Paper deep dive
Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents
Huangwei Chen, Wu Li, Junhao Jia, Yining Chen, Xiaotao Pang, Ya-Long Chen, Li Gonghui, Haishuai Wang, Jiajun Bu, Lei Wu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/14/2026, 1:37:34 AM
Summary
Aegle is a synchronous virtual Multi-Disciplinary Team (MDT) framework that utilizes a graph-based multi-agent architecture to improve clinical intake. By formalizing the consultation state using a structured SOAP representation and employing a dynamic orchestrator to manage specialist agents, Aegle decouples evidence collection from diagnostic reasoning, thereby reducing cognitive biases and improving documentation quality and diagnostic accuracy.
Entities (5)
Relation Signals (3)
Aegle → evaluatedon → ClinicalBench
confidence 100% · Experiments on ClinicalBench and a real-world RAPID-IPN dataset... show that Aegle consistently outperforms
Aegle → uses → SOAP
confidence 100% · Aegle formalizes the consultation state using a structured SOAP representation
Aegle → builtupon → DeepSeek-V3.2
confidence 95% · Built upon DeepSeek-V3.2 (DeepSeek-AI et al., 2025), Aegle instantiates a constrained graph-based agentic topology.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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 this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2604.08927v1
- Canonical: https://arxiv.org/abs/2604.08927v1
Trouble viewing inline? Open PDF directly →
Full Text
99,230 characters extracted from source content.
Expand or collapse full text
Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents Huangwei Chen 1,2,4 , Wu Li 1 , Junhao Jia 1,2,4 , Yining Chen 3 , Xiaotao Pang 2 , Yalong Chen 3 , Gonghui Li 3 , Haishuai Wang 1,† , Jiajun Bu 1,† , Lei Wu 1,2† , 1 Zhejiang Key Laboratory of Accessible Perception and Intelligent Systems, College of Computer Science and Technology, Zhejiang University 2 Hangzhou Pujian Medical Technology Co., Ltd, China 3 Sir Run Run Shaw Hospital, Zhejiang University School of Medicine 4 School of Computer Science and Technology, Hangzhou Dianzi University Correspondence: haishuai.wang@zju.edu.cn, bjj@zju.edu.cn, shenhai1895@zju.edu.cn 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 syn- chronous virtual MDT framework that brings MDT-level reasoning to outpatient consulta- tions via a graph-based multi-agent architec- ture. Aegle formalizes the consultation state using a structured SOAP representation, sepa- rating evidence collection from diagnostic rea- soning to improve traceability and bias control. An orchestrator dynamically activates special- ist agents, which perform decoupled parallel reasoning and are subsequently integrated by an aggregator into a coherent clinical note. Ex- periments on ClinicalBench and a real-world RAPID-IPN dataset across 24 departments and 53 metrics show that Aegle consistently out- performs 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. 1 Introduction The trajectory of clinical care is fundamentally es- tablished during the initial consultation (Starfield et al., 2005). In this pivotal phase, a physician must transmute a patient’s unstructured narrative of symptoms and concerns into a structured med- ical record, crystallizing it as the Initial Progress Note (IPN) in the SOAP (Subjective, Objective, Assessment, Plan) format. This document serves as more than a mere administrative summary; it Patient Input: “Severe chest pain, occasional acid reflux, mild backache. ” chest pain (a) Traditional Single-View Reasoning (Susceptible to Anchoring Bias) Single LLM Fragmented Output Potential cardiac issue. (Other symptoms ignored) (b) Virtual MDT (Decoupled Parallel Reasoning) Patient Input: “Severe chest pain, occa- sional acid reflux, mild backache. ” Orchestrator Cardiology Agent chest pain Gastroenterology Agent acid reflux Musculoskeletal Agent backache Aggregator 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 frag- mented notes. (b) A virtual MDT enables decoupled par- allel specialist reasoning coordinated by an Orchestrator and integrated by an Aggregator, improving evidence coverage and coherence. is the cornerstone for downstream diagnostic deci- sions and treatment planning (Krishna et al., 2021). Consequently, the comprehensiveness and accuracy of this intake process are paramount to effective healthcare delivery, serving as the bedrock upon which the entire clinical pathway rests. However, achieving high-quality intake in rou- tine practice is fraught with cognitive and systemic challenges (You et al., 2025). The traditional pro- cess typically relies on a single physician operating under significant time constraints. When formu- lating diagnoses while simultaneously engaging in empathetic dialogue, individual providers are sus- ceptible to anchoring bias (Croskerry, 2013), fix- ating on prominent symptoms while overlooking subtler, yet critical, diagnostic clues. As illustrated in Fig. 1(a), this “single-view” setting narrows the exploration of the diagnostic space and can lead to fragmented evidence capture in the resulting note. This is not merely a matter of physician compe- arXiv:2604.08927v1 [cs.MA] 10 Apr 2026 tence but a fundamental limit of human cognitive bandwidth when processing high-entropy patient narratives under time pressure. To address complex cases where a single perspective is insufficient, medical practice tra- ditionally turns to a Multi-Disciplinary Team (MDT) (Taylor et al., 2010). By aggregating spe- cialists from diverse fields, MDT enables parallel and complementary reasoning across different clin- ical perspectives, mitigating the anchoring effects inherent to single-view decision making (Fig. 1(b)). While this collaborative model significantly re- duces the risk of oversight, it is inherently resource- intensive, asynchronous, and difficult to scale. Or- ganizing a team of human experts for every routine outpatient consultation is logistically impractical. Thus, a critical gap remains: how can we transpose the systematic depth of MDT-level reasoning to the widespread, real-time outpatient intake phase without prohibitive resource costs? In this paper, we bridge this gap by proposing Aegle, a multi-agent framework that virtualizes the MDT paradigm. Rather than relying on a single Large Language Model (LLM) or a static chain of agents, Aegle introduces a novel computational architecture for medical inquiry. We posit that the essence of effective collaboration lies in decoupled parallel reasoning, where distinct specialist agents analyze the case from their unique domain perspec- tives without interference, followed by a semantic aggregation phase. Furthermore, to address the efficiency issues common in multi-agent systems, we implement a dynamic topology controlled by a meta-cognitive orchestrator. This allows the sys- tem to adaptively scale its reasoning network based on the real-time completeness of the clinical docu- mentation. Our contributions are summarized as follows: •We propose Aegle, a Synchronous Virtual MDT framework that leverages decoupled par- allel reasoning to transcend physical resource constraints. This paradigm transposes system- atic, inpatient-level diagnostic depth into real- time outpatient inquiries, significantly enhanc- ing robustness while mitigating single-view cognitive biases. •We propose a State-Aware Dynamic Topol- ogy that aligns multi-agent collaboration with the evolving clinical document. By imple- menting on-demand specialist activation, this mechanism dynamically constructs reason- ing paths tailored to case-specific ambiguity, thereby maximizing the diagnostic signal-to- noise ratio and ensuring high-density informa- tion gathering. •We conduct a comprehensive evaluation across 24 clinical departments using 53 fine- grained metrics. Empirical results demon- strate Aegle’s superiority over state-of-the-art baselines in diagnostic accuracy and documen- tation quality, establishing a robust benchmark for next-generation clinical AI assistants. 2 Related Works 2.1 LLMs for Clinical Consultation and Documentation LLMs have shown promise in clinical workflow op- timization tasks such as clinical documentation sup- port and conversational assistance during patient intake (Zhou et al., 2025a). In the realm of docu- mentation, models function as semantic compres- sors, transforming unstructured dialogues into stan- dardized formats like SOAP notes (Krishna et al., 2021). While models such as Med-PaLM 2 have achieved accuracy comparable to human scribes in summarizing static records (Singhal et al., 2025), they exhibit significant fragility in temporal reason- ing. Specifically, when summarizing longitudinal patient trajectories, these models often succumb to the “lost-in-the-middle” effect, failing to accurately distinguish between historical ailments and current presenting symptoms, thereby compromising the integrity of the medical record (Kruse et al., 2025; Zeng et al., 2025). Conversely, in interactive consultation, frame- works such as AMIE (Tu et al., 2025) and Health- care Agent (Ren et al., 2025) have attempted to simulate the diagnostic inquiry process. Despite their conversational fluency, a critical limitation persists: these monolithic systems largely oper- ate as passive information receivers (Zhou et al., 2025b). Rather than executing proactively asking rule-out questions to narrow the differential diagno- sis space, they tend to hallucinate details or prema- turely commit to a diagnosis based on incomplete user input (Qiu et al., 2025). This passivity reveals a fundamental misalignment with real-world intake, where the core challenge lies not merely in process- ing available text, but in the strategic elicitation of missing evidence (Brooks et al., 2024). 2.2 Multi-Agent Systems for Clinical Reasoning To transcend the cognitive bounds of single-model architectures, Multi-Agent Systems (MAS) have emerged as a computational analogue to the clini- cal MDT (Han et al., 2025). By assigning special- ized roles such as oncologists, radiologists, and pathologists to distinct agents, frameworks like MedAgents (Tang et al., 2024), MAC (Chen et al., 2025), and MedCollab (Zhan et al., 2026) lever- age dialectical debate or role-specialized collab- oration to decompose complex diagnostic tasks. MDAgents introduces an adaptive topology that dynamically structures collaboration based on the perceived medical complexity of the case, thereby optimizing the trade-off between accuracy and com- putational cost (Kim et al., 2024). Furthermore, frameworks such as ClinicalLab have demonstrated the utility of agentic collaboration in managing multi-departmental diagnostics, simulating the re- ferral and consultation dynamics of a physical hos- pital (Yan et al., 2025). Related ideas have also been explored in the broader multi-agent literature. ChatEval applies multi-agent debate to evaluation rather than clin- ical reasoning (Chan et al., 2024), while sparse communication topologies have been shown to reduce redundant exchanges in debate-based sys- tems (Li et al., 2024). DyLAN dynamically selects agent teams and interaction structures based on the task (Liu et al., 2023), and OSC studies cognitive orchestration through dynamic knowledge align- ment in multi-agent collaboration (Zhang et al., 2025). In contrast, Aegle targets interactive clinical intake, where coordination must remain grounded in an explicitly structured SOAP state and in a staged separation between evidence elicitation and diagnostic synthesis. However, the “black-box” interaction between agents introduces varying degrees of collaborative failure modes. A recent large-scale audit of med- ical MAS reveals that agentic collaboration can lead to “flawed consensus” where agents reinforce each other’s biases, and the suppression of correct minority opinions during the voting process (Gu et al., 2025). Additionally, the interplay of multiple probabilistic models creates a problem of “com- pound opacity”, making it exponentially difficult to trace the provenance of a clinical error (Salehi et al., 2025). These vulnerabilities highlight that while MAS can broaden the hypothesis space, they require rigid structural constraints to prevent unan- chored speculation, a gap that our proposed frame- work specifically addresses. 