Large Language Models for Depression Detection: A Review with Prospects of Incomplete Multimodality

Depression has emerged as one of the leading contributors to the global burden of mental disorders, ranking among the top causes of disability worldwide. With its steadily increasing prevalence, accurate and efficient depression detection has become an urgent yet challenging task. In this paper, we provide a comprehensive review of depression-recognition research from multiple perspectives. Specifically, we first outline the epidemiological status of depression and summarize the commonly used depression assessment scales and public datasets. Then, we structurally review the research progress of depression recognition from the perspectives of unimodal analysis and multimodal fusion, with a particular focus on large language model (LLM)-based methods and their potential in addressing challenges arising from incomplete multimodal data. Furthermore, we identify a critical gap in depression recognition under incomplete modality conditions, which are common in real-world clinical scenarios, and outline future directions toward LLM-driven and clinically applicable solutions. Finally, the clinical translation, ethical considerations, and human-centered deployment of LLM-based depression-recognition systems under real-world healthcare constraints are discussed.