to prevent long-term issues and life-threatening consequences. Social networks have become valuable resources for identifying early signs of anorexia. Individuals frequently share their thoughts, feelings, and behaviors online, creating a digital footprint that can be analyzed for potential warning signs. Psychological research has identified specific linguistic patterns and behavioral markers associated with anorexia, making social media data a promising tool for early detection. In this study, we used the CLEF eRisk 2024 dataset, which consists of writings (posts or comments) from various social media users, to conduct an early risk assessment of anorexia using state-of-the-art large language models. We fine-tuned models such as BERT, RoBERTa, DistilBERT, DistilRoBERTa, MentalBert, ClinicalBert, and GPT-2 to analyze social media posts and identify potential early indicators of anorexia. Our experimental results demonstrate that MentalBERT achieved the highest performance with an accuracy of 98.9% and an F1-score of 98.8% which outperforms previous approaches and achieves superior performance across key evaluation metrics.