3 Methodology We propose Aegle, a multi-agent consultation framework designed to virtualize the cognitive ben- efits of MDT collaboration during early-stage pa- tient encounters. As illustrated in Fig. 2, Aegle integrates structured, virtualized MDT-style col- laboration directly into the consultation workflow. By coordinating multiple specialized agents dur- ing information gathering, the framework aims to surface overlooked considerations earlier and to reduce bias arising from single-perspective reason- ing. Built upon DeepSeek-V3.2 (DeepSeek-AI et al., 2025), Aegle instantiates a constrained graph-based agentic topology. Agent interactions are governed by explicit state representations and execution pro- tocols, enabling controllable, transparent, and bias- aware clinical dialogue. 3.1 Structured Clinical State To ground multi-agent collaboration in established clinical practice, Aegle formalizes the consultation state using the canonical SOAP schema. We denote the clinical state at turntasS t . Beyond its role as a documentation standard, SOAP provides a cog- nitive structure that explicitly separates evidence collection from diagnostic interpretation, thereby supporting bias-aware reasoning. We decomposeS t into two functionally distinct components: •Case Features (F). Corresponding to the Subjective and Objective sections of SOAP,F serves as an incremental repository of factual evidence. It accumulates verifiable patient information throughout the consultation, in- cluding Basic Information, History of Present Illness, Past Medical History, Physical Exami- nation, and Auxiliary Examination results. •Diagnosis and Plan (P). Corresponding to the Assessment and Plan sections,Prepre- sents the analytical output of the consultation. It includes the preliminary diagnosis, diagnos- tic reasoning, and treatment plan, all of which are derived exclusively from the finalized case features inF . Ga te Stage I: InterativeHisto ry TakingStage I: Diagostic Sy nthe sis Sta teSta te 퓕 Orchest ratorSpecialistsAggregatorOrchest ratorSpecialistsAggregatorFrozen 퓕 Transiti on (a) (b) Initial P rogress Note Case Fea tu res Di agnosis and Pla n IPN (c) Di al og ue History Clinical Stat e Orchestrat or Car diol og y Special ist Patholog y Special ist Rad iol og y Special ist Next Inquiry Updated IPN Aggregat or (d) Prelimina ry Diagno sis Diagnostic Discussion Treatment Plan Basic Inf ormation History of Present IllnessPast Medical History Phy sical Examina tionAux iliary Examina tion Case Features 퓕 Hello. May I ask your gender and age first? Could you tell me whait has been bothering you today? Hello, doctor. I’m 34 years old women. I’m here be- cause I’ve been experienc- ing palpitations lately. Initial Progre s Note Case Fea tures Diagnosis and Plan 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 setF. (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 stateS t = [F t ,P t ]separating evidentiary features from diagnostic and planning components. The structured state is defined asS t = [F t ,P t ] and functions as a shared blackboard accessible to all agents. Aegle enforces a unidirectional de- pendency fromFtoPsuch that diagnostic and planning components may only be generated after evidence stabilization. This constraint explicitly links clinical conclusions to accumulated evidence, ensuring traceability and mitigating premature com- mitment to unsupported hypotheses. 3.2 Multi-Agent Graph Topology Aegle operationalizes virtual MDT collaboration through a dynamic multi-agent graph topology composed of three types of nodes, each fulfilling a distinct role in the consultation workflow. Orchestrator.The Orchestrator acts as a routing and coordination policyπ orch that governs agent activation. It does not perform medical reasoning itself. Instead, it allocates computational attention by selecting a subset of specialist agents based on the evolving consultation context: A sub ,ι = π orch (H t ,F t ), A sub ⊆A total ,(1) whereH t denotes the dialogue history andιspec- ifies context-dependent task instructions. This se- lective activation mechanism mirrors real-world MDT practice by engaging specialized expertise only when warranted by the available evidence, thereby avoiding unnecessary or premature expert involvement during early-stage information gather- ing. Specialist Agents.Each specialist agent operates as an independent domain expert, analyzing the clinical state from a distinct medical perspective. Specialists are executed in parallel and generate proposed updates to the clinical state in isolation. This decoupled architecture preserves hypothesis diversity by construction and delays consensus for- mation, reflecting the cognitive advantage of inde- pendent expert opinions in MDT discussions. Aggregator. The Aggregatorπ agg serves as the interface between internal agent reasoning and patient-facing communication. It follows a write- then-speak protocol. First, it validates and inte- grates specialist proposals to update the structured clinical state: S t+1 = π write agg S t ,∆S (a) t a∈A sub .(2) Subsequently, it generates the patient-facing utter- ance conditioned solely on the updated state: u t+1 = π speak agg (S t+1 ).(3) This separation ensures internal consistency and technical precision of the medical record while maintaining clear and empathetic communication with the patient. 3.3 Sequential Clinical Execution Building upon the structured clinical state and de- fined agent roles, Aegle executes consultations through a two-stage finite state machine. This temporal structure enforces a strict separation be- tween evidence acquisition and diagnostic reason- ing, serving as an explicit bias-control mechanism. Stage I: Iterative History Taking.As shown in Fig. 3, the consultation begins with iterative his- tory taking. The Orchestrator activates relevant specialist agents based on patient responses and the current completeness ofF t . Each specialist examines the updated clinical state from its domain perspective and proposes follow-up questions to- gether with evidence-centric revisions to the draft Integrated Patient Note. The Aggregator integrates these parallel propos- als, updates the case features inF, and generates the next consultation question. This process con- tinues until all mandatory fields inFare either populated or explicitly marked as unavailable by the patient, ensuring that downstream diagnostic reasoning is grounded in sufficient evidence. Stage I: Diagnostic Synthesis.OnceFis final- ized, it is frozen to prevent further modification, and the system transitions deterministically to di- agnostic synthesis. In this stage, the Orchestrator commissions specialist agents to perform indepen- dent diagnostic reasoning based on the same fixed evidentiary substrate. Specialists propose diagnos- tic hypotheses and treatment considerations with- out introducing new inquiries. The Aggregator then integrates these heteroge- neous perspectives to produce the final diagnosis and planP, resolving inconsistencies and gener- ating a complete and coherent SOAP note. This staged execution ensures that diagnostic conclu- sions are derived exclusively from stabilized evi- dence, reinforcing traceability and reducing bias induced by early hypothesis fixation. 4 Experiments 4.1 Datasets To evaluate Aegle’s performance across diverse clinical scenarios, we utilize two distinct datasets: ClinicalBench. We employ ClinicalBench (Yan et al., 2025), a comprehensive end-to-end bench- mark derived from de-identified electronic health records (EHRs) of top-tier Grade 3A hospitals in China. It contains 1,500 cases covering 24 clin- ical departments and 150 diseases. Crucially, it enforces a strict data-leakage-free protocol and sup- ports open-ended generation tasks, simulating the whole trajectory from triage to treatment planning. RAPID-IPN. To evaluate complex differen- tial diagnosis,we curated the Real-world Abdominal Pain Integrated Diagnostic-pathway Initial Progress Note (RAPID-IPN) dataset from a top-tier Grade 3A hospital in Eastern China. Span- ning from 2018 to 2024, this cohort comprises 322 patients with abdominal pain across 12 de- partments, encompassing internal medicine (e.g., Cardiology, Gastroenterology) and surgery (e.g., Hepatobiliary Surgery). A rigorous review proto- col by three senior physicians (>5 years of expe- rience) ensured that the standardized SOAP notes and treatment plans strictly aligned with patients’ actual clinical trajectories, thereby guaranteeing real-world fidelity. Due to patient privacy regu- lations and institutional data governance policies, the RAPID-IPN dataset cannot be publicly released. All data were fully de-identified in accordance with local regulations prior to use, and the study proto- col was reviewed and approved by the hospital’s ethics committee. 4.2 Experimental Setup Baselines and Evaluation Paradigm. To rigor- ously benchmark our framework, we report two complementary comparison settings. We first com- pare Aegle against a broad set of proprietary and open-source single-LLM baselines to situate its performance among frontier models. We then conduct a fixed-backbone comparison in which CoT (Wei et al., 2022), ToT (Yao et al., 2023), MDAgents (Kim et al., 2024), MedAgents (Tang Hello. Could you please tell me the patient’s gender and age first? What brings you in today, and what symptoms or concerns are you experiencing? Dear parents, first of all, please rest assured that many children are found to have heart murmurs during physical examinations, but most of them are physiological and do not necessarily indicate serious problems. We need to have a detailed understanding of the child's condition to make an accurate judgment. According to the advice of the specialist doctor, I have a few questions that I need to know 1. You mentioned that you recently had a follow-up echo- cardiography. Could you tell me the specific results of the examination? For instance, have any abnormalities in the heart structure, valve problems been found, or what spe- cific diagnosis did the doctor give? 2. Regarding the child's growth and development, could you tell me what her current height and weight are? Is her growth and development curve normal? Please speak slowly. There's no need to rush. Hello, doctor. I am the child’s mother. My daughter is 3 years and 5 months old. We came to the hospital because a heart murmur was found during her physical examination. It has been noted for about two years now, and we recently had a follow-up echocardiogram. The doctor suggested that we come here for further evalua- tion. Orchestrator Activated specialist: ['cardiology', 'pediatrics'] Instruction: The patient is a 3-year-old and 5-month-old girl. A heart murmur was detected during a physical examination two years ago. After a recent re-examination with a color Doppler ultrasound of the heart, the doctor rec- ommended that she seek medical attention. Please have two experts collaborate to collect a complete medical history. Cardiologist (i) Suggested questions: As a pediatrician, I need to understand the growth and development of children. May I ask what your daughter's height and weight are? Is her growth and development curve normal? (i) Suggestions for draft revision: 1. Basic information: Write "Patient, female, 3 years and 5 months old, visited the hospital due to 'heart murmur detected during physical examination 2 years ago'" into basic_infomation 2. History of present illness: Write "The child was found to have a heart murmur during a physical examination two years ago. After a recent re-examination with a color Doppler ultrasound of the heart, the doctor sug- gested seeking medical attention" into history_of_pre- sent_illness Pediatricians Suggested questions: As a cardiologist, I need to have a detailed understand- ing of the child's condition. First of all, could you tell me the specific result of the recent echocardiogram re-ex- amination? For instance, have any abnormalities in the heart structure, valve problems been found, or what is the specific diagnosis? Suggestions for draft revision: 1. Basic information: Write "Patient, female, ... 