Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents

Huangwei Chen; Wu Li; Junhao Jia; Yining Chen; Xiaotao Pang; YaLong Chen; Gonghui Li; Haishuai Wang

cs.MAApr 10, 2026

Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents

Authors: Huangwei Chen, Wu Li, Junhao Jia, Yining Chen, Xiaotao Pang, YaLong Chen, Gonghui Li, Haishuai Wang,

Organizations: Zhejiang Key Laboratory of Accessible Perception and Intelligent Systems, College of Computer Science and Technology, Zhejiang University · Hangzhou Pujian Medical Technology Co., Ltd, China · School of Computer Science and Technology, Hangzhou Dianzi University · Sir Run Run Shaw Hospital, Zhejiang University School of Medicine

Abstract

The initial outpatient consultation is critical for clinical decision-making, yet it is often conducted by a single physician under time pressure, making it prone to cognitive biases and incomplete evidence capture. Although the Multi-Disciplinary Team (MDT) reduces these risks, they are costly and difficult to scale to real-time intake. We propose Aegle, a synchronous virtual MDT framework that brings MDT-level reasoning to outpatient consultations via a graph-based multi-agent architecture. Aegle formalizes the consultation state using a structured SOAP representation, separating evidence collection from diagnostic reasoning to improve traceability and bias control. An orchestrator dynamically activates specialist agents, which perform decoupled parallel reasoning and are subsequently integrated by an aggregator into a coherent clinical note. Experiments on ClinicalBench and a real-world RAPID-IPN dataset across 24 departments and 53 metrics show that Aegle consistently outperforms state-of-the-art proprietary and open-source models in documentation quality and consultation capability, while also improving final diagnosis accuracy. Our code is available at https://github.com/HovChen/Aegle.

Figures & tables

Figure 1: Single-view vs. virtual MDT reasoning for clinical intake. (a) A single LLM is prone to anchoring, over-focusing on salient symptoms, and producing fragmented notes. (b) A virtual MDT enables decoupled parallel specialist reasoning coordinated by an Orchestrator and integrated by an Aggregator, improving evidence coverage and coherence.

Figure 2: Overview of the Aegle framework. (a) A two-stage consultation workflow consisting of iterative history taking followed by diagnostic synthesis after freezing the case feature set $\mathcal{F}$ . (b) An evolving draft Integrated Patient Note (IPN) that is incrementally updated throughout the consultation. (c) Dynamic multi-agent collaboration, where a context-aware Orchestrator activates relevant specialist agents and an Aggregator integrates their outputs to update the clinical state and generate the next inquiry. (d) Structured clinical state $\mathcal{S}_{t}=[\mathcal{F}_{t},\mathcal{P}_{t}]$ separating evidentiary features from diagnostic and planning components.

Figure 3: Stage I of Aegle, iterative history taking, illustrated with a pediatric heart murmur case. Patient responses are incorporated into the structured SOAP state and analyzed by multiple specialist agents in parallel. Each specialist proposes follow-up questions and targeted updates to case features. The Aggregator integrates these suggestions and generates the next patient-facing inquiry under a write-then-speak protocol.

Model	ClinicalBench												RAPID-IPN
Model	IDEA			SOAP			READ			chrF++			IDEA			SOAP			READ			chrF++
HuatuoGPT-o1-7B † ‡	46.54	$\pm$	11.16	37.98	$\pm$	14.11	68.66	$\pm$	6.30	9.62	$\pm$	1.98	41.89	$\pm$	7.69	40.28	$\pm$	12.96	68.58	$\pm$	3.00	6.66	$\pm$	1.23
Qwen3-8B-think †	43.18	$\pm$	9.22	27.66	$\pm$	10.69	66.74	$\pm$	5.48	10.19	$\pm$	2.11	39.93	$\pm$	8.99	28.92	$\pm$	13.42	66.14	$\pm$	5.67	6.89	$\pm$	1.73
Baichuan-M2-32B † ‡	38.02	$\pm$	9.05	25.58	$\pm$	10.63	64.00	$\pm$	6.93	11.46	$\pm$	2.45	39.40	$\pm$	9.89	31.54	$\pm$	12.25	65.42	$\pm$	5.81	9.24	$\pm$	2.73
Lingshu-32B † ‡	50.50	$\pm$	10.62	39.76	$\pm$	14.92	71.11	$\pm$	5.98	11.53	$\pm$	2.44	51.41	$\pm$	10.66	47.89	$\pm$	13.56	72.07	$\pm$	3.94	8.62	$\pm$	1.89
DeepSeek-V3.2 †	50.51	$\pm$	8.61	38.64	$\pm$	12.57	71.73	$\pm$	4.62	17.33	$\pm$	2.62	54.35	$\pm$	8.76	47.39	$\pm$	10.95	72.14	$\pm$	4.56	14.09	$\pm$	2.69

