Eating disorders (EDs) are one of the major mental health disorders in our society today. Recent works have found that people experiencing an ED often use social media platforms to connect and interact with other like-minded people. These social media data provide a rich source to understand and analyse ED-related issues experienced and discussed by the users. Within this context, a major unresolved research problem is whether ED severity can be predicted from social media posts. To this end, we conduct a thorough study on Twitter (currently$\mathbb {X}$), starting from collecting a large ED-related dataset, manually annotating them, conducting an analysis to investigate the relationship between affective features and ED severity, and developing an advanced deep learning model calledSEEDNet. This model leverages affective computing by embedding emotional and psychological signals such as valence, arousal, and sentiment into both the text representation and lexicon-based graph structure, enabling severity estimation through affect-aware learning. It is used to generate the final output of ED severity through ordinal multiclass classification. In our experiments, we compare our results with several state-of-the-art methods and significantly outperform them (by over 3% in terms of average$F_{1}$-score and accuracy). Beyond this performance gain,SEEDNetoffers additional advantages, including a novel lexicon-guided affective graph and an ordinal-aware loss function. These components not only enhance the model's interpretability but also ensure that predictions better capture the graded nature of ED severity. These findings show the importance of integrating affective computing and deep learning for analysing social media data to improve the assessment of ED severity. This approach offers new opportunities for developing more nuanced and targeted interventions.

Social media platforms provide rich data sources in several domains. In mental health, individuals experiencing an Eating Disorder (ED) are often hesitant to seek help through conventional healthcare services. However, many people seek help with diet and body image issues on social media. To better distinguish at-risk users who may need help for an ED from those who are simply commenting on ED in social environments, highly sophisticated approaches are required. Assessment of ED risks in such a situation can be done in various ways, and each has its own strengths and weaknesses. Hence, there is a need for and potential benefit of a more complex multimodal approach. To this end, we collect historical tweets, user biographies, and online behaviours of relevant users from Twitter, and generate a reasonably large labelled benchmark dataset. Thereafter, we develop an advanced multimodal deep learning model called EDNet using these data to identify the different types of users with ED engagement (e.g., potential ED sufferers, healthcare professionals, or communicators) and distinguish them from those not experiencing EDs on Twitter. EDNet consists of five deep neural network layers. With the help of its embedding, representation and behaviour modeling layers, it effectively learns the multimodalities of social media. In our experiments, EDNet consistently outperforms all the baseline techniques by significant margins. It achieves an accuracy of up to 94.32% and F1 score of up to 93.91% F1 score. To the best of our knowledge, this is the first such study to propose a multimodal approach for user-level classification according to their engagement with ED content on social media.

Eating disorders (EDs) are characterised by abnormal eating habits and obsessive thought about food, weight, shape, and body image. EDs are experienced by a significant portion of our population. Social media is identified as a possible source of influence for EDs, and there is growing evidence of a large amount of ED-related discussions on the Web via social media platforms, such as Twitter. With this growing trend, automatic content analysis for EDs is becoming increasingly important. To date, there does not exist any comprehensive benchmark ED lexicon to identify ED-related conversations that would, in turn, facilitate these content analysis tasks. In this paper, we propose a novel method for generating a lexicon base for ED language, calledEDBase. The method starts with collecting over 3.7 million ED-focused tweets. In order to semantically represent potential ED terminology in a vector space, an ED word embedding model (EDModel) is trained. Then we develop a novel multi-seeded hierarchical density-based algorithm with contrasting corpora for ED lexicon expansion. TheEDModelis queried by the proposed lexicon expansion algorithm to expand the seed terms to a comprehensive lexicon base. OurEDBaseconsists of a (further expandable) list of 3794 high-quality ED terms, quantified by an ED score, and linked to their parent terms. The proposed method significantly outperforms all existing alternative baseline methods and models by over 25% in terms of precision and 1500 in terms of true positives. This research is expected to be impactful in the health data science and healthcare community.