2. History of present illness: Write "The child was found to have a heart murmur ... Initial P rogress Note Case Fea tu res Diagnosis & Plan Basic Information: The patient is a 3-year-5-month-old female who present- ed to the hospital with a chief complaint of "a cardiac murmur detected on physical examination for 2 years." (blank) History of Present Illness: A cardiac murmur was first noted during a routine physi- cal examination 2 years ago. Following a recent follow-up echocardiographic evaluation, the physician recommended further consultation." Past Medical History: (blank) Physical Examination: (blank) Auxiliary Examination: (blank) Aggregator 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. et al., 2024), and Aegle all use DeepSeek-V3.2, allowing us to isolate the effect of reasoning and collaboration structure. All rubric-based results are evaluated under an LLM-as-a-judge paradigm using gpt-4o-mini, with identical scoring prompts across all conditions to ensure fair and consistent comparison. We further conducted a small-scale human evaluation to validate the reliability of the LLM-as-a-judge paradigm; details are provided in Appendix A. EvaluationMetrics. Weadoptamulti- dimensional evaluation framework that assesses both the consultation process and the resulting clinical documentation. Specifically, documenta- tion quality is evaluated along clinical reasoning (IDEA), documentation standardization (SOAP), readability (READ), and surface-level similarity (chrF++).Consultation capability is assessed using a consultation skills rubric covering inquiry skills and humanistic care.To complement these rubric-based assessments with an objective correctness signal, we additionally report final diagnosis accuracy on ClinicalBench. Detailed metric introductions and rubrics are provided in Appendix C. 4.3 Documentation Quality Evaluation The documentation quality results in Table 1 re- veal a clear and consistent performance advantage for Aegle across both evaluation settings. In the frontier model comparison, Aegle remains compet- itive with strong proprietary and open-source mod- els. In the fixed-backbone comparison, it also out- performs strong reasoning-strategy baselines (CoT and ToT) as well as existing medical multi-agent systems (MDAgents and MedAgents). The gains are most pronounced in metrics related to inter- nal coherence and evidential grounding, reflecting a shift from surface-level summarization toward more structured clinical reasoning. In single-model baselines, IPN often exhibit a fa- miliar failure mode: fluent narratives that read well locally but lack global alignment between history, assessment, and plan. While CoT-style prompting partially alleviates this issue by improving local rea- soning consistency, it does not explicitly constrain how evidence is accumulated and reused across sections. Aegle mitigates this limitation by en- forcing an explicit separation between case feature accumulation and diagnostic synthesis. As a result, diagnostic conclusions and management plans are consistently traceable to previously documented evidence, reducing omissions. Model ClinicalBenchRAPID-IPN IDEASOAPREADchrF++IDEASOAPREADchrF++ HuatuoGPT-o1-7B †‡ 46.54 ± 11.16 37.98 ± 14.11 68.66 ± 6.309.62 ± 1.9841.89 ± 7.6940.28 ± 12.96 68.58 ± 3.006.66 ± 1.23 Qwen3-8B-think † 43.18 ± 9.2227.66 ± 10.69 66.74 ± 5.48 10.19 ± 2.1139.93 ± 8.9928.92 ± 13.42 66.14 ± 5.676.89 ± 1.73 Baichuan-M2-32B †‡ 38.02 ± 9.0525.58 ± 10.63 64.00 ± 6.93 11.46 ± 2.4539.40 ± 9.8931.54 ± 12.25 65.42 ± 5.819.24 ± 2.73 Lingshu-32B †‡ 50.50 ± 10.62 39.76 ± 14.92 71.11 ± 5.98 11.53 ± 2.4451.41 ± 10.66 47.89 ± 13.56 72.07 ± 3.948.62 ± 1.89 DeepSeek-V3.2 † 50.51 ± 8.6138.64 ± 12.57 71.73 ± 4.62 17.33 ± 2.6254.35 ± 8.7647.39 ± 10.95 72.14 ± 4.56 14.09 ± 2.69 DeepSeek-V3.2-Thinking † 46.73 ± 9.6834.62 ± 12.71 69.53 ± 4.90 16.17 ± 2.8049.37 ± 11.12 41.98 ± 13.42 70.72 ± 4.69 12.56 ± 2.92 GLM-4.6 † 47.02 ± 8.9634.70 ± 11.87 68.48 ± 5.00 16.73 ± 2.4348.23 ± 11.47 40.44 ± 13.68 68.92 ± 5.43 13.31 ± 2.82 Kimi K2 Thinking † 54.05 ± 9.5742.86 ± 12.70 70.76 ± 6.51 17.10 ± 2.8955.26 ± 10.00 49.45 ± 10.79 70.55 ± 5.77 13.92 ± 2.60 MiniMax-M2 † 57.78 ± 11.02 46.18 ± 12.46 73.87 ± 7.05 16.41 ± 2.9063.01 ± 10.91 56.84 ± 10.52 79.74 ± 9.28 14.81 ± 3.43 GPT-4o ⋆ 41.05 ± 9.7329.38 ± 12.75 67.66 ± 5.20 10.83 ± 1.9844.70 ± 9.8034.79 ± 13.51 69.89 ± 3.66 10.84 ± 2.14 Gemini 2.5 ⋆ 48.35 ± 11.45 35.58 ± 15.63 70.10 ± 9.38 17.82 ± 3.3449.89 ± 11.10 39.03 ± 15.49 71.25 ± 4.75 14.69 ± 3.14 Qwen3-Max ⋆ 61.75 ± 9.5653.40 ± 11.25 74.99 ± 4.57 17.73 ± 2.4860.82 ± 7.8357.84 ± 8.5374.29 ± 4.19 14.66 ± 2.42 Doubao-Seed-1.6 ⋆ 51.51 ± 8.8539.80 ± 11.20 69.03 ± 5.30 18.04 ± 2.7251.11 ± 8.9044.37 ± 9.9568.35 ± 4.63 14.38 ± 2.61 ERNIE-5.0-Preview ⋆ 47.09 ± 9.9935.55 ± 11.93 67.19 ± 5.73 15.82 ± 2.6945.05 ± 9.4940.03 ± 11.59 67.03 ± 5.13 12.41 ± 2.67 CoT64.72±9.0160.34±7.0976.94±4.7425.83±2.9765.61±7.5761.89±7.5478.57±4.9521.23±3.55 ToT66.53±9.7562.34±6.6478.46±5.4925.96±3.1467.58±6.8563.96±6.1779.86±5.5021.77±3.13 MDAgents68.41±9.6156.45±8.5378.81±6.7624.83±2.5366.58±10.0560.98±8.9279.95±8.4121.28±2.73 MedAgents59.85±9.3954.48±7.9572.00±5.8823.66±2.3361.51±8.8157.19±7.4274.30±5.1722.04±2.33 Aegle (Ours)72.78±10.1663.02±5.5777.55±5.4125.83±2.5571.52±8.3563.92±4.7379.93±6.9024.24±2.44 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†are open-source, those with‡are medical-domain models, and those with ⋆ are proprietary models. Mean± standard deviation is reported. On metrics such as READ and chrF++, Aegle is comparable to the strongest reasoning baselines on both datasets, suggesting that linguistic flu- ency has largely saturated among modern LLMs. Consequently, further gains in clinical documen- tation quality depend less on wording and more on how information is structured, prioritized, and constrained. 4.4 Diagnostic Accuracy To complement the rubric-based note-quality met- rics with an objective correctness signal, we addi- tionally evaluate final diagnosis accuracy on Clini- calBench, where standardized diagnostic labels are available. We compare Aegle against DeepSeek- V3.2, CoT, ToT, MedAgents, and MDAgents under a shared DeepSeek-V3.2 backbone. Unlike the original ClinicalBench formulation, our evaluation starts from the consultation phase: the model must first elicit evidence through interaction before pro- ducing a diagnosis, rather than being given the complete post-consultation record as input. We therefore report these results as end-to-end diagno- sis accuracy under the clinical intake setting. As shown in Table 3, Aegle achieves the highest diagnosis accuracy, outperforming the underlying DeepSeek-V3.2 model by 21.33 points and sur- passing both reasoning-strategy baselines and prior medical multi-agent systems. This result is con- sistent with the IDEA and SOAP gains in Table 1, indicating that Aegle improves not only note struc- ture but also final diagnostic decisions under the same backbone. 4.5 Consultation Capability Evaluation The consultation capability results in Table 2 show that Aegle’s strengths are mainly reflected in how consultation information is elicited and validated, rather than in stylistic or expressive aspects of dia- logue. Across both benchmarks, Aegle consistently demonstrates a more directed questioning strategy, in which dialogue turns are organized around re- solving clinically relevant uncertainties. When compared with single-model baselines and reasoning-strategy baselines such as CoT and ToT, Aegle exhibits a more structured pattern of information verification. Its higher VER and QT scores indicate that follow-up questions are more frequently used to confirm or refine patient- provided information, instead of extending the con- versation through loosely related prompts. This pattern is characteristic of specialist-driven history taking, where each question serves a specific diag- nostic purpose. Conversational style and humanistic expression metrics remain close to their upper bounds for most competitive models, which limits their discrimi- native value. In terms of dialogue length, Aegle occupies a middle range, avoiding both very short interactions that may under-verify critical details and excessively long exchanges that dilute diag- nostic focus. By maintaining this balance while Model ClinicalBenchRAPID-IPN CAQTVERPJSPABTurnsCAQTVERPJSPABTurns HuatuoGPT-o1-7B †‡ 4.014.444.884.154.964.9221.674.004.434.894.204.994.9819.42 Qwen3-8B-think † 4.004.104.694.784.944.147.034.004.074.664.684.964.116.86 Baichuan-M2-32B †‡ 3.993.974.474.864.893.9711.544.003.874.324.914.863.9411.10 Lingshu-32B †‡ 4.004.524.914.344.984.918.034.004.484.884.304.994.927.36 DeepSeek-V3.2 † 4.014.554.814.314.934.8620.50 4.024.644.794.254.954.9019.61 DeepSeek-V3.2-Thinking † 4.004.864.964.574.994.978.804.004.844.924.405.004.988.26 GLM-4.6 † 4.004.654.974.795.004.988.214.004.604.944.645.004.988.10 Kimi K2 Thinking † 4.024.884.934.584.994.978.964.014.854.874.385.004.978.14 MiniMax-M2 † 4.044.774.774.444.994.859.144.044.344.054.685.004.5721.23 GPT-4o ⋆ 4.004.744.934.915.004.564.604.004.884.944.845.004.493.84 Gemini 2.5 ⋆ 3.883.584.124.894.384.0029.983.953.604.094.894.283.9630.00 Qwen3-Max ⋆ 4.014.904.964.385.004.966.424.014.844.974.345.004.965.41 Doubao-Seed-1.6 ⋆ 4.004.454.884.614.994.6610.434.004.324.824.565.004.6610.04 ERNIE-5.0-Preview ⋆ 4.004.554.924.594.974.395.064.004.364.844.474.974.364.51 CoT4.004.834.634.054.994.856.404.004.784.484.285.004.866.66 ToT4.004.864.654.004.994.897.374.004.844.564.155.004.917.91 MDAgents4.004.664.854.324.984.936.594.004.764.944.575.004.716.61 MedAgents4.004.584.904.555.004.806.964.004.624.924.575.004.716.61 Aegle (Ours)4.024.954.944.215.005.0010.164.034.964.934.195.005.008.84 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†are open-source, those with‡are medical-domain models, and those with⋆are proprietary models. MethodAcc. (%) DeepSeek-V3.225.60 CoT39.60 ToT38.00 MDAgents25.73 MedAgents39.20 Aegle (Ours)46.93 Table 3: Final diagnosis accuracy on ClinicalBench. All compared methods use DeepSeek-V3.2. achieving strong verification and question cover- age, Aegle demonstrates a consultation behavior that is both efficient and clinically grounded, which is consistent across the two benchmarks. 