Large Language Model–Driven Multimodal Framework for Social Media Depression Detection

Depression, the second leading cause of disability worldwide, is a prevalent and increasingly widespread mental disorder. Traditional diagnostic methods rely on patient communication and subjective cooperation, incurring significant costs in labor, resources, and time. With the accumulation of user data on social media and advancements in large language models (LLMs), computer-aided diagnosis has become feasible, offering more objective and efficient analysis for depression detection and presenting novel research directions. We propose a novel Large Language Model (LLM)-based multimodal framework integrating textual content from historical social media posts and nonlinguistic user behavioral data. Our method detects depressive tendencies with improved accuracy. Experimental results on a public multimodal Twitter dataset demonstrate substantial performance gains over standalone LLMs, achieving state-of-theart effectiveness.