Paper deep dive
Evidence, Logic, and Compliance: Multi-Agent Structured Graph Reasoning with Expert Arbitration for Medical Referral
Qi Peng, Yi Cai, Jialin Cui, Tong Zhu, Yujuan Ding, Qingbao Huang, Tao Wang, Jiayuan Xie, Changmeng Zheng, Qing Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 9/1/2026, 4:28:23 AM
Summary
The paper introduces MASGR (Multi-Agent Structured Graph Reasoning), a framework designed to improve medical referral decisions by addressing information overload and unstructured collaboration in existing LLM-based systems. MASGR treats referral as a structured graph construction problem, using specialized agents to extract evidence from multimodal data (patient narratives, lab, radiology, pathology) and coordinate them via a clinical reasoning graph. It employs a knowledge-guided arbitration mechanism to prioritize patient safety rules over standard diagnostic classification, significantly outperforming state-of-the-art LLMs and multi-agent systems on real-world medical records.
Entities (10)
Relation Signals (7)
MASGR β addresses β Unstructured Collaboration
confidence 95% Β· MASGR... coordinates them through a clinical reasoning graph... replaces unstructured dialogue with a collaborative cross-domain linking mechanism.
MASGR β addresses β Information Overload
confidence 95% Β· MASGR deploys specialized agents to extract evidence from distinct modalities... This graph forces agents to establish explicit logical connections between conflicting evidence.
MASGR β uses β Clinical Reasoning Graph
confidence 95% Β· MASGR deploys specialized agents... and coordinates them through a clinical reasoning graph.
MASGR β prioritizes β Patient Safety Rules
confidence 92% Β· integrates a knowledge-guided arbitration mechanism that prioritizes patient safety rules over standard diagnostic classification.
MASGR β outperforms β Large Language Models
confidence 90% Β· Extensive experiments on real-world medical records demonstrate that MASGR significantly outperforms state-of-the-art LLMs
MASGR β uses β Referral Knowledge Base
confidence 90% Β· we integrate a knowledge-guided arbitration mechanism... By referencing a curated knowledge base... we construct a referral knowledge base (RKB)
Referral Knowledge Base β constructedusing β MedRAG
confidence 85% Β· we integrate MedRAG... as an external reference source... To construct this knowledge base...
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Medical referral (directing patients to the appropriate hospital department) is a complex decision-making process requiring the synthesis of multimodal data, including patient narratives, laboratory indicators, and radiology imaging. While Large Language Models (LLMs) have advanced medical dialogue systems, they struggle with real-world referral tasks due to two primary limitations: (1) Information Overload, where models fixate on high-frequency disease terms while overlooking subtle but critical urgency indicators; and (2) Unstructured Collaboration, where existing multi-agent frameworks rely on loose dialogue that leads to semantic drift and confirmation bias. To address these challenges, we introduce MASGR (Multi-Agent Structured Graph Reasoning), a framework that treats referral not as a classification task but as a structured graph construction problem. MASGR deploys specialized agents to extract evidence from distinct modalities and coordinates them through a clinical reasoning graph. This graph forces agents to establish explicit logical connections between conflicting evidence. Furthermore, we integrate a knowledge-guided arbitration mechanism that prioritizes patient safety rules over standard diagnostic classification. Extensive experiments on real-world medical records demonstrate that MASGR significantly outperforms state-of-the-art LLMs and existing multi-agent systems, particularly in complex cases requiring the balancing of chronic disease management and emergency intervention. The AI contribution lies in the Multi-Agent Structured Graph Reasoning framework that transforms unstructured multi-agent dialogue into a verifiable logical graph construction. The engineering application is demonstrated through its deployment in a complex healthcare decision-making system to optimize the precision of complex medical referrals.
Tags
Links
- Source: https://arxiv.org/abs/2608.30938v1
- Canonical: https://arxiv.org/abs/2608.30938v1
Trouble viewing inline? Open PDF directly β
Full Text
71,825 characters extracted from source content.
Expand or collapse full text
Evidence, Logic, and Compliance: Multi-Agent Structured Graph Reasoning with Expert Arbitration for Medical Referral Qi Peng a,b , Yi Cai a , Jialin Cui a , Tong Zhu c , Yujuan Ding b , Qingbao Huang d , Tao Wang e , Jiayuan Xie b,β , Changmeng Zheng b,β and Qing Li b a School of Software Engineering, South China University of Technology, Guangzhou, China b Department of Computing, Hong Kong Polytechnic University, Hong Kong, China c School of Computer Science and Network Engineering, Guangzhou University, Guangzhou China d School of Electrical Engineering, Guangxi University, Guangxi China e Department of Biostatistics & Health Informatics (BHI), Kingβs College London, London A R T I C L E I N F O Keywords: Medical Referral Multi-Agent Systems Graph Reasoning A B S T R A C T Medical referral (directing patients to the appropriate hospital department) is a complex decision-making process requiring the synthesis of multimodal data, including patient narra- tives, laboratory indicators, and radiology imaging. While Large Language Models (LLMs) have advanced medical dialogue systems, they struggle with real-world referral tasks due to two primary limitations: (1) Information Overload, where models fixate on high-frequency disease terms while overlooking subtle but critical urgency indicators; and (2) Unstructured Collaboration, where existing multi-agent frameworks rely on loose dialogue that leads to semantic drift and confirmation bias. To address these challenges, we introduce MASGR (Multi- Agent Structured Graph Reasoning), a framework that treats referral not as a classification task but as a structured graph construction problem. MASGR deploys specialized agents to extract evidence from distinct modalities and coordinates them through a clinical reasoning graph. This graph forces agents to establish explicit logical connections between conflicting evidence. Furthermore, we integrate a knowledge-guided arbitration mechanism that prioritizes patient safety rules over standard diagnostic classification. Extensive experiments on real-world medical records demonstrate that MASGR significantly outperforms state-of-the-art LLMs and existing multi-agent systems, particularly in complex cases requiring the balancing of chronic disease management and emergency intervention. The AI contribution lies in the Multi-Agent Structured Graph Reasoning framework that transforms unstructured multi-agent dialogue into a verifiable logical graph construction. The engineering application is demonstrated through its deployment in a complex healthcare decision-making system to optimize the precision of complex medical referrals. 