A Multi-Modal Few-Shot Learning Framework for Depression Detection Using Video and Audio Features

Early detection of depression plays a crucial role in timely interventions and improved mental health outcomes. Conventional diagnostic approaches, which typically rely on face-to-face interviews and self-reported questionnaires, are not only time-consuming but also vulnerable to subjective bias. Moreover, current automated detection methods face significant challenges due to the scarcity of annotated data and the difficulties in effectively integrating heterogeneous multimodal signals. In this paper, we propose a novel multimodal few-shot learning frame-work designed to address these issues. Our approach introduces two key innovations: a contrastive prototypical learning mechanism that jointly optimizes prototype-based classification and contrastive objectives to enhance feature discrimination, and a dual-stream architecture that independently processes video and audio signals with cross-modal attention fusion to capture complementary information. Extensive experiments on standard depression detection datasets demonstrate that our framework substantially improves detection performance under limited data conditions. Our work not only offers a promising new solution for automated depression detection but also provides valuable insights for future research in the field.