Research map / Multi-agent systems and games / LLM agents
MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MoodAngels, a retrieval-augmented multi-agent LLM framework with item-level scale analysis and debate for mood disorder diagnosis, plus the MoodSyn synthetic psychiatric case dataset (1,173 cases).
Problem
Automated mood disorder diagnosis is hard due to subjective clinical assessments, symptom overlap across disorders, and privacy limits on sensitive psychiatric data.
Approach
Decomposes psychiatric scales into item-level questions selected by Pearson correlation, pairs DSM-5 criteria retrieval and case retrieval across three agent variants with different reliance on past cases, and aggregates their diagnoses through a judge-led debate; a TabSyn-based synthetic dataset supports privacy-preserving research.
Key results
- Angel.R with DeepSeek-V3 reaches ACC 0.927 and MCC 0.841 on real data versus 0.847 ACC for DeepSeek-V3 prompted directly; GPT-4o-based Angel.R reaches ACC 0.920 vs GPT-4o 0.797 (about 12.3 points).
- multi-Angels with GPT-4o achieves the best real-data recall of 0.881 among GPT-4o-based variants.
- On synthetic MoodSyn test data, Angel.C and multi-Angels (DeepSeek-V3) reach ACC 0.821 and Macro F1 0.821 versus 0.821 ACC for DeepSeek-V3 prompted baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply MoodAngels instead of in-context GPT-4o | Mood disorder diagnosis · Real clinical cases | accuracy: 0.797 -> 0.920 | Helped |
| Add retrieved cases as context (Angel.D rather than Angel.R) | MoodAngels with GPT-4o · Real clinical cases | accuracy: 0.920 -> 0.923 | Helped |
| Compare retrieved cases with the current case (Angel.C rather than Angel.R) | MoodAngels with GPT-4o · Real clinical cases | accuracy: 0.920 -> 0.914 | Hurt |
| Add debate-based synthesis of the independent agents | MoodAngels with GPT-4o · Real clinical cases | accuracy: 0.920 -> 0.925 | Helped |
| Add retrieved cases as context (Angel.D rather than Angel.R) | MoodAngels with Deepseek-V3 · MoodSyn test | accuracy: 0.8 -> 0.807 | Helped |
| Add debate-based synthesis of the independent agents | MoodAngels with Deepseek-V3 · MoodSyn test | accuracy: 0.8 -> 0.821 | Helped |
| Use structured rather than unstructured records for symptom matching (ablation) | Angel.R · Real clinical cases | accuracy: 0.920 -> 0.918 | Hurt |
| Return structured rather than unstructured records to the agent (ablation) | Angel.R · Real clinical cases | accuracy: 0.918 -> 0.914 | Hurt |
| Use unfiltered total scores from all 13 scales instead of selected questions (ablation) | Angel.R · Real clinical cases | accuracy: 0.920 -> 0.852 | Hurt |
Surprising findings
- Unfiltered total scale scores lowered Angel.R accuracy by 6.8 points, showing that item-level selection matters more than total scores.
- Narrative-style medical records gave better diagnostic performance than structured records passed to the agent.
Limitations
- Only medical records and scale scores are used, without clinician interview information, limiting replication of full diagnosis.
- Evaluated on a single hospital's data, with a small mood-disorder test subset (56 cases), so results may not generalize.
- Binary mood disorder classification only; no subtype diagnosis.
Experimental setup
Datasets: real-world hospital clinical cases (2,804 entries; 2,243 for retrieval, 561 for test) DSM-5 diagnostic criteria
Benchmarks: real-world clinical test set (561 cases) MoodSyn synthetic test set (140 cases)
Metrics: Recall ACC MCC Macro F1
Baselines: LLaMA3-8B-Instruct Mistral-7B-Instruct-v0.3 GPT-4o DeepSeek-V3 medgemma-27b-text-it
Builds on: TabSyn BGE-M3 DSM-5 criteria
Closest papers by meaning
- AgentMental: An Interactive Multi-Agent Framework for Explainable and Adaptive Mental Health Assessment (AAAI 2026 · similarity 0.88)
- MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents (AAAI 2025 · similarity 0.87)
- From Medical Records to Diagnostic Dialogues: A Clinical-Grounded Approach and Dataset for Psychiatric Comorbidity (ICLR 2026 · similarity 0.87)
- MedAgent-Pro: Towards Evidence-based Multi-modal Medical Diagnosis via Reasoning Agentic Workflow (ICLR 2026 · similarity 0.84)
- Mental-Perceiver: Audio-Textual Multi-Modal Learning for Estimating Mental Disorders (AAAI 2025 · similarity 0.84)
- Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.