1. Introduction Medical referral, the process of directing patients to the most appropriate hospital department, is a critical juncture in healthcare delivery (Anyanwu, Abedi Harrison and Onohwakpor Efe, 2015). An accurate referral not only optimizes hospital resource allocation but, more importantly, ensures timely and life-saving treatment for patients with complex conditions. In real-world hospitals, this decision-making process is inherently multimodal and cognitively demanding (Kline, Wang, Li, Dennis, Hutch, Xu, Wang, Cheng and Luo, 2022; Shaik, Tao, Li, Xie and VelΓ‘squez, 2024). It requires physicians to synthesize fragmented information from heterogeneous sources, ranging from subjective outpatient narratives to quantitative laboratory indicators, spatial radiology imaging, and microscopic pathology findings. Recent advances in Large Language Models (LLMs) have shown great impressive capabilities across a wide range of medical applications (Liu, McCoy and Wright, 2025; Toma, Lawler, Ba, Krishnan, Rubin and Wang, 2023; Yang, Wei, Li, Liu, Liu, Hu, He, Ju, Zhou, Liu et al., 2025; Kumari, Chauhan, Jain and Garg, 2025; Xu, Chen, Chai, Xiao, Yan, Ding, Wang, Jin, Jiao, Yue et al., 2026). However, applying general-purpose LLMs to complex medical referral tasks presents significant challenges. First, the information overload problem: when creating a prompt with all available patient data, single-agent LLMs often suffer from attention dispersion (Vishwanath, Alyakin, Alber, Lee, Kondziolka and Oermann, 2025; Fan, Sun, Xue, Zhang, Zhang and Ruan, 2025). As illustrated in Figure β Corresponding author jiayuan.xie@polyu.edu.hk (J. Xie); changmeng.zheng@polyu.edu.hk (C. Zheng) ORCID(s): : Preprint submitted to ElsevierPage 1 of 18 arXiv:2608.30938v1 [cs.MA] 31 Aug 2026 Information Data: - Gender: Female - Main complaint: Persistent chest pain with shortness of breath ... - Present Illness: ... The patientβs partial pressure of oxygen is 53 mmHg and oxygen saturation is 85.2%, meeting ICU emergency criteria (pressure of oxygen < 60 mmHg or oxygen saturation < 90%). Therefore, the most appropriate department is the Intensive Care Unit (ICU). Radiology Data: Laboratory Data: Pathology Data: - Oxygen partial pressure: 53 (mmHg) - (CEA): 21.8 ng/mL - ... - Arterial Blood Oxygen Saturation (SpOβ): 78% - ... - Microscopic findings: Tumor cells... distinct nucleoli, increased mitosis, and partial necrosis ... - X-ray Findings: ... A mass-like high-density shadow in the left, ... - CT Findings: ... density lesion located in the left lower lobe, showing lobulation and spicule signs. ... Clinical Reasoning Graph Constructed by Our Framework Determined by expert Main Complaint: Chest pain with shortness of breath Present Illness: Chest tightness and pain significantly aggravated in the past 3 days X-ray Findings: A mass-like high-density shadow in the left lower lobe with blurred margins and slightly elevated ... CT Findings: A 4.2 Γ 3.8 cm irregular left lower-lobe mass with lobulation, spiculation, focal necrosis, ... Blood Routine: SpOβ: 78% Oxygen partial pressure: 53 (mmHg) Tumor markers: Carcinoembryonic Antigen (CEA): 21.8 ng/mL Possibility: Hypoxemia Possibility: Lung Cancer Transfer Department: Department of ICU Transfer Department: Department of Oncology Pathological Diagnosis: Tumor cells form solid and gland-like nests ... (a) Single LLM: (c) Our Framework: (b) Traditional Multi-Agent Framework: - Information: Chest pain - Radiology: Mass; Lobulation - Pathology: Tumor cells Dominated Input Focus (Red Data) Conclusion: Highly suspeet lung cancer. Refer to Oncology Department. FLAW: Information Overload & Attention Bias to High-Frequency Terms Chest pain and shortness of breath. 40-year smoking history indicates high risk for lung cancer. CT shows lobulation, which strongly supports malignancy. Invasive adenocarcinoma. Oncology referral is needed. Tumor markers are elevated (CEA = 21.8 ng/mL). This also supports cancer progression. Agreed, the correct department is Oncology. FLAW: Sequential Unstructured Dialogue Leads to Semantic Drift and Herding Figure 1: A comparison of our framework with existing decision-making paradigms for a patient with lung cancer complicated by severe hypoxemia. 1(a), a model may fixate on salient, high-frequency terms such as βmass-like high-density shadowβ in radiology or βtumor cellsβ in pathology reports. Consequently, it may overlook subtle but critical urgency indicators (e.g., a drop in blood oxygen in lab results), which lead to technically correct disease identification but clinically disastrous referral decisions (e.g., sending a patient in respiratory failure to Oncology instead of the ICU). Second, the lack of structured reasoning: existing solutions, including recent medical multi-agent frameworks (Tang, Zou, Zhang, Li, Zhao, Zhang, Cohan and Gerstein, 2024; Kim, Park, Jeong, Chan, Xu, McDuff, Lee, Ghassemi, Breazeal, Park et al., 2024; Peng, Cai, Liu, Zou, Chen, Zhong, Wang, Xie and Li, 2025a; Yan, Liu, Wu, Chen, Wang, Chai, Wang, Zhao, Zhang, Zhang et al., 2024), typically rely on natural language dialogue for collaboration. This unstructured interaction resembles a disorganized discussion; information is lost in translation, and there is no explicit verification of logical consistency (Ma, Huang, Yang, Zhang, Wu, Shen, Fan, Zhang and Zang, 2025). As shown in Figure 1(b), agents may engage in βherdingβ behavior: seeing elevated tumor markers and CT lobulation, they mutually reinforce a diagnosis of invasive adenocarcinoma. This semantic drift allows the conversation to bypass the patientβs critical hypoxemia entirely. Without a structured βcognitive bridgeβ connecting evidence to conclusions, the reasoning process remains opaque and prone to hallucination. To address these limitations, we introduce the Multi-Agent Structured Graph Reasoning (MASGR) framework. Inspiring from the workflow of a multidisciplinary team (MDT) of specialists (Taberna, Moncayo, JanΓ©-Salas, Antonio, Arribas, Vilajosana, Torres and MesΓa, 2020), our framework models the referral process not as a classification task, but as a structured graph construction and reasoning problem. The MASGR framework operates on three pillars: β’ Domain-Specific Local Graph Construction: Instead of a single overwhelmed model, we deploy specialized agents to process distinct data modalities (i.e., Outpatient, Lab, Radiology, and Pathology). Each agent acts as a domain expert, extracting evidence to construct a structured local reasoning graph. This ensures that subtle cues, such as the specific texture of a lung nodule in radiology or a critical drop in blood gas values, are captured with high fidelity before global integration. : Preprint submitted to ElsevierPage 2 of 18 β’ Global Graph Co-Building via Structured Relations: We replace unstructured dialogue with a collaborative cross-domain linking mechanism. Agents communicate by establishing explicit logical relationships between clin- ical evidence nodes: cascade (causal chains), syndrome (mutual reinforcement), exclusion (conflict identification), and context (conditional validity). This structured approach forces the system to align conflicting data, such as identifying a βdiagnostic discordanceβ between a chronic tumor diagnosis and an acute respiratory crisis, into a coherent global reasoning graph, serving as a transparent investigation board. β’ Expert Knowledge Arbitration: A correct diagnosis does not always equal a correct referral. To bridge the gap between medical facts and hospital operations, we introduce an expert knowledge arbitration mechanism. By referencing a curated knowledge base, an expert agent evaluates candidate reasoning graphs against three dimensions: evidence completeness, reasoning coherence, and knowledge compliance. For instance, as shown in Figure 1(c), the system applies the strict compliance rule that βpressure of oxygen < 60 mmHgβ constitutes a medical emergency, correctly overriding the cancer diagnosis to prioritize the Intensive Care Unit (ICU). We constructed a dataset based on real-world inpatient records from a tertiary hospital to evaluate our approach. Extensive experiments demonstrate that MASGR significantly outperforms state-of-the-art LLMs and existing medical multi-agent systems. Furthermore, ablation studies confirm that explicitly modeling logical relationships and integrating expert arbitration are crucial for handling complex cases. 