Table 1: Documentation quality evaluation on ClinicalBench and RAPID-IPN. Unshaded rows report frontier model comparisons across proprietary and open-source baselines. Shaded rows denote the comparison setting with matched base models, where light gray rows denote reasoning-strategy baselines and light blue rows denote medical multi-agent systems. Models marked with

\dagger

are open-source, those with

\ddagger

are medical-domain models, and those with

\star

are proprietary models. Mean

\pm

standard deviation is reported.

Model	ClinicalBench							RAPID-IPN
Model	CA	QT	VER	PJ	SP	AB	Turns	CA	QT	VER	PJ	SP	AB	Turns
HuatuoGPT-o1-7B † ‡	4.01	4.44	4.88	4.15	4.96	4.92	21.67	4.00	4.43	4.89	4.20	4.99	4.98	19.42
Qwen3-8B-think †	4.00	4.10	4.69	4.78	4.94	4.14	7.03	4.00	4.07	4.66	4.68	4.96	4.11	6.86
Baichuan-M2-32B † ‡	3.99	3.97	4.47	4.86	4.89	3.97	11.54	4.00	3.87	4.32	4.91	4.86	3.94	11.10
Lingshu-32B † ‡	4.00	4.52	4.91	4.34	4.98	4.91	8.03	4.00	4.48	4.88	4.30	4.99	4.92	7.36
DeepSeek-V3.2 †	4.01	4.55	4.81	4.31	4.93	4.86	20.50	4.02	4.64	4.79	4.25	4.95	4.90	19.61

Table 2: Consultation Capability Evaluation Results on ClinicalBench and RAPID-IPN. Unshaded rows report frontier model comparisons across proprietary and open-source baselines. Shaded rows denote the comparison setting with matched base models, where light gray rows denote reasoning-strategy baselines and light blue rows denote medical multi-agent systems. Metrics: CA = Conversation Arrangement; QT = Question Types; VER = Verifications; PJ = Professional Jargon; SP = Speech; AB = Amiable Behavior; Turns = number of dialogue turns. Models marked with

\dagger

are open-source, those with

\ddagger

are medical-domain models, and those with

\star

are proprietary models.

Method	Acc. (%)
DeepSeek-V3.2	25.60
CoT	39.60
ToT	38.00
MDAgents	25.73
MedAgents	39.20
Aegle (Ours)	46.93

Table 3: Final diagnosis accuracy on ClinicalBench. All compared methods use DeepSeek-V3.2.

Figure 4: Ablation study results on ClinicalBench and RAPID-IPN. (a) Performance Degradation: The score drop of ablated variants compared to Aegle across five documentation quality metrics. Higher bars indicate a larger contribution of that component to the model’s performance. (b) Dialogue Efficiency: Comparison of consultation turns. Removing Generative Inquiry (w/o GI) or Structured State (w/o SS) leads to significantly longer and less efficient dialogues. (c) Average Drop: The average performance degradation across all metrics, highlighting the Structured State as the most critical component for overall quality.

Model	Architecture	Experts per Case	Experts per Round
MDAgents	Static Multi-Agent	3.702	3.702
MedAgents	Static Multi-Agent	4.968	4.968
Aegle (Ours)	Dynamic Virtual MDT	2.416	1.423

Table 4: Comparison of specialist activation across different multi-agent frameworks. Architecture indicates the underlying expert coordination scheme.