4.6 Expert Activation Efficiency Beyond output quality, the practicality of multi- agent systems critically depends on the efficient utilization of expert resources during reasoning. Table 4 compares the average number of activated specialists across MDT-style frameworks. Static multi-agent baselines such as MDAgents and MedAgents activate a fixed set of experts at every dialogue turn, resulting in identical expert counts per case and per round. In contrast, Aegle employs a state-aware dynamic topology that acti- vates specialists on demand. This design substan- ModelArchitectureExperts per CaseExperts per Round MDAgentsStatic Multi-Agent3.7023.702 MedAgentsStatic Multi-Agent4.9684.968 Aegle (Ours)Dynamic Virtual MDT2.4161.423 Table 4: Comparison of specialist activation across dif- ferent multi-agent frameworks. Architecture indicates the underlying expert coordination scheme. tially reduces redundant expert invocation, achiev- ing fewer activated experts per case and, more im- portantly, per round, without sacrificing diagnos- tic performance. The results highlight that Aegle improves not only clinical reasoning quality but also computational efficiency, which is essential for real-time outpatient deployment. 4.7 Ablation Study To reveal the roles each component plays in Aegle, we conduct comprehensive ablation experiments and visualize the results in Figure 4. As shown in Figure 4(a) and (c), among all vari- ants, removing the structured clinical state leads to the most severe degradation across both datasets, particularly in IDEA and SOAP scores. This confirms that explicitly separating case features from diagnostic outputs is essential for maintain- ing evidence-grounded reasoning and standardized 0102030 ClinicalBench +16.08 +22.27 +4.39 +12.34 IDEA 02040 +12.11 +24.38 +0.95 +13.27 SOAP 0.02.55.07.5 +4.24 +5.82 +0.87 +3.14 READ 051015 +1.18 +8.50 -0.78 +2.98 chrF++ 0.00.20.4 +0.20 +0.11 +0.05 +0.26 Cons. 01020 RAPID-IPN +14.27 +17.17 +4.05 +10.60 01020 +10.27 +16.53 +1.20 +7.79 0510 +6.55 +7.79 +2.25 +3.62 051015 +7.95 +10.15 +0.86 +7.07 0.00.20.4 +0.24 +0.09 +0.02 +0.26 w/o generative inquiryw/o structured statew/o dynamic topologyw/o decoupled reasoning (a) 102030 w/o GI w/o S w/o DT w/o DR +23.8 +10.3 -1.5 -1.6 ClinicalBench Aegle Ablated Variant 102030 +25.2 +10.8 -0.9 -0.5 RAPID-IPN Aegle Ablated Variant w/o GIw/o SSw/o DTw/o DR 0.0 2.5 5.0 7.5 10.0 12.5 ClinicalBenchRAPID-IPN (b) (c) 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 S) 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. documentation. Eliminating generative inquiry produces a dif- ferent failure mode. In this setting, history taking follows a fixed template based on Bates’ Guide to Physical Examination and History Taking (Bick- ley and Szilagyi, 2012). While surface-level docu- mentation quality remains relatively high, reason- ing quality deteriorates and, as illustrated in Fig- ure 4(b), dialogue length increases dramatically. The sharp rise in dialogue turns indicates inefficient and unfocused information gathering, suggesting that without context-aware questioning, the system struggles to converge on a sufficiently informative case representation. By comparison, removing dynamic topology or decoupled reasoning results in more moderate but systematic performance drops (Figure 4(a)). With- out dynamic specialist activation, the system loses adaptability to case-specific ambiguity, while re- moving decoupled reasoning reduces hypothesis diversity and increases the risk of premature con- vergence. Taken together, these results suggest that Aegle’s performance gains do not arise from any single mechanism, but from the coordinated interaction of structured state representation, active inquiry, adaptive expert selection, and independent specialist reasoning. 5 Conclusion In this paper, we presented Aegle, a virtualized MDT framework that elevates the quality of outpa- tient consultation by enabling synchronous, multi- perspective reasoning within the intake workflow. By enforcing a structural separation between evi- dence acquisition and diagnostic synthesis within a dynamic multi-agent topology, Aegle effectively mitigates the cognitive biases and premature clo- sure inherent in single-view models. Extensive evaluations on ClinicalBench and RAPID-IPN dataset demonstrate that Aegle consistently out- performs state-of-the-art baselines in documenta- tion quality and consultation capability, while an additional diagnostic-accuracy study on Clinical- Bench further shows improved final decisions. To- gether, these results establish a robust and scalable paradigm for next-generation clinical decision sup- port systems. Acknowledgments This work was supported by the National Nat- ural Science Foundation of China (Grant No. 62372408) and Hangzhou Pujian Medical Tech- nology Co., Ltd, China and ZJU-Pujian Research & Development Center of Medical Artificial Intel- ligence for Hepatobiliary and Pancreatic Disease. Limitations Despite its strong empirical performance, Aegle has several limitations that warrant careful consid- eration. First, the multi-agent paradigm inevitably introduces additional inference overhead. Dynamic routing, parallel specialist execution, and struc- tured aggregation increase end-to-end latency com- pared with single-model generation, which may prolong user waiting time in real-time outpatient settings where responsiveness is critical. Second, maintaining a state-aware collaboration grounded in a continuously evolving SOAP record leads to longer effective contexts. As the dialogue pro- gresses, the accumulated structured state and inter- mediate agent outputs expand the context window, increasing token consumption and computational cost, and potentially constraining deployment un- der strict resource budgets. Third, while fully de- coupled parallel reasoning is central to preserv- ing hypothesis diversity and mitigating premature convergence, it can also yield redundant or over- lapping recommendations across specialists. Such repetition may complicate aggregation by diluting genuinely novel signals. Future work should there- fore explore mechanisms that better balance cogni- tive bias mitigation and content redundancy, such as diversity-aware expert prompting, redundancy- penalized aggregation, or adaptive expert selection strategies. References Elizabeth A. Baker, Cynthia H. Ledford, Louis Fogg, David P. Way, and Yoon Soo Park. 2015. The IDEA assessment tool: Assessing the reporting, diagnostic reasoning, and decision-making skills demonstrated in medical students’ hospital admission notes. Teach- ing and Learning in Medicine, 27(2):163–173. Lynn Bickley and Peter G. Szilagyi. 2012. Bates’ guide to physical examination and history-taking. Lippin- cott Williams & Wilkins. Katherine C. Brooks, Katie E. Raffel, David Chia, Ab- hishek Karwa, Colin C. Hubbard, Andrew D. Auer- bach, and Sumant R. Ranji. 2024. Stigmatizing lan- guage, patient demographics, and errors in the diag- nostic process. JAMA Internal Medicine, 184(6):704– 706. Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. 2024. ChatEval: Towards better LLM-based eval- uators through multi-agent debate. In The Twelfth International Conference on Learning Representa- tions. Xi Chen, Huahui Yi, Mingke You, WeiZhi Liu, Li Wang, Hairui Li, Xue Zhang, Yingman Guo, Lei Fan, Gang Chen, Qicheng Lao, Weili Fu, Kang Li, and Jian Li. 2025. Enhancing diagnostic capability with multi- agents conversational large language models. npj Digital Medicine, 8(1):159. Pat Croskerry. 2013. From mindless to mindful practice: Cognitive bias and clinical decision making. New England Journal of Medicine, 368(26):2445–2448. DeepSeek-AI, Aixin Liu, Aoxue Mei, Bangcai Lin, Bing Xue, Bingxuan Wang, Bingzheng Xu, Bochao Wu, Bowei Zhang, Chaofan Lin, Chen Dong, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chen- hao Xu, Chong Ruan, Damai Dai, Daya Guo, De- jian Yang, and 193 others. 2025. DeepSeek-V3.2: Pushing the frontier of open large language models. Preprint, arXiv:2512.02556. Lei Gu, Yinghao Zhu, Haoran Sang, Zixiang Wang, Dehao Sui, Wen Tang, Ewen Harrison, Junyi Gao, Lequan Yu, and Liantao Ma. 2025. MedAgentAu- dit: Diagnosing and quantifying collaborative failure modes in medical multi-agent systems. Preprint, arXiv:2510.10185. Xudong Han, Xianglun Gao, Xiaoyi Qu, and Zhenyu Yu. 2025. Multi-agent medical decision consensus matrix system: An intelligent collaborative frame- work for oncology MDT consultations. Preprint, arXiv:2512.14321. Health Human Resources Development Center, Na- tional Health Commission of China. 2024. Standard scheme for the final clinical practice ability assess- ment of standardized resident training. Yubin Kim, Chanwoo Park, Hyewon Jeong, Yik Siu Chan, Xuhai Xu, Daniel McDuff, Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, and Hae Won Park. 2024. MDAgents: An adaptive collaboration of LLMs for medical decision-making. In Advances in Neural Information Processing Systems. Oral pre- sentation. Kundan Krishna, Sopan Khosla, Jeffrey Bigham, and Zachary C. Lipton. 2021. Generating SOAP notes from doctor-patient conversations using modular summarization techniques. In Proceedings of the 59th Annual Meeting of the Association for Compu- tational Linguistics and the 11th International Joint Conference on Natural Language Processing (Vol- ume 1: Long Papers), pages 4958–4972, Online. As- sociation for Computational Linguistics. Maya Kruse, Shiyue Hu, Nicholas Derby, Yifu Wu, Samantha Stonbraker, Bingsheng Yao, Dakuo Wang, Elizabeth M. Goldberg, and Yanjun Gao. 2025. Large language models with temporal reasoning for longitudinal clinical summarization and prediction. In Findings of the Association for Computational Lin- guistics: EMNLP 2025, pages 20715–20735, Suzhou, China. Association for Computational Linguistics. Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. 2024. Improv- ing multi-agent debate with sparse communication topology. In Findings of the Association for Com- putational Linguistics: EMNLP 2024, pages 7281– 7294, Miami, Florida, USA. Association for Compu- tational Linguistics. Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. 2023. A dynamic LLM-powered agent net- work for task-oriented agent collaboration. CoRR, abs/2310.02170. Pengcheng Qiu, Chaoyi Wu, Junwei Liu, Qiaoyu Zheng, Yusheng Liao, Haowen Wang, Yun Yue, Qianrui Fan, Shuai Zhen, Jian Wang, Jinjie Gu, Yanfeng Wang, Ya Zhang, and Weidi Xie. 2025. Evolving diagnostic agents in a virtual clinical environment. Preprint, arXiv:2510.24654. Zhiyao Ren, Yibing Zhan, Baosheng Yu, Liang Ding, Pingbo Xu, and Dacheng Tao. 2025. Healthcare agent: Eliciting the power of large language models for medical consultation. npj Artificial Intelligence, 1(1):24. Sara Salehi, Yashbir Singh, Parnian Habibi, and Bradley J. Erickson. 2025. Beyond single systems: How multi-agent AI is reshaping ethics in radiology. Bioengineering, 12(10):1100. Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R. Pfohl, Heather Cole-Lewis, Dar- lene Neal, Qazi Mamunur Rashid, Mike Schaeker- mann, Amy Wang, Dev Dash, Jonathan H. Chen, Nigam H. Shah, Sami Lachgar, Philip Andrew Mans- field, and 16 others. 2025. Toward expert-level medi- cal question answering with large language models. Nature Medicine, 31(3):943–950. Barbara Starfield, Leiyu Shi, and James Macinko. 2005. Contribution of primary care to health systems and health. The Milbank Quarterly, 83(3):457–502. Xiangru Tang, Anni Zou, Zhuosheng Zhang, Ziming Li, Yilun Zhao, Xingyao Zhang, Arman Cohan, and Mark Gerstein. 2024. MedAgents: Large language models as collaborators for zero-shot medical rea- soning. In Findings of the Association for Com- putational Linguistics: ACL 2024, pages 599–621, Bangkok, Thailand. Association for Computational Linguistics. Cath Taylor, Alastair J. Munro, Rob Glynne-Jones, Clive Griffith, Paul Trevatt, Michael Richards, and Amanda J. Ramirez. 