2. Related Work 2.1. Medical Large Language Models Large Language Models (LLMs) have been widely adopted in domains such as software development (Hou, Zhao, Liu, Yang, Wang, Li, Luo, Lo, Grundy and Wang, 2024; Fan, Gokkaya, Harman, Lyubarskiy, Sengupta, Yoo and Zhang, 2023), fundamental sciences (Zhang, Wei, He and Yu, 2024), and social simulation (Park, OβBrien, Cai, Morris, Liang and Bernstein, 2023). More recently, they have shown great promise in the medical domains, with applications such as medical question answering (Singhal, Tu, Gottweis, Sayres, Wulczyn, Amin, Hou, Clark, Pfohl, Cole-Lewis et al., 2025; Lucas, Yang, Pomeroy and Yang, 2024) and report generation (Liu, Tian, Chen, Song and Zhang, 2024b). Medical LLMs typically follow two main development strategies: (1) training specialized models on domain-specific medical data (Wang, Zhao, Ouyang, Wang and Shen, 2023), and (2) adapting general-purpose LLMs through prompt engineering and retrieval-augmented generation (RAG) techniques (Wen, Wang and Sun, 2024). Initial efforts in medical LLMs centered on pre-training and fine-tuning models using domain-specific medical data. With the advent of large-scale general-purpose LLMs, however, training-free methods have gained attraction, which leverage both their inherent capabilities and external medical knowledge. For example, GPT-5 (Achiam, Adler, Agarwal, Ahmad, Akkaya, Aleman, Almeida, Altenschmidt, Altman, Anadkat et al., 2023), when guided by well-designed prompts, has demonstrated superior performance compared to specialized fine-tuned models such as Med-PaLM (Nori, King, McKinney, Carignan and Horvitz, 2023). Though LLMs have shown strong performance in the medical domain, they face limitations in complex scenarios involving multi-source medical data (Peng et al., 2025a). To address this, we propose a multi-agent framework where multiple LLMs collaboratively process diverse medical data, extract key information, and explicitly construct inter-data associations, improving the accuracy in complex medical scenarios. 2.2. Multi-Agent Frameworks Previous studies have shown that multi-agent collaboration can significantly improve performance in complex tasks, such as large-scale code development (Hong, Zheng, Chen, Cheng et al., 2023), strategic game playing (Zheng, Liang, Zhang, Wei, Chua and Li, 2024; Liang, Wei and Zheng, 2026), and multi-robot coordination (Heuer, Palmieri, Mannucci, Koenig and Magnusson, 2024). In recent years, this approach has been extended to the medical domain. For example, Tang et al. (Tang et al., 2024) propose a Multidisciplinary Collaboration (MC) framework, where multiple agents engage in collaborative debate to reach medical diagnoses. Kim et al. (Kim et al., 2024) introduce MDAgents, a system that decomposes medical tasks into different complexity levels and assigns agent teams accordingly. These studies demonstrate the effectiveness of designing role-specific doctor agents to collaboratively perform medical decision-making. However, existing multi-agent approaches often employ unstructured natural language interactions, causing topic drift or herd effects, which may lead to suboptimal referral decisions. To address this issue, we propose a multi-agent reasoning framework. In this framework, multiple agents collaboratively extract key information from different types of medical data and finally integrate them into a reasoning graph dynamically. This reasoning : Preprint submitted to ElsevierPage 3 of 18 Table 1 Comparison with existing datasets. Our dataset covers all types of medical data. DatasetPatient Info Radiology Laboratory PathologyAvg. Words Departments PMC-VQA (2023)%!%!9.962 Slake (2021)%!%%4.471 Path-VQA (2021)%%%!6.231 VQA-RAD (2018)%!%%6.801 MedQA (2021)!%!%116.18- Open-XDDx (2025)!%!%113.739 MedThink-Bench (2025)!%!!151.1610 Ours!!!!1956.0137 graph models the causal relationships among symptoms, examinations, and diagnoses in clinical settings, improving inference accuracy and enhancing the interpretability of the final decision. 3. Data Construction 3.1. Dataset 3.1.1. Data Collection Since no publicly available dataset exists for evaluating large models in medical referral scenario, we collected a real-world dataset from a major general hospital. We focused on inpatients during a selected time window and accessed the hospitalβs electronic medical record (EMR) system to retrieve all relevant clinical information from admission to department transfer. The collected data include patient information, laboratory test results, imaging reports, pathology reports, and official referral records. All files are indexed and organized using each patientβs unique hospitalization ID to ensure consistency and facilitate downstream modeling and analysis. 3.1.2. Data processing After obtaining the raw data, we first structured the information of all patients. Specifically, we saved each patientβs medical data type as the key and the corresponding result as the value in a JSON format file. We then formulated the medical referral task as a multi-class classification problem. To enable objective evaluation, we generated four candidate department options for each case. These included the ground truth departments from the transfer record and other randomly selected distractors from the hospitalβs remaining departments. The final structured JSON file for each patient contains demographic information, categorized medical test results, candidate department options, and the corresponding labels. 3.1.3. Data Statistic The final dataset consists of 952 patients, covering 37 clinical departments and 4,823 distinct disease types. The top five most common diagnoses are hepatic cyst (2.8%), hypertension (2.3%), hypoproteinemia (1.7%), type 2 diabetes mellitus (1.7%), and chronic gastritis (1.6%). In terms of gender distribution, female patients account for 46.1%, while male patients account for 53.9%. The three most frequently visited departments are oncology (15.76%), gastrointestinal surgery (13.55%), and thyroid surgery (8.82%). The overall distribution of departments is shown in Figure 2. Table 1 compares our proposed dataset with existing benchmarks. We categorize medical data into four distinct types: Patient Info (e.g., chief complaint, present illness), Radiology (e.g., X-ray, CT reports), Laboratory (e.g., blood tests, urinalysis), and Pathology (e.g., biopsy findings). Additionally, the column βAvg. Wordsβ indicates the average length of the input text for each dataset, and βDepartmentsβ means the number of departments. As observed in the table, most existing datasets focus on limited data modalities. In contrast, our dataset achieves comprehensive coverage of all four data types. Moreover, each case in our dataset consistently contains all four modalities, whereas cases in existing datasets typically include only one or two types of examinations. This design leads to substantially more complex clinical scenarios, which require more advanced reasoning and decision-making capabilities to integrate heterogeneous medical evidence. 