Research map / Multi-agent systems and games / LLM agents

From Medical Records to Diagnostic Dialogues: A Clinical-Grounded Approach and Dataset for Psychiatric Comorbidity

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

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.

Builds 502 synthetic psychiatric EMRs from Reddit posts and uses a multi-agent LLM framework (HDSM + diagnostic context tree, SCID-5-RV-based) to generate PsyCoTalk, 3,000 Chinese multi-turn comorbidity diagnostic dialogues.

Problem

Existing mental-health dialogue corpora focus on single disorders and lack structured, clinically grounded multi-disorder (comorbidity) diagnostic dialogues for training and evaluating LLM-based screening.

Approach

Converts the SCID-5-RV clinical interview into a hierarchical diagnostic state machine plus a diagnostic context tree that drives a doctor/patient/tool agent system grounded in synthetic EMRs, with a disease-symptom knowledge graph to suppress hallucinated symptom agreement.

Key results

• HDSM-guided multi-agent system reaches 0.31 exact-match diagnostic accuracy vs 0.22 for zero-shot Qwen2.5-72B (McNemar p<0.001).
• Per-label F1 is 0.92 MDD, 0.81 AD, 0.64 ADHD, 0.40 BD.
• Dialogues average 45.9 turns; in a psychiatrist AB test PsyCoTalk scores 5/10 realism vs 6 for real-world dialogues and 1 for MDD-5k.

Surprising findings

• GPT-4o-mini and Deepseek-v3 score below 0.1 exact-match and Qwen3-32B below 0.02 on the diagnostic task, far below the HDSM-guided system.

Limitations

• Covers only four disorders (MDD, AD, BD, ADHD) and six comorbidity combinations.
• Main dataset is Chinese; English experiment is small (205 dialogues).
• Reddit-derived symptom cues may carry demographic or cultural skew; the paper does not validate against real patient outcomes.

Experimental setup

Datasets: PsySym PsyCoProfile (502 synthetic EMRs) D4 MDD-5k AlexanderStreet

Benchmarks: PsyCoTalk diagnostic exact-match

Metrics: exact-match accuracy per-label F1 AB-test realism score 6-dimension expert rating normalized entropy hapax proportion semantic diversity

Baselines: Qwen2.5-72B zero-shot GPT-4o-mini Deepseek-v3 Qwen3-32B Qwen3-8B D4 MDD-5k

Builds on: SCID-5-RV AMC framework MDD-5k PsySym

Closest papers by meaning

• MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents (AAAI 2025 · similarity 0.89)
• MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis (NeurIPS 2025 · similarity 0.87)
• Thousand Voices of Trauma: A Large-Scale Synthetic Dataset for Modeling Prolonged Exposure Therapy Conversations (NeurIPS 2025 · similarity 0.86)
• PatientVLM Meets DocVLM: Pre-Consultation Dialogue Between Vision-Language Models for Efficient Diagnosis (AAAI 2026 · similarity 0.85)
• MedExAgent: Training LLM Agents to Ask, Examine, and Diagnose in Noisy Clinical Environments (NeurIPS 2026 · similarity 0.85)
• TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots (NeurIPS 2026 · similarity 0.84)