2010. Multidisciplinary team working in cancer: What is the evidence?BMJ, 340:c951. Tao Tu, Mike Schaekermann, Anil Palepu, Khaled Saab, Jan Freyberg, Ryutaro Tanno, Amy Wang, Brenna Li, Mohamed Amin, Yong Cheng, Elahe Vedadi, Nenad Tomašev, Shekoofeh Azizi, Karan Singhal, Le Hou, Albert Webson, Kavita Kulkarni, S. Sara Mahdavi, Christopher Semturs, and 7 others. 2025. Towards conversational diagnostic artificial intelligence. Na- ture, 642(8067):442–450. Chen Wang, Shuang Li, Ning Lin, Xin Zhang, Yu Han, Xia Wang, Dong Liu, Xue Tan, Di Pu, Kun Li, Gang Qian, and Rui Yin. 2025. Application of large lan- guage models in medical training evaluation: Us- ing ChatGPT as a standardized patient—multimetric assessment. Journal of Medical Internet Research, 27:e59435. Yijie Wang, Yining Chen, and Jifang Sheng. 2024. As- sessing ChatGPT as a medical consultation assis- tant for chronic hepatitis B: Cross-language study of english and chinese. JMIR Medical Informatics, 12:e56426. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompt- ing elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Associates, Inc. Weixiang Yan, Haitian Liu, Tengxiao Wu, Qian Chen, Wen Wang, Haoyuan Chai, and Jiayi Wang. 2025. ClinicalLab: Aligning agents for multi-departmental clinical diagnostics in the real world. In The Thirty- ninth Annual Conference on Neural Information Pro- cessing Systems. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models. In Advances in Neural Information Processing Systems, volume 36, pages 11809–11822. Curran Associates, Inc. Jacqueline G. You, Reema H. Dbouk, Adam Land- man, David Y. Ting, Sayon Dutta, Julie C. Wang, Amanda J. Centi, Molly Macfarlane, Eran Bechor, Jonathan Letourneau, Gabrielle Choo-Kang, Es- ther H. Kim, Cordula Magee, Brian J. Lang, Laura Angelo, Jackson Olin, Michelle Frits, Christine Ian- naccone, Angela Rui, and 7 others. 2025. Ambient documentation technology in clinician experience of documentation burden and burnout. JAMA Network Open, 8(8):e2528056. Sihang Zeng, Yujuan Fu, Sitong Zhou, Zixuan Yu, Lu- cas Jing Liu, Jun Wen, Matthew Thompson, Ruth Etzioni, and Meliha Yetisgen. 2025. Traj-CoA: Pa- tient trajectory modeling via chain-of-agents for lung cancer risk prediction. In NeurIPS 2025 Workshop on GenAI4Health. Yuqi Zhan, Xinyue Wu, Tianyu Lin, Yutong Bao, Xi- aoyu Wang, Weihao Cheng, Huangwei Chen, Feiwei Qin, and Zhu Zhu. 2026. Medcollab: Causal-driven multi-agent collaboration for full-cycle clinical di- agnosis via ibis-structured argumentation. Preprint, arXiv:2603.01131. Jusheng Zhang, Yijia Fan, Kaitong Cai, Xiaofei Sun, and Keze Wang. 2025. OSC: Cognitive orchestra- tion through dynamic knowledge alignment in multi- agent LLM collaboration. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2025, pages 6320–6337, Suzhou, China. Association for Computational Linguistics. Juexiao Zhou, Haoyang Li, Siyuan Chen, Zhangtianyi Chen, Zhongyi Han, and Xin Gao. 2025a. Large language models in biomedicine and healthcare. npj Artificial Intelligence, 1(1):44. Zhanke Zhou, Xiao Feng, Zhaocheng Zhu, Jiangchao Yao, Sanmi Koyejo, and Bo Han. 2025b. From pas- sive to active reasoning: Can large language models ask the right questions under incomplete informa- tion? In Proceedings of the Forty-second Interna- tional Conference on Machine Learning. 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.664 (p <1e-4) ρ=0.632 (p <1e-4) ClinicalBench | READ Model DeepSeek-V3.2 GPT-4o Qwen3-Max 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.769 (p <1e-4) ρ=0.810 (p <1e-4) ClinicalBench | SOAP Model DeepSeek-V3.2 GPT-4o Qwen3-Max 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.862 (p <1e-4) ρ=0.866 (p <1e-4) ClinicalBench | IDEA Model DeepSeek-V3.2 GPT-4o Qwen3-Max 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.798 (p <1e-4) ρ=0.809 (p <1e-4) RAPID-IPN | SOAP Model DeepSeek-V3.2 GPT-4o Qwen3-Max 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.840 (p <1e-4) ρ=0.853 (p <1e-4) RAPID-IPN | IDEA Model DeepSeek-V3.2 GPT-4o Qwen3-Max 0.00.20.40.60.81.0 LLM 0.0 0.2 0.4 0.6 0.8 1.0 Doctor n=150 r=0.625 (p <1e-4) ρ=0.595 (p <1e-4) RAPID-IPN | READ Model DeepSeek-V3.2 GPT-4o Qwen3-Max 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 = 150per dataset). Pearson (r) and Spearman (ρ) correlation coefficients are reported for IDEA, SOAP, and READ metrics. A Analysis of the reliability of LLM-as-a-judge To assess the reliability of the LLM-as-a-judge eval- uation paradigm used throughout our experiments, we conducted a targeted human evaluation study and compared physician ratings with the scores produced by the judge model. For each dataset (ClinicalBench and RAPID- IPN), we randomly sampled 50 cases and se- lected three representative models for compari- son (DeepSeek-V3.2, GPT-4o, and Qwen3-Max), resulting in a total of 150 evaluated instances per dataset. Licensed physicians independently scored the generated IPNs using the same evalu- ation rubrics as those employed in the automatic assessment, covering IDEA, SOAP, and READ di- mensions. Figure 5 presents the correlation analysis between physician scores and LLM-as-a-judge scores. Across both datasets and all evaluated metrics, we observe statistically significant positive correla- tions. On ClinicalBench, Pearson correlation coef- ficients range from 0.664 (READ) to 0.862 (IDEA), with corresponding Spearman rank correlations ranging from 0.632 to 0.866. On RAPID-IPN, Pear- son correlations range from 0.625 (READ) to 0.840 (IDEA), with Spearman correlations ranging from 0.595 to 0.853. All correlations are significant with p < 10 −4 . These results indicate that the LLM-as-a-judge scores are broadly consistent with physician judg- ments across different datasets and evaluation di- mensions. While minor discrepancies remain, the overall alignment supports the validity of using LLM-based evaluation as a scalable proxy for large- scale comparative experiments. B Prompt Design We present the prompt design for Standardized Pa- tient, Orchestrator, Specialist, and Aggregator in this section. Standardized Patient You are a professionally trained standardized patient actor, simulating a realistic medical consultation with a physician. Below is your complete case information: <Case Information> case_infomation </Case Information>t **Role-Playing Principles** 1. **Based on the information**: All of your responses must be strictly based on the case text above. 2. **No fabrication**: Do not imagine, extend, or invent any medical history or details that are not provided in the case. If the doctor asks about information that is completely absent from the materials, respond naturally with phrases such as “I’m not sure,” “I didn’t notice,” or “I don’t have that issue.” 3. **Non-professional stance**: You are an ordinary patient and should not provide medical explana- tions, professional interpretations, or diagnostic suggestions. 4. **Natural expression**: Use conversational, natural, emotionally realistic first-person language (“I...”). Do not list information like a robot. **Expression Requirements** - Respond as a real patient speaking to a doctor, avoid mechanical answers, lists, or summaries. - When describing experiences, symptoms, and feelings, focus on subjective experience (e.g., severi- ty, duration, impact, emotions). - If professional terms are involved (such as test results mentioned in the case), paraphrase them natu- rally, or only read them out in detail when the doctor specifically asks for the report. **Exceptions** - For objective information such as test results, laboratory data, or imaging findings, you may report them truthfully according to the case when the doctor asks. - If the doctor presses for specific details that do not exist in the case materials (e.g., exact numerical values or redacted information), clearly state that you cannot provide them. You will now answer the doctor’s questions in the role of the patient. Orchestrator You are a Medical MDT Orchestrator. **Current phase: current_phase_description** Phase goal: phase_goal Your responsibilities: 1. **Efficient decision-making**: For simple confirmation questions, routine information gathering, or single-dimension follow-up inquiries, you should generate the question directly without activating specialist physicians. 2. For complex situations, dynamically dispatch specialist Agents (Specialists) and issue them instruc- tions that **align with the current phase goal**. **Decision logic**: - **Simple cases**: When only specific numerical values need to be supplemented, past medical history needs to be confirmed, or simple details need to be clarified → **keep `active_specialists` empty** and place the question directly into `suggested_question`. - **Complex cases**: When differential diagnosis, multi-system symptoms, or in-depth analysis is required → activate the relevant Specialists. **Strictly output in the specified JSON format** **You do not make medical diagnoses; you are responsible only for coordination and orchestra- tion.** historytakingdiagnostic_synthesis current_phase_description: Phase 1: History Taking phase_goal: Efficiently collect the patient’s case features, fill in information gaps, and clarify symptom details. Focus on documenting the patient’s actual clinical information: basic information, history of present illness, past medical history, physical examination, and ancillary tests. Note: this is an information-gathering and documen- tation phase, not a phase for formulating inves- tigation plans. current_phase_description: Phase 2: Diagnosis & Plan phase_goal: Based on the established case features, orga- nize discussions among specialists to provide a preliminary diagnosis, differential diagnostic considerations, and a diagnostic and treatment plan. Specialist You are a specialist physician focusing on the field of spec_id. As a member of a multidisciplinary team (MDT), your core responsibility is to assist in refining the case information and to provide precise diagnostic and treatment recommendations from the professional perspective of spec_id. Tasks of the current phase: phase_instructions Medical record documentation guidance: soap_guidance Your workflow: 1. Analyze the input information and determine which SOAP field each piece of information should be written into. 2. Strictly output in the JSON format defined by `SpecialistOutput`, and in `draft_modifications` clearly specify the exact SOAP field to which each item of information belongs, for example: - “Write ‘chest pain for 2 hours’ into `history_of_present_illness`” - “Write ‘blood pressure 150/90 mmHg’ into `physical_examination`” 3. next_question_instruction historytakingdiagnostic_synthesis phase_instructions: From the professional perspective of spec_id, carefully review the current case features. Focus on clinical information that is directly relevant to your specialty, including basic patient infor- mation, history of present illness, past medical history, physical examination findings, and ancillary test results. If the existing information is not sufficient to fully and accurately document the patient’s condition, raise specific and concrete interview questions to supplement the missing details. If the available information is already adequate, clearly state that there are no additional ques- tions. At this stage, your role is strictly limited to recording and refining clinical information; do not propose diagnostic tests or treatment plans. next_question_instruction: Fill in the questions that need to be raised in 'suggested_questions' to advance the consulta- tion. If there are no questions, you