3.2. Referral Knowledge Base To equip the expert agent with reliable domain knowledge and to ensure clinically grounded arbitration, we construct a referral knowledge base (RKB) that provides a minimally complete and interpretable framework for : Preprint submitted to ElsevierPage 4 of 18 Figure 2: The distribution of departments in dataset. department-level referral reasoning. For each department, the RKB incorporates eight categories of structured knowledge: Department Overview and Major Conditions and Diagnoses delineate the clinical scope and diagnostic boundaries, enabling precise semantic alignment between patient-specific evidence and departmental expertise. Recommended Patient Transfer Criteria and Cases Not Suitable for This Department offer positive and negative referral constraints, ensuring logical consistency and preventing misclassification. Urgency Guidance introduces emergency transfer indicators essential for identifying time-critical referral scenarios, while Interdepartmental Boundaries and Collaborations captures cross-specialty interactions essential for handling ambiguous or overlapping presentations. Finally, Key Diagnostic Clues or Tests and Typical Case Examples provide evidence-based anchors that align model-generated reasoning chains with established clinical patterns. To construct this knowledge base, we employ GPT-5 with carefully designed prompt templates tailored to each knowledge field. To minimize hallucinations and enforce factual grounding, we integrate MedRAG (Xiong, Jin, Lu and Zhang, 2024) as an external reference source. MedRAG retrieves authoritative clinical materials from resources such as PubMed 1 , StatPearls 2 and Wikipedia 3 , ensuring that the generated content remains objective, verifiable, and aligned with medical guidelines. For example, a partial prompt used for constructing Department Overview field is: You are a medical knowledge extraction agent. Your task is to generate an accurate, concise, and guideline-aligned Department Overview for a given medical department. ...... You will be given: (1) The target department name. (2) A set of reference passages retrieved from MedRAG ...... Return the Department Overview in the following format: ...... To rigorously validate the quality of the constructed RKB, we conducted a human evaluation involving three board- certified clinicians. We randomly sampled 10% of the department entries and established a standardized evaluation protocol based on a 5-point Likert scale (1 = poor, 5 = excellent). The experts assessed each entry across three dimensions: (1) Clinical Accuracy, measuring the factual correctness and absence of hallucinations in the generated medical content; (2) Information Completeness, evaluating whether the entry comprehensively covers all eight structured categories defined in our schema without omitting critical diagnostic criteria; and (3) Referral Utility, 1 https://pubmed.ncbi.nlm.nih.gov/ 2 https://w.statpearls.com/ 3 https://w.wikipedia.org/ : Preprint submitted to ElsevierPage 5 of 18 assessing the practical value and distinctiveness of the guidance for distinguishing inter-departmental boundaries. The evaluation yielded average scores of 4.8 for accuracy, 4.7 for completeness, and 4.7 for utility, demonstrating the high reliability of our construction pipeline. Furthermore, the Fleissβ Kappa score (Falotico and Quatto, 2015) was calculated to be 0.76, indicating substantial inter-annotator agreement and confirming the robustness of the knowledge base as a trustworthy clinical reference. 4. Problem Formulation 4.1. Task Input and Objective Task Input: Each case involves a multi-source patient record denoted as ν = ν 1 ,ν 2 ,...,ν ν , where each ν ν represents a specific information source, such as laboratory test results, imaging reports or prior medical history. These heterogeneous data sources collectively describe the patientβs current clinical status. The system operates over a predefined set of candidate medical departments ξ° = ν·ννν‘ 1 ,ν·ννν‘ 2 ,...,ν·ννν‘ ν , encompassing common hospital divisions such as Cardiology, Pulmonology, and Emergency Medicine. Task Objective: Given the patient record ν , the goal is to infer the most appropriate target department ν dept β ξ°, that can provide optimal diagnosis or treatment for the patientβs current condition. 4.2. Task Formulation: Reasoning Graph Construction To address this task, we re-formulate the medical referral task as a multi-stage, structured graph generation problem, defining the task as the following pipeline: Stage 1: Local Reasoning Graph Generation. This stage is executed by each domain agent ν΄ ν β ξ. The task for each agent ν΄ ν is to independently analyze its designated data ν ν and construct a preliminary reasoning graph ξ³ ννννν,ν . ξ³ ννννν,ν = ν΄ ν (ν ν ) Here, ξ³ ννννν,ν = (ξ ν ,ξ± ν ) comprises the evidence nodes ξ ν and their internal logical relationships ξ± ν . The collective output of this stage is a set of preliminary reasoning graphs ξ³ ννννν,1 ,ξ³ ννννν,2 ,...,ξ³ ννννν,ν . Stage 2: Global Reasoning Graph Co-Building. This stage is executed by the multi-agent system ξΉ through collaboration among all agents ξ. The task is to define a collaborative generation function Ξ¦ ννν that takes all preliminary graphs ξ³ ννννν,ν as input. Through agent collaboration, this function constructs a set of candidate global reasoning graphs,ν νννν . ν νννν = Ξ¦ ννν (ξ³ ννννν,1 ,...,ξ³ ννννν,ν ,ξΉ) In this formulation,ν νννν represents the set of all plausible global graphs ξ³ νννννν,ν . Each candidate ξ³ νννννν,ν = (ξ νννννν ,ξ± νννννν ) must be a logically coherent graph merging evidence from multiple sources. The key challenge is designing Ξ¦ ννν to effectively merge, link, and resolve the evidence and logic from multiple ξ³ ννννν,ν . Stage 3: Expert-Knowledge Arbitration. This final stage is executed by the Expert Agent ν΄ νν₯νννν‘ augmented with an Expert Knowledge Base ξ· νν₯νννν‘ . The objective is to evaluate the logical validity of the candidate graphs in ν νννν and identify the most reasonable reasoning chain, denoted as ξ³ ννν‘νννν . ξ³ ννν‘νννν = Ξ¦ ννν (ν νννν ,ξ· νν₯νννν‘ ) Here, Ξ¦ ννν represents the arbitration process where the expert agent filters candidates based on domain knowledge. Finally, the specific department recommendation ν νννν‘ is directly derived from the diagnostic path within this optimal reasoning chain. ν νννν‘ = Ξ¨ νννννν (ξ³ ννν‘νννν ) where Ξ¨ νννννν maps the confirmed diagnosis path to the target department ν νννν‘ β ξ°. : Preprint submitted to ElsevierPage 6 of 18 Expert Doctor Prompt: Answer: Department of Neurosurgery. : Outpatient Doctor : Laboratory Doctor : Radiology Doctor : Pathology Doctor Referral Knowledge: Graph Generation by Multi-Agent graph TD P1[Main complaint: recurrent headache, blurred vision] -->|Cascade| S1[Neurological symptoms] P2[Current medical history: progressively worsening symptoms, accompanied by nausea and vomiting] -- >|Cascade| S2[Increased intracranial pressure] C_CT[Head CT: right temporal lobe mixed density mass, surrounding edema, midline compression] --> M_MRI -->|Cascade| E2[Neurological disorders] ... Overall Clinical Reasoning Graph Information: - Age: 48 - Sex: Female - Chief Complains: Recurrent headache and blurred vision for .... - Present Illness: ... Prompt: You are an experienced attending physician responsible for clinical.. ... Patient information is as follows:Patient Information ... Output MUST use EXACTLY this graph TD format ... Outpatient Doctor Clinical Chain graph TD P1[Main Complaint: ...] -->|Cascade| S1[...] ... P2[Present Illness: ...] -->|Cascade| S2[...]... Laboratory: - White Blood Cell Count (WBC): 6.8 Γ10βΉ/L - Hemoglobin: 12.5 g/dL - NSE: 18.5 ng/mL - CEA: 21.6 ng/mL - ... Prompt: You are an experienced laboratory physician ... ... The laboratory test results are ...:Laboratory Results ... Output MUST use EXACTLY this graph TD format ... Outpatient Doctor Clinical Chain Radiology: - CT Findings: An irregular mixed-density mass measuring ... - MRI Findings: The lesion in the right temporal lobe appears ... Prompt: You are an experienced imaging physician responsible for ... The imaging test results are: Radiology Results ... Output MUST use EXACTLY this graph TD format ... Outpatient Doctor Clinical Chain Pathology: - Microscopic Findings: Tumor cells exhibit diffuse infiltrative ... - Pathological Diagnosis: Malignant glioma (WHO Grade IV). ... Prompt: You are an experienced pathology physician responsible ...The pathology test results are: Pathology Results ... Output MUST use EXACTLY this graph TD format ... Outpatient Doctor Clinical Chain graph TD L1[NSE: ...] -->|Cascade| S3[...] ... L2[CEA: ...] -->|Exclusion| S4[...]