Figure 5: Correlation between LLM-as-a-judge scores and physician ratings on ClinicalBench and RAPID-IPN. Each point represents one evaluated IPN instance ( $n=150$ per dataset). Pearson ( $r$ ) and Spearman ( $\rho$ ) correlation coefficients are reported for IDEA, SOAP, and READ metrics.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10

Figure 11

Figure 12

Figure 13

Figure 14

Sec.	Item	Definition	Max	0	1	2	3	4
1. Present Illness & Comprehensive History
1	1.1 Detailed HPI	Complaint: location, quality, severity, duration, onset, radiation, aggravating/relieving factors.	4	Incorrect/illogical; key elements missing	Some elements present	Most elements present	All elements present	Concise, organized, diagnostically salient
1	1.2 Prior Dx/Tx Course	Prior evaluations/treatments: time, location, tests, meds, interventions; full if first visit.	4	Incorrect/illogical description	Some elements present	Most elements present	All elements present	Concise, structured, informative
1	1.3 Descriptive HPI Language	Medical descriptors (e.g., acute/chronic, sharp/dull, constant/intermittent).	3	Inappropriate/none	Minimal but appropriate	Frequent, appropriate	Consistent, accurate, concise	–
1	1.4 Chronological Organization	Temporal ordering and coherent illness narrative.	3	Temporal contradictions	Disorganized timeline	Mostly coherent; minor gaps	Clear and consistent chronology	–
1	1.5 Contextualized HPI	Integrates relevant PMH, family/social history, and associated symptoms.	4	No/incorrect integration	Partial integration	Comprehensive integration	Clear, accurate, concise	Key info prioritized

Appendix

Table 5: IDEA scoring rubric for clinical note evaluation.

Item	Max	Rubric
S: Subjective
S-1. Format	5	Each major health problem is described separately with clear categorization (e.g., somatic vs. psychological). Fully described problems score 5. If categories are mostly clear but some descriptions are brief, deduct 2–3. If key visit information or diagnostic/treatment details are omitted, categories are confused, or descriptions are fragmented, deduct 4–5.
S-2.1 Chief complaint	2	Concise and accurate summary of the primary discomfort and duration (2). If it is generally clear but not concise, or the duration is vague, deduct 1. If unclear or fails to reflect the main problem, score 0.
S-2.2 Symptoms and clinical course	5	Detailed symptom characteristics (location, quality, severity), frequency, aggravating/relieving factors, and illness trajectory (5). If key information is partially missing, deduct 2–3. If only symptoms are briefly mentioned without describing progression, score 0–2.
S-2.3 Prior evaluation and treatment	3	Prior care is documented, including facility, tests (name/time), diagnoses, medications (name/dose/duration), and response (3). If brief, deduct 1–2. If absent, score 0.
S-2.4 Relevant medical history	3	Comprehensive and accurate past history, including prior diseases, surgeries/trauma, and allergies (3). If 1–2 important elements are missing, deduct 1–2. If largely absent, score 0.

Appendix

Table 6: SOAP scoring rubric for clinical note evaluation.

Code	Meaning	Code	Meaning
A1	Misuse of terminology	A2	Vague expression
B1	Missing important positive findings	B2	Redundant minor positive findings
C1	Negative stated as positive	C2	Positive stated as negative
C3	Missing important negative findings	D1	Irrelevant information
E1	Missing time information	E2	Vague time information
F1	Incorrect order/sequence	F2	Incorrect time value

Appendix

Table 7: Deduction codes used for error annotation.