should state the end. phase_instructions: Based on the finalized case features, provide your professional opinion from the perspective of spec_id. In the suggestions_for_draft_revision section, structure your output strictly as follows: - Preliminary diagnosis - Diagnostic discussion (including diagnosis basis and differential diagnosis) - Treatment plan, covering further investiga- tions, pharmacologic therapy, non-pharmaco- logic interventions, and follow-up require- ments next_question_instruction: Fill in 'N/A' in the 'suggested_questions' field and no longer ask questions to the patient. historytakingdiagnostic_synthesis soap_guidance: Basic Information (basic_information): - Patient demographics and chief complaint - Includes: age, sex, occupation, visit date, main symptoms - Keep it concise and focused on the main prob- lem History of Present Illness (history_of_pre- sent_illness): - Onset and course of the current illness - Includes: how it started, symptom characteris- tics, progression, prior evaluation and treat- ment, and response - Describe in chronological order with clear logic Past Medical History (past_medical_history): - Patient’s previous health history - Includes: past diseases, surgeries, trauma, allergies, personal and family history - Focus on information related to the current condition Physical Examination (physical_examination): - Objective findings from the physician’s exam- ination - Includes: vital signs and system-based exam findings - Describe objectively; avoid subjective judg- ments Auxiliary Examination (auxiliary_examina- tion): - Laboratory and imaging results - Includes: test values and examination dates - Report objective data accurately DIAGNOSIS & PLAN: Preliminary Diagnosis (preliminary_diagnosis): - List diagnoses line by line in standard clinical format - Include major and relevant secondary diagno- ses - Order by importance or timeline Diagnostic Discussion (diagnosis_discussion): - Explain diagnostic reasoning and differential diagnosis - Summarize key evidence: patient profile, symptoms, exam findings, and test results - Clearly mention: - Risk factors or health problems (if unknown, state “unclear”) - Complications or related conditions - Treatment adherence - Available family or social support - Do not invent tests or list unasked negative findings Treatment Plan (treatment_plan): 1. Further evaluation: - Planned tests, purpose, and timing, with guideline-based rationale - If surgery is indicated, include indication and approach 2. Medications: - Drug classes with indications, contraindica- tions, and monitoring points - Avoid specific dosages 3. Non-drug management: - Hydration goals, diet, lifestyle advice, pain management, and warning signs 4. Follow-up: - Follow-up timeframe, repeat tests, and visit format (in-person or remote) soap_guidance: Aggregator You are a recorder and decision-maker in a medical MDT. **Current phase: phase_name** As a professional physician, you must organize and document patient information strictly according to the standard SOAP format for the initial medical record: soap_guidance Your task: aggregator_task_description **Work requirements:** 1. **Information integration**: Carefully review the input from specialist physicians (if any). **If `specialist_outputs` is empty, extract information directly from the patient’s latest response and the coordinator’s suggestions, and update the SOAP record accordingly.** 2. **Empathy**: You are the doctor directly communicating with the patient; your tone should be warm, professional, and patient. 3. **Logical consistency**: Maintain a complete and coherent medical record. If the patient has clearly stated that certain information cannot be provided, document this truthfully and do not ask again. **Strictly output in the specified JSON format** historytakingdiagnostic_synthesis soap_guidance: (the same as specialist) aggregator_task_description: Summarize the question suggestions from specialist physicians and ask the patient accord- ingly. When the patient mentions key symptoms, make sure to confirm them. The patient is a layperson, so avoid overly technical or uncommon medical terms. soap_guidance: (the same as specialist) aggregator_task_description: Integrate all specialist opinions to generate preliminary diagnosis, diagnostic discussion, and treatment plan. C Evaluation Metrics We adopt a multi-dimensional evaluation frame- work that assesses both the consultation process and the resulting clinical documentation.All models are evaluated under an LLM-as-a-judge paradigm using gpt-4o-mini. Identical scoring prompts and rubric definitions are applied across all settings to ensure consistency and fair comparison. Documentation Quality. We evaluate the qual- ity of generated SOAP notes from complementary perspectives covering clinical reasoning, documen- tation standardization, readability, and reference similarity. Specifically, we use the following met- rics. •IDEA Score (Baker et al., 2015). IDEA eval- uates the completeness and coherence of clini- cal reasoning by examining alignment among history taking, physical examination, diagno- sis and differential, and care planning. Higher scores require detailed and well-organized HPI, complete and diagnostically relevant physical examinations, diagnoses supported by objective evidence with clear reasoning and ranked differentials, and comprehensive, appropriate care plans. Internal inconsisten- cies across sections are explicitly penalized. The detailed rubric is provided in Table 5. •SOAP Score (Health Human Resources De- velopment Center, National Health Com- mission of China, 2024). SOAP Score mea- sures adherence to standardized SOAP docu- mentation practices. The rubric evaluates com- pleteness and accuracy within each section. The Subjective component emphasizes struc- tured problem descriptions, concise chief com- plaints, detailed symptom characterization, prior evaluations and treatments, and relevant medical, family, and social histories. The Ob- jective component focuses on complete phys- ical examinations and accurate reporting of laboratory and ancillary tests. The Assess- ment component rewards clear diagnoses jus- tified by clinical evidence, analysis of risk factors and comorbidities, and evaluation of adherence and family resources. The Plan component emphasizes guideline-consistent diagnostic and management plans, detailed treatment strategies, non-pharmacologic inter- ventions, and explicit follow-up requirements. Predefined deduction codes are used to anno- tate common documentation errors for fine- grained analysis. The detailed rubric and de- duction codes are shown in Tables 6 and 7. •READ Score (Wang et al., 2024). READ Score assesses presentation quality and clini- cal usability, focusing on structural complete- ness, logical coherence, terminology accuracy, information redundancy, and salience of key findings. It reflects how easily a note can be read, understood, and safely used in practice. The detailed rubric is presented in Table 8. •chrF++. chrF++ measures surface-level sim- ilarity between generated notes and gold- standard documentation using charactern- gram overlap, providing a complementary lex- ical similarity signal. Consultation Capability.We operationalize con- sultation capability using the standardized patient (SP) consultation skills grading and scoring crite- ria from (Wang et al., 2025). The rubric adopts a five-tier scale (5 = best) and covers two domains: inquiry skills and humanistic care. For inquiry skills, we score (i) conversation arrangement, fo- cusing on whether the consultation has a clear open- ing, structured middle, and explicit closing with an orderly question flow; (i) question types, empha- sizing appropriate and balanced use of open-ended and closed-ended questions while avoiding sequen- tial leading questions; (i) verifications, assessing whether the clinician adequately verifies key infor- mation and cross-checks details through follow-up and reference; and (iv) professional jargon, reward- ing clear patient-friendly explanations with mini- mal unnecessary medical terminology. For human- istic care, we score (v) speech, evaluating whether tone and pace are comfortable and appropriate, and (vi) amiable behavior, assessing whether the clin- ician provides empathetic responses and comfort when appropriate. We score each item by matching dialogue behaviors to tier descriptors, then aggre- gate item scores into an overall consultation capa- bility score. Sec.ItemDefinitionMax01234 1. Present Illness & Comprehensive History 11.1 Detailed HPIComplaint: location, quality, severity, duration, onset, radiation, aggravating/relieving factors. 4Incorrect/illogical; key elements missing Some elements present Most elements present All elements presentConcise, organized, diagnostically salient 11.2 Prior Dx/Tx Course Prior evaluations/treatments: time, location, tests, meds, interventions; full if first visit. 4Incorrect/illogical description Some elements present Most elements present All elements presentConcise, structured, informative 11.3 Descriptive HPI Language Medical descriptors (e.g., acute/chronic, sharp/dull, constant/intermittent). 3Inappropriate/noneMinimal but appropriate Frequent, appropriate Consistent, accurate, concise – 11.4 Chronological Organization Temporal ordering and coherent illness narrative. 3Temporal contradictions Disorganized timeline Mostly coherent; minor gaps Clear and consistent chronology – 11.5 Contextualized HPI Integrates relevant PMH, family/social history, and associated symptoms. 4No/incorrect integration Partial integrationComprehensive integration Clear, accurate, concise Key info prioritized 11.6 Comprehensive History PMH, family history, social history, review of systems. 3Major errors/omissions Significant missing content Mostly completeThorough and complete – 11.7 CalibrationInternal consistency across the note. –Deduction: -2 points per internal inconsistency. 2. Physical Examination 22.1 Complete Physical Exam Comprehensive documentation of physical examination. 4Errors/not addressedMajor components missing Mostly complete; minor omissions Complete examWell-organized; professional terms 22.2 Key Physical Findings Highlights diagnostically relevant positive/negative findings. 3Missing/incorrect emphasis Partial emphasisComprehensive emphasis Prioritized by relevance – 3. Diagnosis & Differential 33.1 Diagnostic Completeness Primary, secondary, and additional diagnoses. 4Primary missing/incorrect Primary onlyPrimary + some secondary Primary + all secondary Includes additional diagnoses 33.2 Objective Evidence Evidence from history, exam, and investigations. 4Missing/incorrect evidence One domainTwo domainsAll relevant domains Also supports other dx 33.3 Diagnostic Reasoning Reasoning for primary diagnosis. 3None/incorrectBasic, partial explanation Rigorous explanation Links features to presentation – 33.4 Explanatory Summary Links diagnoses, risks, and complications. 3None/incorrectPartial analysisAll associations discussed Clear, logical synthesis – 33.5 Differentials≥3 relevant differentials ranked by likelihood. 3Irrelevant<3 or missing key alts All key alts includedOrdered by likelihood – 33.6 Differential Reasoning Inclusion/exclusion rationale; confounders. 3None/incorrectExclusion onlyAdequate exclusion only Inclusion + exclusion; confounders – 33.7 Overall Impression Professionalism, clarity, logical rigor. 4Poor professional- ism/logic Adequate professionalism Clear and consistent Strong professional quality Concise, polished, structured 4. Plan 44.1 Plan Completeness Investigations, treatment, lifestyle, follow-up. 4Missing/incorrectSingle aspectMulti-dimensionalDynamic assess- ment/prognosis Concise, rigorous 44.2 Plan Appropriateness Evidence/reasoning supports key decisions. 