Nature子刊·npj Digital Medicine | 复旦大学陈颉院士团队提出WiseMind框架：为AI装上“理性大脑”与“情感心灵”

【研究摘要】

Multi-Agent LLM Foundation for Psychiatric Diagnosis: Building a “Cognition–Emotion Synergy” AI Doctor Framework. As large language models are increasingly applied to psychiatric diagnosis, two critical limitations hinder real-world use: the lack of structured clinical reasoning pathways and insufficient affective interaction to sustain patient trust during multi-turn interviews. To address this scientific challenge, our team proposed a knowledge-guided multi-agent framework for psychiatric diagnosis, WiseMind, inspired by the Dialectical Behavior Therapy (DBT) theory of “Reasonable Mind” and “Emotional Mind.” The framework adopts a dual-agent collaborative architecture to decouple and coordinate clinical reasoning and empathic communication: the Reasonable Mind Agent conducts structured diagnostic reasoning following DSM-guided pathways, while the Emotional Mind Agent transforms diagnostic intent into empathetic, patient-facing interviewing language, thereby maintaining rigorous diagnostic logic while improving conversational rapport and perceived respect. The novelty lies in introducing a three-layer verification mechanism and, for the first time in psychiatry, adopting AISP as a new testing paradigm for diagnostic systems, establishing a scalable and stress-testable evaluation framework for LLM-based psychiatric diagnosis. Empirically, WiseMind achieved 85.6% first-order diagnostic accuracy across 1,206 simulated clinical dialogues and 180 real-user interactions, significantly outperforming single-agent LLM baselines, while also improving diagnostic node coverage, empathy ratings, and clinician professionalism scores. The structured knowledge-guided reasoning and multi-agent collaborative diagnostic architecture developed in this work provide essential technical foundations for the proposed project’s closed-loop validation system involving AI-doctor, AI-patient, and human expert clinicians. This work was published in npj Digital Medicine (SCI Q1, Top-tier medical journal, IF=15.1), providing strong methodological references for scalable and trustworthy conversational psychiatric diagnosis.

【核心发现】

诊断性能逼近人类医生水平：在针对抑郁、双相、焦虑障碍的模拟鉴别诊断评估中（63类鉴别诊断），WiseMind的Top-1诊断准确率达到85.6%，显著优于各类单智能体基线（提升17-58个百分点），逼近人类精神科医生平均水平（86.0%）。关键节点召回率高达95.6%，证明其推理路径高度贴合临床标准。