... graph TD C_CT [Head CT: ...] -->|Cascade| D_Occupy [...] M_MRI[...] --> |Cascade| D_Tumor[...] graph TD E2 -->|Cascade| R_Path[...] ... R_Path -->|Cascade| F_Glioma[...] ... Main Complaint: Recurrent headaches, blurred vision Present Illness: Symptoms progressively worsen, with nausea and vomiting. CT Findings: Mixed density mass in the right temporal lobe, surrounding edema, ... MRI Findings: The right temporal lobe has uneven enhancement, necrotic cystic changes, and ... Pathology Findings: Malignant glioma WHO grade IV. Tumor markers: CEA: 21.6 ng/mLβ (Possibility of cancer) Tumor markers: NSE: 18.5 ng/mLβ (Neurological disorders) Transfer Department: Department of Oncology Transfer Department: Department of Neurosurgery Department Overview Transfer Criteria Urgency Guidance Key Diagnostic Clues ...... Single-Agent Generation Multi-Agent Collaboration Deteminded by Expert Doctor As the head of the medical team, please integrate all opinions from other doctors to obtain the final opinion on which department the patient should be transferred to the most: ...retrieved medical knowledge:Referral Knowledge ...reasoning graphs: Reasoning_Graphs ...options available for selection: Options ... Return the final referral decision derived from the optimal reasoning graph. The output MUST be... Figure 3: The overview of Multi-Agent Structured Graph Reasoning (MASGR) framework. The upper part is the process of single-agent reasoning graph generation. The middle part is multi-agent collaboration to correct and integrate the final reasoning graphs. The lower part is how our expert agent selects final department. 5. Methodology To address the medical referral task, we propose a Multi-Agent Clinical Reasoning Graph Construction Framework, which models clinical decision-making as a multi-stage graph construction process. As illustrated in Figure 3, the framework dynamically constructs and aggregates reasoning structures across heterogeneous medical domains. It operates in three sequential stages: (1) Local Reasoning Graph Construction by single agent, (2) Global Reasoning Graph Co-Building through multi-agent collaboration, and (3) Expert Knowledge Arbitration for final decision. 5.1. Stage 1: Local Reasoning Graph Construction Considering real-world clinical workflows, medical data can be categorized into four main types: outpatient records, laboratory tests, radiology reports, and pathology reports. Accordingly, we define four domain agents: ν΄ = ν΄ νν’ν‘ ,ν΄ ννν ,ν΄ ννν ,ν΄ ννν‘ : Preprint submitted to ElsevierPage 7 of 18 Each agent ν΄ ν is responsible for interpreting data from its respective domain ν ν and constructing a structured local reasoning graph. Formally, for each agent ν΄ ν , we define: ξ³ ννννν,ν = ν΄ ν (ν ν ) = (ξ ν ,ξ± ν ) where ξ³ ννννν,ν is the local reasoning graph, ξ ν denotes the set of extracted clinical evidence nodes, and ξ± ν represents the logical relationships among them. To ensure consistent and controllable reasoning behavior across heterogeneous domains, each agent ν΄ ν is guided by a structured prompt that specifies how raw clinical inputs are abstracted into evidence nodes, how intermediate clinical inferences are formed, and how referral-oriented conclusions are represented in a graph form. The prompt template shared by all agents is: You are domain_specific physician responsible for ... and generate a reasoning graph to recommend the most appropriate specialty department. ... The domain-specific data are as follows: domain_specific_data Generate a department referral reasoning graph by following these principles: 1. Start with abnormal test results and their deviations from reference ranges. . . . Output MUST use EXACTLY this graph TD format (with line breaks between nodes) and the content is for reference only: . . . 5.2. Stage 2: Global Reasoning Graph Building This stage aims to integrate the local graphs into the global reasoning graphs through multi-agent collaboration. The process consists of two key mechanisms: 1) Cross-Domain Linking, where agents exchange and reconcile their reasoning structures through relationship-aware updates, and 2) Graph Aggregation, where the converged local graphs are merged into logically coherent global graph. 5.2.1. Cross-Domain Linking At iteration ν‘, each agent ν΄ ν maintains its current reasoning graph ξ³ (ν‘) ννννν,ν = (ξ (ν‘) ν ,ξ± (ν‘) ν ). During each round of collaboration, every agent interacts with all others, updating its graph: ξ³ (ν‘+1) ννννν,ν = Ξ¦ ν ( ξ³ (ν‘) ννννν,ν ,ξ³ (ν‘) ννννν,ν β£ ν β ν ) , where Ξ¦ ν (β ) denotes the graph update operator of agent ν΄ ν . To ensure that cross-agent interactions lead to consistent and interpretable structural updates, the operator Ξ¦ ν (β ) is implemented via a structured interaction prompt, which instructs each agent to revise its local reasoning graph by incorporating new evidence and feedback from other agents while preserving explicit reasoning relations. A generic interaction prompt shared by all agents is summarized below: As the doctor_type doctor, please generate an UPDATED reasoning graph based on the original assessment, new medical data, and feedback from other doctors. Output MUST use EXACTLY this graph TD format (with line breaks between nodes) and the content is for reference only: ... Use specific relationship types between nodes: ... Please ensure that the output strictly follows the above format. Avoid any additional text or explanations outside the graph structure. Let ν βΆ β ν ξ (ν‘) ν β ξ― denote the mapping from agent-specific evidence to a shared clinical concept space ξ―. For each iteration ν‘, agent ν΄ ν updates its reasoning structure based on the following relational mappings: : Preprint submitted to ElsevierPage 8 of 18 οΏ½ οΏ½ οΏ½ οΏ½ οΏ½ support Cas. (a) Cascade οΏ½ οΏ½ οΏ½ οΏ½ οΏ½ s u p p o r t s u p p o r t (b) Syndrome οΏ½ οΏ½ οΏ½ οΏ½ οΏ½ supportCon. (d) Context (c) Exclusion E x c . οΏ½ οΏ½ οΏ½ οΏ½ οΏ½ s u p p o r t E.g.:Long-term hypertension (οΏ½ οΏ½ ) causes structural heart changes (οΏ½ οΏ½ ), which eventually results in heart failure (οΏ½). E.g.:The Elevated Troponin Levels (οΏ½ οΏ½ ) and ST-segment elevation on ECG (οΏ½ οΏ½ ) provide cross-modal validation, strongly confirming the diagnosis Myocardial Infarction (οΏ½). E.g.:While symptoms Acute Dyspnea (οΏ½ οΏ½ ) suggest a clot, the Normal D-dimer level (οΏ½ οΏ½ ) acts as strong negative evidence, making Pulmonary Embolism (οΏ½) highly unlikely. E.g.:A fever (οΏ½ οΏ½ ) is generic, but when interpreted within the context of βRecent travel to endemic tropical regionsβ (οΏ½ οΏ½ ), it becomes a strong indicator for Malaria (οΏ½). Syn. Figure 4: The four logical relationships among clinical evidence. (a) Cascade: νΈ ν΄ leads to νΈ ν΅ , which ultimately supports outcome ν . (b) Syndrome: νΈ ν΄ and νΈ ν΅ jointly point to outcome ν . (c) Exclusion: νΈ ν΅ conflicts with outcome ν , which was derived from νΈ ν΄ . (d) Context: νΈ ν΅ can only lead to outcome ν when νΈ ν΄ serves as a precondition. (i) Cascade: This relation captures a causal trajectory where an antecedent condition (νΈ ν΄ ) triggers a subsequent pathological state (νΈ ν΅ ), which directly leads to the final diagnosis (ν ). If two agents discover sequential or causal chains across their domains: ν (ν) ν Cas. ββ ν (ν) ν , ν (ν) ν Cas. ββ ν (ν) ν and their mapped concepts satisfy ν(ν (ν) ν ) = ν(ν (ν) ν ),ν(ν (ν) ν ) = ν(ν (ν) ν ), then a unified progressive edge is established: [ν ν ] Cas. ββ [ν ν ] β ξ± (ν‘+1) ν βͺ ξ± (ν‘+1) ν . Here, the notation [ν ν ] Cas. ββ [ν ν ] denotes that a pathological evolution is formed between the semantic equivalence classes [ν ν ] and [ν ν ], which represent concept-level evidence clusters shared across agents. The edge set ξ± (ν‘+1) ν βͺξ± (ν‘+1) ν indicates that this newly unified edge is jointly integrated into the updated reasoning graphs of agents ν΄ ν and ν΄ ν at iteration (ν‘ + 1), thus synchronizing