3 Inappropriate/incorrectVague/unsupportedGenerally appropriate Clear; strong evidence – 5. Overall Competency 55.1 Presentation Skill Quality of written presentation. 3–Basic: partialGood: mostExcellent: nearly all – 55.2 Reasoning SkillQuality of diagnostic reasoning. 3–Basic reasoningRelevant comparison Comprehensive, rigorous – 55.3 Decision SkillQuality of decisions in the plan. 3–List actions onlyPartial reasoningEvidence-based; patient-centered – Table 5: IDEA scoring rubric for clinical note evaluation. ItemMaxRubric S: Subjective S-1. Format5Each 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 complaint2Concise 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 5Detailed 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 3Prior 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 history3 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. S-2.5 Family history2 Clear documentation of heritable diseases in family members (2). If the key hereditary history is missing, deduct 1. If absent, score 0. S-2.6 Lifestyle, psychological, and social factors 5Comprehensive description of diet, sleep, exercise, smoking/alcohol use, mental status, work stress, family relationships, and financial situation (5). If 1–2 key elements are missing, deduct 2–3. If only briefly listed, score 0–2. O: Objective O-1. Physical examination8Vital signs and system examinations are accurately and completely documented; abnormal findings are described in detail (8). If 1–2 items are missing or inaccurate, deduct 2–4. If largely missing or incorrect, score 0–3. O-2. Laboratory and ancillary tests 5Test items, timing, and results (values or abnormal flags) are complete and accurate (5). If 1–2 results are missing or incorrectly transcribed, deduct 2–3. If absent or disorganized, score 0–2. O-3. Psychological tests/other assessments 2 If performed, psychological tests are documented with name and results (score/conclusion) (2). If incomplete, deduct 1. If not performed or not documented, score 0. A: Assessment A-1. Preliminary diagnoses4Primary diagnosis and comorbid/secondary diagnoses are clear and complete (4). Primary diagnosis correct (2). Some secondary diagnoses are missing (1). Secondary diagnoses complete (1). A-2.1 Diagnostic evidence4Diagnoses are justified using symptoms, signs, and test results with standard terminology (4). If evidence is insufficient or terminology is non-standard, deduct 1–2. If the diagnosis is incorrect or unsupported, score 0–1. A-2.2 Risk factors and health problems 10Disease-related risk factors and other potential health problems are comprehensively identified and their relationships analyzed (10). If 1–2 items are missing or the analysis is weak, deduct 3–5. If only listed without analysis, score 0–4. A-2.3 Complications and comorbidities 4Existing or potential complications and comorbidities are accurately identified and interactions analyzed (4). If important conditions are missed, deduct 2. If not analyzed, score 0–2. A-2.4 Adherence/compliance2 Treatment adherence is assessed based on clinical course with reasonable analysis (2). If brief, deduct 1. If incorrect or absent, score 0. A-2.5 Family resources1Available family support resources (human, financial, informational) are clearly described (1). If vague, deduct 0.5. If absent, score 0. P: Plan P-1. Further diagnostic and management plan 6Guideline-consistent plans specify required tests, follow-up timing, and necessary consultations (6). If 1–2 key elements are missing or timing is unclear, deduct 2–3. If disorganized or generic, score 0–3. P-2.1 Treatment plan (medications/surgery) 10Medication or surgical plans match diagnoses, with complete details and cited guideline sources and evidence levels (10). If key information is missing, deduct 3–5. If unreasonable or largely missing, score 0–4. P-2.2 Non-pharmacologic treatment 15Behavioral, dietary, and exercise interventions are specific and feasible, with precautions and cited evidence (15). If overly general, deduct 5–8. If empty or vague, score 0–6. P-3. Follow-up requirements4Follow-up timing and content (re-evaluation items and assessment focus) are clearly specified (4). If either timing or content is missing, deduct 2. If absent, score 0. Table 6: SOAP scoring rubric for clinical note evaluation. CodeMeaningCodeMeaning A1Misuse of terminologyA2Vague expression B1Missing important positive findingsB2Redundant minor positive findings C1Negative stated as positiveC2Positive stated as negative C3Missing important negative findingsD1Irrelevant information E1Missing time informationE2Vague time information F1Incorrect order/sequenceF2Incorrect time value G1Incomplete citation of external recordsG2Incorrect paraphrase of external records G3Non-standard citation formatHCompound error ILogical inconsistency/disorderJRedundant/verbose expression Table 7: Deduction codes used for error annotation. Item12345 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 coherenceNo 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 sufficiencyKey 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. Table 8: Readability rubric for clinical note evaluation. ItemTierRubric (English) Inquiry skills Conversation arrangement5The beginning, middle, and end of the consultation are clear and precise, with questions asked in an orderly manner. 4Between 5-point and 3-point. 3Most of the consultation is conducted in an orderly fashion, but the beginning and ending are not clearly defined. 2Between 3-point and 1-point. 1The consultation lacks coherence and organization. Question types5Reasonable use of open-ended or closed-ended questions. 4Between 5-point and 3-point. 3No open-ended questions, directly asking with closed-ended questions. 2Between 3-point and 1-point. 1Frequently uses sequential and leading questions. Verifications5Conduct a comprehensive and thorough verification and reference. 4Between 5-point and 3-point. 3The verification and reference are incomplete and not sufficient. 2Between 3-point and 1-point. 1Did not conduct verification and reference. Professional jargon5The explanation is clear and easy to understand, not using complicated medical terminology. 4Between 5-point and 3-point. 3The explanation is understandable, with minimal use of complex medical terminology. 2Between 3-point and 1-point. 1Frequently uses complicate medical terminology. Humanistic care Speech5Appropriate speech speed and tone. 4Between 5-point and 3-point. 3The speech speed and tone are mildly uncomfortable. 2Between 3-point and 1-point. 1The speech speed and tone are noticeably uncomfortable. Amiable behavior5Appropriate response and comfort. 4Between 5-point and 3-point. 3Provides responses and comfort. 2Between 3-point and 1-point. 1No response or comfort. Table 9: Standardized patient consultation skills grading and scoring criteria. D Analysis D.1 Department-wise Performance Analysis Figure 6 presents a fine-grained, department-wise comparison of documentation quality across 24 clinical specialties on ClinicalBench, evaluated us- ing IDEA and SOAP metrics. This analysis reveals that Aegle’s performance gains are not confined to a small subset of domains, but instead generalize consistently across all departments. On the IDEA metric, which emphasizes evidence-grounded clinical reasoning and diag- nostic coherence, Aegle outperforms MDAgents and MedAgents in the vast majority of depart- ments. The advantage is particularly pronounced in cognitively complex or high-ambiguity settings such as gastroenterology, neurology, cardiology, endocrinology, and hepatobiliary surgery. These departments typically involve heterogeneous symp- tom presentations and overlapping differential diag- noses, where single-perspective reasoning is espe- cially vulnerable to anchoring bias. The consistent IDEA improvements suggest that Aegle’s decou- pled parallel specialist reasoning and evidence-first state design effectively enhance hypothesis cover- age and diagnostic traceability under such complex- ity. In surgical departments (e.g., thoracic surgery, vascular surgery, neurosurgery, and gastrointesti- nal surgery), Aegle also demonstrates stable gains, despite these domains being traditionally more procedure-driven and less conversational. This indi- cates that the framework does not merely improve dialogue fluency, but meaningfully strengthens the structured capture of perioperative history, risk fac- tors, and decision rationales. Notably, the confi- dence intervals of Aegle are generally narrower than those of baseline multi-agent systems, sug- gesting reduced variance and more stable behavior across cases within the same department. The SOAP results exhibit a similar but slightly more conservative trend. While baseline multi- agent systems already perform competitively in highly standardized departments (e.g., pediatrics, obstetrics, and hematology), Aegle still achieves ei- ther the highest or statistically comparable scores in most cases. The gains are especially evident in de- partments where documentation structure is more heterogeneous, such as otolaryngology, urology, and respiratory medicine. This pattern indicates that Aegle’s explicit separation between case fea- tures and diagnostic outputs contributes to more consistent adherence to SOAP conventions when documentation norms are less rigid. Across both metrics, there is no department in which Aegle exhibits systematic degradation rela- tive to other multi-agent baselines. Instead, the im- provements scale with clinical complexity: depart- ments with broader diagnostic spaces and higher information entropy tend to benefit more from vir- tualized MDT-style collaboration. This observation aligns with the design motivation of Aegle, namely to mitigate single-view cognitive bias by distribut- ing reasoning across independent specialists while maintaining a shared, structured clinical state. Overall, the department-wise analysis substanti- ates that Aegle’s advantages are robust, generaliz- able, and clinically meaningful, rather than being driven by a small number of favorable scenarios. It further supports the claim that structured state- aware multi-agent collaboration is particularly ef- fective for complex, multi-system clinical intake tasks. D.2 Case Study: High-Risk Prostate Cancer To qualitatively demonstrate the advantages of Ae- gle’s virtual MDT framework, we analyze a com- plex real-world case from the RAPID-IPN dataset involving a 73-year-old male presenting with pro- gressive lower urinary tract symptoms (LUTS) and a PSA level> 155ng/mL (Fig 7 and Fig. 8). This case requires integrating urological history, oncol- ogy pathology, and imaging evidence to formulate a high-risk management plan. Precision in Evidence Acquisition. As illus- trated in Table 10, the primary challenge in this case was not the diagnosis of prostate cancer, which had already been confirmed by biopsy, but the ac- curate characterization of risk stratification and the severity of urinary obstruction. Reasoning-strategy baselines (CoT and ToT) captured high-level symp- toms but failed to record granular metrics required for surgical planning. Specifically, both CoT and ToT omitted the quantitative International Prostate Symptom Score (IPSS) and the exact blood pres- sure measurement documented during the physical examination. In contrast, Aegle’s dynamic topol- ogy activated a specialized Urologist Agent during the inquiry phase (Stage I). Whereas MDAgents