Detecção precoce de transtornos de saúde mental em português

Abstract

The present study focuses on the early detection of mental health disorders along the lines of the eRisk shared task series (originally devoted to online mental health discussion domain in the English language) in general-purpose Portuguese social media. More specifically, we adapt a strategy that has won a number of shared tasks to the case of early detection of depression and anxiety disorders in the Brazilian Twitter/X domain, using for this purpose a novel approach based on LLMs and prompt engineering. Our results indicate that the use of LLMs affords greater power to anticipate diagnosis if compared to traditional approaches in the field, and that detection based on general-purpose social media text is potentially more challenging than in the original problem formulation, being dependent on the proximity of the messages to the moment of diagnosis in the chronological order of the timeline on Twitter/X.