Depression remains widely underdiagnosed and undertreated because stigma and subjective symptom ratings hinder reliable screening. To address this challenge, we propose a coarse-to-fine, multi-stage framework that leverages large language models (LLMs) for accurate and interpretable detection. The pipeline performs binary screening, five-class severity classification, and continuous regression. At each stage, an LLM produces progressively richer clinical summaries that guide a multimodal fusion module integrating text, audio, and video features, yielding predictions with transparent rationale. The system then consolidates all summaries into a concise, human-readable assessment report. Experiments on the E-DAIC and CMDC datasets show significant improvements over state-of-the-art baselines in both accuracy and interpretability.
@article{arxiv.2604.11334,
title = {Dynamic Summary Generation for Interpretable Multimodal Depression Detection},
author = {Shiyu Teng and Jiaqing Liu and Hao Sun and Yu Li and Shurong Chai and Ruibo Hou and Tomoko Tateyama and Lanfen Lin and Yen-Wei Chen},
journal= {arXiv preprint arXiv:2604.11334},
year = {2026}
}