failed to capture the critical urinary retention met- ric that determines the urgency of decompression, Aegle successfully incorporated this information into the Objective section of the IPN. 556065707580 breast surgical department endocrinology department orthopedics department cardiac surgical department hernia surgical department vascular surgical department thyroid surgical department gastroenterology department urinary surgical department thoracic surgical department neurology department gynecology department otolaryngology head and neck surgical department neurosurgery department pediatrics department hematology department obstetrics department nephrology department respiratory medicine department cardiovascular medicine department gastrointestinal surgical department hepatobiliary and pancreas surgical department anus and intestine surgical department IDEA 5055606570 SOAP AegleMDAgentsMedAgents 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 represen- tative medical multi-agent baselines (MDAgents and MedAgents). Higher scores indicate better clinical reasoning quality (IDEA) and documentation standardization (SOAP). Handling Diagnostic Ambiguity. The pathol- ogy report described multifocal adenocarcinoma with heterogeneous Gleason scores across biopsy cores. Standard baselines generally summarized these findings as “Prostate Cancer” without further differentiation. MedAgents and Aegle were the only models that retained core-level involvement percentages ranging from 20 to 70%. Importantly, Aegle extended beyond data retention in the As- sessment section by synthesizing these findings to correctly classify the patient as high risk accord- ing to EAU guidelines. This clinically meaningful distinction was not captured by the generic summa- rization produced by CoT. Plan Coherence.The improved evidence ground- ing directly translated into a higher-quality treat- ment plan. Because Aegle explicitly encoded the markedly elevated PSA level (> 155) and the Glea- son score of 8 within the clinical stateF, the Oncol- ogist Agent in Stage I generated a comprehensive staging strategy. This plan included a whole-body bone scan and pelvic MRI to exclude metastatic disease prior to scheduling radical prostatectomy. This case demonstrates how Aegle’s decoupled rea- soning framework preserves low-salience yet high- impact clinical details, ensuring that the final IPN satisfies the standards required for specialist refer- ral. Case in RAPID-IPN (Urology + Oncology) ## Case Features 1. **Patient profile:** Male, 73 years old, admitted due to “difficulty in urination for half a year and a confirmed diagnosis of prostate cancer for 2 months.” 2. **History of present illness:** The patient began experiencing progressive difficulty in urination half a year ago, accompanied by urinary frequency and nocturia 3–4 times per night. Two months ago, he was hospitalized at X Municipal Central Hospital for **acute urinary retention**. Laboratory testing revealed a PSA level greater than 155 μg/L. A prostate biopsy was performed, and pathology confirmed **prostate cancer**. Recently, the patient developed hematuria characterized by light red urine, beginning 5–6 days ago, occurring occasionally 1–2 times per day, without blood clots. He has also experienced fatigue for several months, with no significant impact on daily activities. He is now admitted for further evaluation and treatment with a provisional diagnosis of “prostate cancer.” 3. **Past medical history:** 3.1 **Chronic diseases:** History of gout for over 30 years and hypertension for more than 20 years, currently well controlled with oral medications. 3.2 **Surgical, trauma, and transfusion history:** History of open appendectomy more than 30 years ago. Denies other surgeries or blood transfusions. 3.3 **Marital, reproductive, and family history:** Married at age 26, with one son and one daughter. Family members are generally healthy. 3.4 **Smoking, alcohol, and substance use history:** Denies smoking, alcohol consumption, and use of addictive substances. 3.5 **Vaccination history:** No vaccinations received within the past year. 3.6 **Allergy history:** Denies any known drug or food allergies. 3.7 **Personal and occupational history:** Long-term resident of X City, Zhejiang Province. Retired company employee. Denies exposure to toxic chemicals or radiation. Denies residence in epidemic areas. 4. **Physical examination:** Conscious and alert, in fair general condition. Pain score: 0. Respiratory rate: 18 breaths/min; Oral temperature: 36.7°C; Pulse: 69 beats/min; Blood pressure: 131/71 mmHg. No cyanosis of the lips. No palpable superficial lymphadenopathy. Cardiac and pulmonary auscultation revealed no significant abnormalities. Abdomen soft, without tenderness or rebound tenderness; no palpable abdominal masses. No costovertebral angle tenderness bilaterally. No tenderness along the ureters. No palpable bladder distension above the pubic symphysis. **Digital rectal examination:** Enlarged prostate with a shallow central sulcus; no distinct nodules palpated. No blood noted on the examining glove. 5. **Auxiliary examinations:** - **Haining Central Hospital (2017-11-11):** Urinary system ultrasound showed mild bilateral hydronephrosis, prostatic hyperplasia, and post-void residual urine volume of approximately 300 mL. - **Tumor markers (2017-11-12):** Total PSA >155.00 μg/L; free PSA 8.87 μg/L. - **Prostate biopsy pathology (2017-11-24):** 1. *Left inner prostate:* 3 cores obtained; prostate cancer identified in 2 cores; Gleason score 4+3=7; tumor involvement 30%. 2. *Left outer prostate:* 3 cores obtained; prostate cancer in all 3 cores; Gleason score 4+4=8; tumor involvement 50%. 3. *Right inner prostate:* 3 cores obtained; prostate cancer in all 3 cores; Gleason score 4+3=7; tumor involvement 20%. 4. *Right outer prostate:* 3 cores obtained; prostate cancer in all 3 cores; Gleason score 5+3=8; tumor involvement 60%. 5. *Suspicious area of left outer prostate:* 3 cores obtained; prostate cancer in 2 cores; Gleason score 5+3=8; tumor involvement 70%. ------ ## Preliminary Diagnosis 1. Prostate cancer 2. Hypertension 3. Gout 4. Status post appendectomy ------ ## Diagnostic Discussion ### Diagnostic Basis The patient is a 72-year-old male admitted due to “difficulty in urination for half a year and a confirmed diagnosis of prostate cancer for 2 months.” He developed progressive urinary obstruction with urinary frequency and nocturia half a year ago, consistent with symptoms associated with prostate cancer. Two months prior, he presented with acute urinary retention, and external hospital testing revealed a markedly elevated PSA level (>155 μg/L). Prostate biopsy confirmed prostate cancer with Gleason scores ranging from 7 to 8 and tumor involvement of 20%–70%. Recently, hematuria has developed. Digital rectal examination revealed an enlarged prostate with a shallow central sulcus. Urinary ultrasound demonstrated mild bilateral hydronephrosis and a residual urine volume of 300 mL. These findings strongly support the diagnosis of prostate cancer. Figure 7: The IPN of a 73-year-old male patient suffering from prostate cancer. Case in RAPID-IPN (Urology + Oncology) Regarding risk factors and health issues, advanced age (72 years) is a major risk factor for prostate cancer. The patient also has comorbid hypertension and gout. In terms of complications, mild bilateral hydronephrosis suggests the presence of obstructive uropathy. The patient demonstrates good treatment compliance, seeks medical care promptly, and is receptive to therapeutic plans. In terms of family and social resources, he is a retired employee with stable financial support and medical insurance, ensuring good access to medical care. ------ ### Differential Diagnosis 1. **Benign prostatic hyperplasia (BPH):** The patient presents with urinary obstruction, frequency, and nocturia, and digital rectal examination shows prostate enlargement, which may resemble BPH. However, the markedly elevated PSA and biopsy-confirmed malignancy exclude isolated BPH. 2. **Urinary tract infection:** May cause urinary symptoms and hematuria; however, the patient has no fever or dysuria, and the significantly elevated PSA favors malignancy. 3. **Bladder tumor:** Can present with hematuria and urinary obstruction, but prostate biopsy has already confirmed prostate cancer, and there is no current evidence of bladder mass lesions. 4. **Neurogenic bladder:** Can result in urinary retention and increased residual urine volume, but the patient has no history of neurologic disease and normal anal sphincter tone. 5. **Urethral stricture:** May cause urinary obstruction, but in an elderly male with markedly elevated PSA, prostate cancer is more consistent with the clinical picture. ------ ## Treatment Plan ### Further diagnostic evaluation According to EAU guidelines, the patient has high-risk prostate cancer (Gleason score 7–8, PSA >155 μg/L). Comprehensive staging examinations are required, including: - Pelvic MRI to assess local tumor invasion and pelvic lymph node involvement - Whole-body bone scan to evaluate for bone metastases - Chest CT to assess for pulmonary metastases - Abdominal CT to evaluate for liver metastases and lymph node involvement The patient meets surgical indications. **Radical prostatectomy** is considered for localized or locally advanced prostate cancer, including removal of the prostate and seminal vesicles with pelvic lymph node dissection. ### Pharmacologic treatment Androgen deprivation therapy (ADT) may be considered as neoadjuvant or adjuvant therapy for high-risk prostate cancer to reduce tumor volume and surgical difficulty. Liver function, lipid profile, and bone mineral density should be monitored. Optional agents include gonadotropin-releasing hormone (GnRH) agonists or antagonists. ### Non-pharmacologic interventions and health education Daily fluid intake is recommended at 2,000–2,500 mL, adjusted based on renal function. Given the patient’s gout and hypertension, a low-fat diet is advised, with restricted red meat intake and controlled purine consumption. Moderate physical activity is encouraged to maintain a healthy body weight. The patient should receive education on pain self-management and be instructed to seek immediate medical attention if severe bone pain, complete urinary obstruction, or fever occurs. ### Follow-up The first postoperative follow-up is recommended 1–2 weeks after surgery, followed by regular follow-up every 3–6 months. Follow-up evaluations should include PSA levels, liver and renal function tests, and complete blood counts. Given the patient’s good compliance and adequate family support, in-person follow-up is preferred, with telemedicine consultations as needed. Figure 8: The IPN of a 73-year-old male patient suffering from prostate cancer (continued). Key Clinical EvidenceAegleCoTToTMDAgentsMedAgents Age: 73-year-old male✓ Progressive LUTS (weak stream, hesitancy, intermittency, straining, dribbling)✓ Nocturia 3–4 times/night✓ Intermittent painless light-red hematuria (1–2×/day, no clots)✓ Acute urinary retention 2 months ago (catheterized)✓ PSA >155 ng/mL✓ Prostate biopsy: adenocarcinoma, Gleason 7–8, multifocal, 20–70% involvement✓ △ △ △✓ Urinary ultrasound: bilateral mild hydronephrosis✓ Post-void residual≈300 mL✓ ×✓ No CT / MRI / bone scan performed yet✓ △✓ Renal function & baseline labs not yet available✓ △✓ Blood pressure from original exam✓ ×✓△ IPSS score mentioned✓ × × × Table 10: Coverage of key clinical evidence across different reasoning frameworks.✓= explicitly documented;△ = partially mentioned or ambiguous;× = missing;