Item	1	2	3	4	5
1. Structural completeness	Severe omission of core modules (e.g., no HPI, PMH, or physical exam); structure is chaotic, and the basic framework is unrecognizable.	Incomplete core modules (e.g., missing treatment or family history); module order reversed, impairing information retrieval.	Major core modules present (HPI, PMH, physical exam), but minor modules missing (e.g., allergy history); order mostly reasonable.	Core modules complete and in standard order; occasional minor omissions that do not affect understanding.	All modules complete (including auxiliary ones such as personal and reproductive history); strictly follows standard order with clear structure.
2. Logical coherence	No clear timeline or causal relationships; symptom sequence is contradictory, and disease course cannot be reconstructed.	Timeline is vague; symptom evolution contains clear contradictions.	Timeline mostly complete, but relationships between some symptoms are unclear, with occasional logical gaps.	Clear timeline with explicit causal links between symptoms and management; only minor logical issues.	Strict adherence to onset - progression - management - outcome logic with precise timestamps and rigorous causal descriptions.
3. Terminology accuracy	Frequent misuse of medical terms or self-created abbreviations renders core information uninterpretable.	Multiple terminology errors or non-standard abbreviations without clarification, requiring repeated inference.	Occasional imprecise terms or abbreviations that generally follow conventions but need clarification.	Accurate and standardized terminology; all abbreviations are commonly accepted and unambiguous.	Highly precise, condition-specific terminology with clearly defined abbreviations and professional expression.
4. Information redundancy	Large amounts of irrelevant information obscure core content; excessive verbosity overwhelms key findings.	Substantial redundancy or irrelevant content; non-essential information exceeds 20% of the note.	Occasional redundancy or repetition; irrelevant information below 10% and does not impair extraction.	Concise information with no irrelevant content; only minor expressions could be further streamlined.	Highly distilled information with prominent key content and no redundancy or repetition.
5. Information sufficiency	Key information is buried among secondary content and not emphasized, making it easy to miss.	Some key findings are insufficiently highlighted and require careful searching to identify.	Most key information is reasonably placed but not emphasized through formatting or structure.	Key information (e.g., diagnostic evidence or critical values) is clearly highlighted and easy to identify.	All critical information is prominently presented through emphasis, prioritization, or separate sections for immediate recognition.

Appendix

Table 8: Readability rubric for clinical note evaluation.

Item	Tier	Rubric (English)
Inquiry skills
Conversation arrangement	5	The beginning, middle, and end of the consultation are clear and precise, with questions asked in an orderly manner.
	4	Between 5-point and 3-point.
	3	Most of the consultation is conducted in an orderly fashion, but the beginning and ending are not clearly defined.
	2	Between 3-point and 1-point.
	1	The consultation lacks coherence and organization.

Appendix

Table 9: Standardized patient consultation skills grading and scoring criteria.

Figure 6: Department-wise documentation quality comparison on ClinicalBench. IDEA (left) and SOAP (right) scores with 95% confidence intervals are reported across 24 clinical departments. Aegle is compared with representative medical multi-agent baselines (MDAgents and MedAgents). Higher scores indicate better clinical reasoning quality (IDEA) and documentation standardization (SOAP).

Key Clinical Evidence	Aegle	CoT	ToT	MDAgents	MedAgents
Age: 73-year-old male	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Progressive LUTS (weak stream, hesitancy, intermittency, straining, dribbling)	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Nocturia 3–4 times/night	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Intermittent painless light-red hematuria (1–2 $\times$ /day, no clots)	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Acute urinary retention 2 months ago (catheterized)	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
PSA $>$ 155 ng/mL	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

Appendix

Table 10: Coverage of key clinical evidence across different reasoning frameworks.

\checkmark

= explicitly documented;

\triangle

= partially mentioned or ambiguous;

\times

= missing;

Figure 7: The IPN of a 73-year-old male patient suffering from prostate cancer.

Figure 8: The IPN of a 73-year-old male patient suffering from prostate cancer (continued).

Clinical Reasoning

Clinical reasoning studies how AI systems interpret patient evidence, form and revise diagnoses, and support safe decisions across care settings.

See all 180 papers

Jun 2, 2026

MeDxAgent: Multi-Agent Consultation for Interactive Medical Diagnosis

Large language models (LLMs) are increasingly used for health-related decision support. Yet most evaluations treat diagnosis as a single-shot task with complete information provided upfront, often as a multiple-choice selection. This diverges from clinical practice, where diagnosis is interactive and open-ended, involving sequential hypothesis refinement through targeted questioning. We address this gap. We build MeDxBench, a large-scale benchmark of 4,421 clinical cases across 20 specialties. We further propose MeDxAgent, a multi-agent consultation system for interactive diagnosis, and systematically study its prompt-, flow- and agent-level design choices. MeDxAgent achieves a 10.3% accuracy gain over the baseline on MeDxBench, closing 52.3% of the gap to a full-information oracle. We find that specific design choices: collecting demographics first, passing summarized dialogue for diagnosis, and feeding candidate diagnoses for targeted questioning, improve accuracy, mirroring how physicians reason, though their effect emerges fully only in combination. Code and dataset will be released upon publication.