their causal reasoning structures within the multi-agent system. (i) Syndrome: This relation describes a consensus mechanism where independent clinical findings (νΈ ν΄ and νΈ ν΅ ) mutually reinforce each other to heighten the confidence of the diagnosis (ν ). If multiple agents provide : Preprint submitted to ElsevierPage 9 of 18 complementary evidence ν (ν) 1 ,ν (ν) 2 ,...,ν (ν) ν supporting the same conclusion ν(ν£): ν(ν (ν) 1 ),ν(ν (ν) 2 ),...,ν(ν (ν) ν ) β ν(ν£), then a conjunctive structure is formed in the next update: [ν 1 ],[ν 2 ],...,[ν ν ] syn. ββ [ν£] β ξ± (ν‘+1) ν . (i) Exclusion: This relation represents a reasoning conflict where positive evidence (νΈ ν΄ ) suggests the diagnosis (ν ), while contradictory evidence (νΈ ν΅ ) acts as a counter-indicator to rule it out. If cross-agent evidence pairs are known to be mutually exclusive according to the medical ontology ξ : ν(ν (ν) ν ),ν(ν (ν) ν ) β ξ, then a contradiction edge is added: [ν ν ]β [ν ν ] β ξ± (ν‘+1) ν βͺ ξ± (ν‘+1) ν . (iv) Context: This relation indicates that the diagnostic validity of a specific symptom (νΈ ν΅ ) towards the diagnosis (ν ) is conditional upon or amplified by the patientβs background context (νΈ ν΄ ). When the validity of an evidence ν (ν) ν depends on a prerequisite condition ν (ν) ν : ν (ν) ν Con. ββ ν (ν) ν , the conditional dependency is preserved: [ν ν ] Con. ββ [ν ν ] β ξ± (ν‘+1) ν . After ν iterations, all agents reach consensus, yielding a converged set of reasoning graphs ξ³ β ν : ξ³ (ν) ννννν,ν = ξ³ (νβ1) ννννν,ν = ξ³ β ννννν,ν . 5.2.2. Graph Aggregation The global reasoning graphs are obtained via an aggregation operator Ξ(β ) that merges the converged local graphs into a set of candidate global reasoning graphs: ν νννν = Ξ ( ξ³ (ν) ννννν,ν 4 ν=1 ) = ξ³ νννννν,1 ,...,ξ³ νννννν,ν Each candidate ξ³ νννννν,ν βν νννν represents a specific, logically connected diagnostic pathway derived from the multi-agent consensus. This setν νννν serves as the input for the subsequent Expert Knowledge Arbitration stage. 5.3. Stage 3: Expert Knowledge Arbitration In this final stage, we employ an Expert Knowledge Arbitration mechanism to identify the most clinically valid reasoning path from the set of candidate global graphsν νννν . The arbitration process, formalized as function Ξ¦ ννν , evaluates each candidate graph along three critical dimensions to ensure high-quality clinical decision-making: β’ Evidence Completeness: Verifying whether the reasoning graph sufficiently incorporates all key clinical evidence from the patient case, including symptoms, medical history, diagnostic tests, and examination results. β’ Reasoning Coherence: Assessing whether the causal and logical structure of the graph is medically reasonable, internally consistent, and free from logical conflicts. β’ Knowledge Compliance: Ensuring that the diagnostic path and referral conclusions strictly conform to established clinical guidelines and expert rules defined in the Expert Knowledge Base (ξ· νν₯νννν‘ ). : Preprint submitted to ElsevierPage 10 of 18 To operationalize Ξ¦ ννν , the expert agent is guided by a structured arbitration prompt, which instructs it to semantically validate candidate reasoning graphs against expert knowledge and select the optimal referral outcome. The generic prompt used by ν΄ νν₯νννν‘ is summarized below: As the head of the medical team, please integrate all opinions from other doctors to obtain the final opinion on which department the patient should be transferred to the most. Inputs: 1. Patient case information: patient_case 2. Retrieved expert medical knowledge: retrieved_info 3. Reasoning graphs: ... ... Arbitration principles: β Evaluate each reasoning graph for evidence completeness, reasoning coherence, and knowledge compliance. β ... Output requirements: Return the final referral decision derived from the optimal reasoning graph. The output MUST be ... Through this arbitration process, the system filters out hallucinatory or incomplete reasoning paths and outputs the optimal reasoning graph: ξ³ ννν‘νννν = Ξ¦ ννν (ν νννν ,ξ· νν₯νννν‘ )(1) Finally, the department recommendation is determined by traversing the diagnostic path of the verified optimal graph. We define a decision function Ξ¨ νννννν that maps the conclusion of the reasoning chain to a specific medical department: ν νννν‘ = Ξ¨ νννννν (ξ³ ννν‘νννν ), ν νννν‘ β ξ°.(2) 6. Experiment 6.1. Experimental Settings 6.1.1. Dataset Considering there is no dataset for complex medical referral task, we first construct a corresponding dataset based on inpatient records from a real-world hospital, which is mentioned in Section 3. This dataset includes all relevant information prior to patient department transfer, such as basic patient information (e.g., sex, age, present illness), laboratory indicators (e.g., blood tests), radiology findings (e.g., X-ray, MRI), and pathology findings (e.g., frozen pathology). It also contains records of patient department transfers. The dataset covers 952 patients across 37 departments and includes data from various medical examination modalities with temporal information. 6.1.2. Baseline Methods We compare our framework with the following methods: β’ Large Language Models: Considering the advancements of general-purpose LLMs in medical domain, we select several models for comparison, including GPT-5 (Achiam et al., 2023), Qwen-Turbo (Team, 2024), Glm- 4-Flash (GLM, Zeng, Xu, Wang, Zhang, Yin, Zhang, Rojas, Feng, Zhao et al., 2024), DeepSeek-V3 (Liu, Feng, Xue, Wang, Wu, Lu, Zhao, Deng, Zhang, Ruan et al., 2024a), and DeepSeek-R1 (Guo, Yang, Zhang, Song, Zhang, Xu, Zhu, Ma, Wang, Bi et al., 2025). β’ Prompt-Based Methods: We compare against prompt-enhanced methods for LLMs, such as chain-of-thought (CoT) (Wei, Wang, Schuurmans, Bosma, Xia, Chi, Le, Zhou et al., 2022), tree-of-thought (ToT) (Yao, Yu, Zhao, Shafran, Griffiths, Cao and Narasimhan, 2023), and self-consistency (Wang, Wei, Schuurmans, Le, Chi, Narang, Chowdhery and Zhou, 2022). : Preprint submitted to ElsevierPage 11 of 18 β’ Multi-Agent Systems: We also compare with medical multi-agent frameworks, such as MedAgents (Tang et al., 2024), MDAgents (Kim et al., 2024), MMA (Peng et al., 2025a) and ToR (Peng, Cui, Xie, Cai and Li, 2025b). These frameworks are based on multi-agent collaboration to complete medical tasks. 6.1.3. Experimental Details For Large Language Models, we directly use the corresponding API interfaces from the respective platforms, including OpenAI 4 , Qwen 5 , ChatGLM 6 and DeepSeek 7 . For prompt-based methods and multi-agent systems, we use DeepSeek-V3 as the base model to ensure a fair comparison. Similarly, we access the API interfaces provided by the DeepSeek platform to implement these methods. All methods are executed in a zero-shot setting. 6.1.4. Evaluation Metrics In this study, we select both objective and subjective metrics to evaluate proposed methods. For objective evaluation, we model the task as a multi-class classification problem, where the goal is to select appropriate departments for patient based on all available medical examination data. Therefore, we use precision, recall, and F1 score as objective evaluation metrics to assess the performance of different methods. For subjective evaluation, we assess the clinical reasoning quality of outputs from different methods. Specifically, two evaluation dimensions are defined: 1. Evidence Completeness (EC): β’ Definition: This metric evaluates whether the reasoning graph sufficiently incorporates all key clinical evidence from the patient case. β’ Scoring Guidelines (0-5 points): β 5.0 points: The graph captures all essential clinical findings. Every key piece of evidence required for the diagnosis is explicitly represented and correctly categorized. β 4.0 points: Includes the vast majority of core evidence. Only negligible details are omitted, which does not affect the final clinical decision. β 3.0 points: Captures the primary complaint and major symptoms but misses supporting diagnostic details. The basis for diagnosis is present but not robust. β 2.0 points: Misses significant clinical evidence or introduces irrelevant βnoiseβ nodes that distract from the main diagnosis. The graph fails to reflect the complexity of the patientβs condition. β 1.0 points: Fails to incorporate essential clinical data or contains hallucinations (evidence not present in the source text). The graph is medically insufficient. 