HiMA-MDD Detects Depression Using Hierarchical Multi-Agent AI.

Assessing depression from multimodal clinical interviews is a complex task that requires synthesizing diverse evidence from multiple symptoms to construct a coherent PHQ-8 profile. This process is inherently hierarchical: specific evidence is often sparse and context-dependent within individual question-answer exchanges, multiple exchanges collectively inform symptom-level judgments, and the final assessment relies on the consistency of the complete symptom profile. Existing large language model (LLM) systems either process interviews holistically or distribute tasks among generic agents, often lacking an explicit mechanism to coordinate evidence access, scoring authority, feedback, and state recording across these hierarchical levels. To bridge this gap, the authors present HiMA-MDD, a hierarchical multi-agent harness designed to align the assessment hierarchy with three distinct agent layers. After initial non-agentic preprocessing converts multimodal data into context-preserving question-answer units, Layer 1 agents identify potential QA-to-item relationships and route evidence. Layer 2 assigns symptom groups to specialized agents, each responsible for a provisional item score. Finally, Layer 3 audits the complete provisional profile, requests targeted revisions if needed, and reconstructs the verified PHQ-8 profile. This layered architecture naturally produces a Hierarchical Evidence Trace, ensuring that all intermediate evidence, judgments, and revisions are preserved for auditability. The final item scores then deterministically generate the total score and screening decision. Using Qwen2.5-72B-Instruct as the backbone, experiments on the E-DAIC dataset demonstrate that HiMA-MDD significantly outperforms existing state-of-the-art methods, highlighting its effectiveness and interpretability.