Akshat Sanghvi, Naren Akash, Raza Imam +2

Microsoft Research India · Institute of AI and Innovation, AIG Hospitals · Mohamed bin Zayed University of Artificial Intelligence

Multi-Agent LLM Systems Medical Diagnosis AI Agent Benchmarks Clinical Reasoning Medical AI Agents

Ask PDF

Jun 2, 2026

D2MDT: Department-aware Multidisciplinary Team Consultation with Deliberation for Efficient Clinical Prediction

Electronic health records (EHRs) are central to clinical prediction, but existing methods either rely on correlation-driven deep models or use single large language models (LLMs), making it difficult to support multidisciplinary clinical reasoning. Recent multi-agent systems (MAS) provide a promising alternative, yet current EHR-grounded MAS methods still suffer from weak evidence differentiation across agents and redundant multi-round interaction. We propose D2MDT, a Department-aware MultiDisciplinary Team Consultation with Deliberation for Efficient clinical prediction. D2MDT first constructs structured EHR evidence and consultation-ready semantic evidence for multi-agent consultation. It then assigns patient-specific department perspectives to doctor agents and retrieves complementary evidence for collaborative consultation. To improve efficiency, D2MDT further introduces residual deliberation, which updates only unresolved consensus rather than replaying the full discussion history. Finally, D2MDT fuses the refined consensus report with structured EHR representations for prediction. Experiments on mortality prediction show that D2MDT improves both predictive performance and consultation efficiency. We release the code online to ease the reproducibility of this paper.

Yongqi Liang, Qidong Liu, Chunze Yang +4

Xi’an Jiaotong University, Xi’an, China

Clinical Outcome Prediction Multi-Agent Collaboration Clinical Reasoning Electronic Health Records Mortality Prediction

Ask PDF

Mar 5, 2026

MedCoRAG: Interpretable Hepatology Diagnosis via Hybrid Evidence Retrieval and Multispecialty Consensus

Diagnosing hepatic diseases accurately and interpretably is critical, yet it remains challenging in real-world clinical settings. Existing AI approaches for clinical diagnosis often lack transparency, structured reasoning, and deployability. Recent efforts have leveraged large language models (LLMs), retrieval-augmented generation (RAG), and multi-agent collaboration. However, these approaches typically retrieve evidence from a single source and fail to support iterative, role-specialized deliberation grounded in structured clinical data. To address this, we propose MedCoRAG (i.e., Medical Collaborative RAG), an end-to-end framework that generates diagnostic hypotheses from standardized abnormal findings and constructs a patient-specific evidence package by jointly retrieving and pruning UMLS knowledge graph paths and clinical guidelines. It then performs Multi-Agent Collaborative Reasoning: a Router Agent dynamically dispatches Specialist Agents based on case complexity; these agents iteratively reason over the evidence and trigger targeted re-retrievals when needed, while a Generalist Agent synthesizes all deliberations into a traceable consensus diagnosis that emulates multidisciplinary consultation. Experimental results on hepatic disease cases from MIMIC-IV show that MedCoRAG outperforms existing methods and closed-source models in both diagnostic performance and reasoning interpretability.

Zheng Li, Jiayi Xu, Zhikai Hu +4

School of Cyber Science and Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China · School of Artificial Intelligence, Jilin University, Changchun, 130015, China · Department of Neurology, China-Japan Union Hospital of Jilin University, Changchun, 130033, China +2

Medical Diagnosis Clinical Decision Support Retrieval-Augmented Generation Clinical Reasoning Multi-Agent LLMs

Ask PDF

Abstract

Clinical Reasoning

Explore similar work

MeDxAgent: Multi-Agent Consultation for Interactive Medical Diagnosis

D2MDT: Department-aware Multidisciplinary Team Consultation with Deliberation for Efficient Clinical Prediction

MedCoRAG: Interpretable Hepatology Diagnosis via Hybrid Evidence Retrieval and Multispecialty Consensus