2. Reasoning Coherence (RC): β’ Definition: This metric assesses whether the logical connections among different clinical evidence are medically reasonable and consistent with established diagnostic knowledge. β’ Scoring Guidelines (0-5 points): β 5.0 points: The reasoning chain is logically flawless and explicitly causal. All deductions align perfectly with standard clinical guidelines, offering a clear interpretability path. β 4.0 points: The reasoning is generally sound and medically plausible. There may be minor logical gaps or implicit steps, but the overall diagnostic direction is correct and justifiable. β 3.0 points: The logical path is traceable but lacks depth. The connection between evidence and diagnosis relies on weak justifications or generic associations rather than specific pathological reasoning. β 2.0 points: Contains noticeable logical contradictions or significant leaps (non sequiturs). The graph connects unrelated evidence to conclusions without valid medical basis. β 1.0 points: The reasoning is medically invalid, dangerous, or completely unstructured. The relationships between nodes violate basic medical principles. 4 https://openai.com/ 5 https://chat.qwen.ai/ 6 https://chatglm.cn/ 7 https://w.deepseek.com/ : Preprint submitted to ElsevierPage 12 of 18 Table 2 Overall performance of different methods, including precision, recall and F1. EC means Evidence Completeness and RC means Reasoning Coherence. The best results are marked in bold. Statistical significance is evaluated using a paired t-test, and improvements are considered significant at ν < 0.01. CategoryMethodsPrecision(%) Recall(%) F1(%) EC RC LLMs GPT-5 (2025) (Achiam et al., 2023)85.095.189.73.833.4 Qwen-Turbo (2024) (Team, 2024)82.997.089.43.83.36 Glm-4-Flash (2024) (GLM et al., 2024)65.195.077.23.433.1 DeepSeek-V3 (Liu et al. 2024) (Liu et al., 2024a)82.993.888.03.83.33 DeepSeek-R1 (Guo et al. 2025) (Guo et al., 2025)74.691.081.943.5 Prompt-Based Self-Consistency (Wang et al. 2022) (Wang et al., 2022)34.289.349.43.53.1 CoT (Wei et al. 2022) (Wei et al., 2022)87.392.990.03.873.5 ToT (Yao et al. 2023) (Yao et al., 2023)52.981.064.03.473.17 Multi-Agent MedAgents (Tang et al. 2024) (Tang et al., 2024)90.989.790.34.13.47 MDAgents (Kim et al. 2024) (Kim et al., 2024)88.289.288.74.13.5 MMA (Peng et al. 2024) (Peng et al., 2025a)90.492.591.44.473.8 ToR (Peng et al. 2025) (Peng et al., 2025b)91.893.392.64.33.83 OurMulti-Agent Structured Graph Reasoning (MASGR)95.295.495.3 4.57 4.2 Table 3 Ablation study on different Agents. βββ means presence of the doctor agent and βΓβ means absence of the doctor agent. Outpatient Laboratory Radiology Pathology Expert P.(%) R.(%) F1(%) βΓ82.993.888.0 βΓ83.195.088.7 βΓ89.3 97.093.0 βΓ90.396.093.1 β95.295.4 95.3 Specifically, three experienced clinicians independently assessed each generated reasoning graph based on the two dimensions. Scoring is precise to one decimal place, and the final score for each dimension is computed as the average rating across all evaluators. 6.2. Main Results As shown in Table 2, we compared the performance of LLMs, prompt-based methods, multi-agent methods, and our framework using the evaluation metrics of precision, recall, and F1. We present the following key findings: (i) For complex medical tasks involving different types of medical data, multi-agent methods outperform most of LLMs and prompt-based methods. Specifically, the F1 score of MedAgents is 90.3%, whereas DeepSeek-R1 and CoT have F1 scores of 81.9% and 90.0%, respectively. Unlike traditional medical tasks, this task involves multiple types of medical data, which need to be simultaneously analyzed and correlated to determine the most appropriate department for transfer. As for LLM methods, they face the challenge of processing multiple medical data types from different modalities simultaneously, which can cause the modelβs attention to be dispersed and reduce its ability to effectively focus on the most relevant information. This leads to a decrease in overall performance. In contrast, in the multi-agent framework, each agent focuses on its specific task without interference from other medical data. Through collaboration, agents provide consistent diagnostic results, allowing for more comprehensive analysis and a more complete clinical reasoning output. (i) Our framework outperforms existing multi-agent systems. The F1 score of MASGR is 95.3%, compared to the F1 score of 91.4% for MMA. Unlike other multi-agent frameworks, our method proposes a collaborative multi- agent approach that constructs a clinical reasoning graph, associating key clinical information from different types of medical data, which can better capture and analyze complex medical data, improving reasoning accuracy. Additionally, we introduce an expert agent with clinical referral knowledge, which provides essential information on departmental functions, transfer criteria, and urgency guidance, thereby facilitating accurate referral decisions. : Preprint submitted to ElsevierPage 13 of 18 Figure 5: The ablation experimental result on four relationship types. The vertical axis represents the F1 score, while the horizontal axis illustrates the relationship types included in the reasoning graph. A relationship matrix is used for visualization, where greenβ indicates that a specific relationship type is incorporated and red Γ denotes its exclusion. Figure 6: Ablation study on different mechanism or module. βw/oβ means our framework without a certain part. (i) For the subjective evaluation metrics, our framework achieved the highest scores in both Evidence Completeness (EC) and Reasoning Coherence (RC). Specifically, the proposed method obtained an EC score of 4.57 and an RC score of 4.20, outperforming GPT-5 by 0.74 and 0.80, respectively, and exceeding the best-performing multi-agent baseline (ToR) by 0.27 and 0.37. These results demonstrate that, compared with existing single-LLM and multi-agent approaches, our method captures more comprehensive clinical evidence and establishes more coherent relationships among the evidence to support accurate decision-making. This capability is particularly crucial in real- world medical scenarios, as it enables the generation of more trustworthy reasoning chains, thereby offering high value for clinical decision support systems. 