5/ The meta-review agent achieved 78% accuracy in binary classification, remarkably matching the performance of a human evaluator. This system may serve as an interpretable, scalable decision-support tool, offering insights into patients' mental health conditions.

4/ The quantitative assessment agent used an embedding-based few-shot prompting approach to predict symptom-specific scores, achieving an average MAE of 0.619, compared with 0.796 for zero-shot prompting.

3/ The qualitative assessment agent produced specific, coherent, complete, and clinically meaningful assessments, confirmed by both a human reviewer and the judge agent. The feedback loop effectively improved all evaluation metrics.

Can AI agents assess mental disorders like human experts?

Our #ML4H2025 paper presents an LLM-based multi-agent system for assessing depression from clinical interviews. It combines four collaborative agents, iterative self-refinement, and few-shot prompting, achieving human-level performance.