Mixture of Experts for Depression and Anxiety Disorder Prediction From Textual and Non-Textual Social Media Data

In Natural Language Processing (NLP) and related fields, computational models of mental health statuses prediction aim to detect early signs of mental health issues based on an individual’s behaviour on social media. Models of this kind, which are mostly devoted to depression disorders and to the English language, present many open research questions. First, since context-free social media data are prone to noise, the task may involve processing large amounts of data with little or no relation to mental health, which may hinder both model efficiency and accuracy. Second, existing work has been largely devoted to text processing, even though social media also include a wide range of non-textual information, which may be predictors of mental health in their own right. Finally, existing models are usually validated in a single domain, often involving one dataset of a particular text genre and language, and it is not clear whether their results may generalise to other scenarios. Based on these observations, the present work introduces a number of computational models for the prediction of depression and anxiety disorder using large language models (LLM) to handle noisy, context-free social media data. These combine textual and non-textual information with the aid of mixture of experts (MoE), and are evaluated in both the Twitter/X domain in Portuguese and in the Reddit domain in English using both machine learning metrics and human assessment provided by mental health specialists. Results show a number of improvements over the previous work and suggest new lines of investigation.