6.3. Ablation Study To explore the impact of different agents on the final results, we conduct the ablation study shown in Table 3. In this experiment, different agents were excluded from the system to observe the changes in precision, recall, and F1 score. When a specific doctor agent is excluded, the corresponding medical data or responsibility are assigned to the outpatient doctor agent. We made the following findings: : Preprint submitted to ElsevierPage 14 of 18 Figure 7: Performance comparison of different backbone models across varying discussion rounds. The plot illustrates the F1 score trends of DeepSeek-V3, Qwen-Turbo, GPT-4o-mini, and Gemini-2.0-Flash as the number of multi-agent discussion rounds increases from 1 to 5. (i) As the number of agents increases, the performance improves. Specifically, when the laboratory doctor is introduced, the F1 score increases by 0.7%. Further improvements are seen when the radiology and pathology agents are added, with F1 scores increasing by 4.3% and 0.1%, respectively. After involving the expert doctor agent, the F1 scores continuely increases by 2.2%. Further analysis reveals that when more agents are introduced, the data from different domains are handled by their respective agents, allowing for more precise extraction of key information from each data type, which improves the accuracy of the clinical reasoning graph. (i) When the laboratory, radiology, pathology, and expert doctors are excluded, our framework reduces to a single LLM system, resulting in an F1 score of 88.0%. When faced with complex medical reasoning tasks, a single LLM must process multiple types of medical data simultaneously, which can cause the modelβs attention to shift and lead to the omission of key information from multi-source medical data, thus decreasing model performance. To explore the impact of different modules on our framework, we conduct the corresponding ablation study, as shown in Figure 6. βw/o Reasoning Graphβ refers to a scenario where the multi-agent system generates outputs in natural language rather than using a structured reasoning graph. In this case, the F1 score decreased by 5.1%, indicating that the construction of a reasoning graph enhances the analytical capability of the multi-agent framework in complex medical scenarios, leading to more accurate clinical reasoning. βw/o Multi-Agentβ refers to a situation where each agent constructs its reasoning graph independently and directly merges the results, without any inter- agent collaboration or update of the reasoning graph. In this case, the F1 score decreased by 6.1%, suggesting that the reasoning graphs built by individual agents have limited coverage. Multi-agent collaboration allows for a more comprehensive analysis of key clinical evidence across different data sources, thereby improving the overall accuracy. βw/o Referral Knowledgeβ refers to a scenario where the expert agent does not refer to the clinical guideline when making decisions. In this case, the F1 score dropped by 2.3%. The introduction of referral knowledge provides the expert agent with department transfer criteria and urgency guidance, enabling the agent to recommend the most appropriate department for a given medical scenario and thus improving the accuracy of medical referral decisions. To investigate the impact of the four proposed relationship types (Cascade, Syndrome, Exclusion, and Context) on graph construction, we conduct ablation experiments as shown in Figure 5. Within the multi-agent framework, we explicitly remove selected relationship types to observe their influence on the final performance. Results demonstrate that the overall F1 score consistently increases as more relationship types are incorporated, indicating that the four defined relationship types align well with the characteristics of clinical reasoning and effectively support structured inference. Furthermore, among all relationship types, Cascade exhibits the most substantial contribution to performance improvement. For example, in the βremove 1 relationshipβ setting, excluding Cascade leads to a drop in F1 from 95.3% to 90.6%. Similar significant declines are observed in both the βremove 2β and βremove 3β settings when Cascade is omitted. This finding suggests that, in complex medical referral scenarios, long-chain dependencies frequently exist across clinical evidence. By explicitly modeling these continuous inferential links through the proposed Cascade relation, our framework enhances the ability of multi-agent framework to perform coherent and clinically reliable reasoning. : Preprint submitted to ElsevierPage 15 of 18 Information Data: ... Recurrent cough, sputum and exertional dyspnea for 2 months, worsened for 1 week; ... Fever 38.8Β°C... Laboratory Data: WBC: 12.5 Γ 10βΉ/L; Neutrophil %: 82%; CRP: 55 mg/L; ESR: 30 m/h; Arterial oxygenation: mild hypoxemia; ... Radiology Data: Chest X-ray: patchy high-density opacity in the right upper lobe with poorly defined margins ... Which department should the patient be referred to? A. Intensive Care Unit (ICU) B. Department of Oncology C. Department of Pulmonology D. Gastrointestinal Surgery Pathology Data: ... loose fibroblastic connective tissue (Masson bodies) and chronic inflammatory cell infiltration ... small clusters of epithelial cells displaying a hobnail appearance ... GPT-5: CoT: MMA: Ours: C. Department of Pulmonology Step-by-step Reasoning: 1. cough, sputum, and dyspnea ...; 2.Elevated WBC, ...; 3.Chest X-ray shows patchy opacity ...These findings suggest an infectious process involving the lungs; 4.Pathology shows Masson bodies. Final referral department is Pulmonology. C. Department of Pulmonology C. Department of Pulmonology Outpatient Doctor: ... indicate lower-respiratory disease. Laboratory Doctor: WBC, ..., supporting infection. Radiology Doctor: right upper lobe patchy consolidation ... Pathology Doctor: ... characteristic of organizing pneumonia. ... All agents support a diagnosis of pulmonary infection/organizing pneumonia, indicating referral to Pulmonology department.. C. Department of Pulmonology ... chronic productive cough, ... , right upper-lobe consolidation ... WBC and CRP point to an infectious pulmonary ..., suggesting involvement of the lower respiratory tract. ... Chief Complaint: Recurrent cough, sputum and exertional dyspnea for 2 months ... Lab Results: WBC β ; CRP β X-Ray Findings: patchy high-density opacity in the right upper lobe with poorly defined margins ... Chief Complaint: Fever 38.8Β°C Department: Department of Pulmonology Possibility: bacterial pneumonia GPT-5: CoT: MMA: Ours: Information Data: ... Recurrent cough, sputum, shortness of breath for 3 weeks; The patient developed cough and white sputum 3 weeks ago, ... Laboratory Data: WBC: 13.2 Γ10βΉ/L; ... CRP: 82 mg/L; ... Radiology Data: CT: ...Segmental consolidation in the right lower lobe ....; Carotid Doppler Ultrasound: ... A small, unstable hypoechoic plaque with a micro-ulceration in the left carotid bulb(3 m)... Which department should the patient be referred to? A. Department of Pulmonology B. Department of Cardiology C. Department of Oncology D. Department of Endocrinology A. Department of Pulmonology ... progressive cough, sputum production, ...segmental consolidation ... elevated WBC (13.2Γ10βΉ/L) further support an infectious etiology. ... Pulmonology is the referral department. Step-by-step reasoning: 1.cough, sputum, shortness of breath ... 2.CT demonstrates consolidation ... 3.Elevated inflammatory markers (WBC, CRP) ... These findings are suggestive of a pulmonary infection, and referral to the Pulmonology department is recommended A. Department of Pulmonology A. Department of Pulmonology Outpatient Doctor: The patient has cough, sputum, ... Laboratory Doctor: Lab values show elevated WBC (13.2Γ10βΉ/L), ...; Radiology Doctor: CT shows segmental consolidation with air bronchogram. Based on the consensus of all agents, the findings indicate a pulmonary infection, and referral to the Pulmonology department is recommended. B. Department of Cardiology Chief Complaint: cough, sputum, shortness of breath for 3 weeks X-Ray Findings: Patchy high-density opacity in the right lower lobe. CT Findings: Segmental consolidation in the right lower lobe ... Lab Results: WBC β ; CRP β; PCT β Department: Department of Pulmonology Present Illness: Hypertension for 10+ years; blood pressure fluctuating between 150β160/90β95 mmHg ... Ultrasound Findings: hypoechoic plaque with a micro-ulceration in the left carotid bulbοΌ3 mοΌ.. Department: Department of Cardiology Possibility: Systemicperfusion/vascular risk Possibility: Pulmonary complication risk Figure 8: Qualitative comparison of case studies. The figure is divided into two parts: Case 1 (Left) and Case 2 (Right). In each part, the upper section presents the multi-source medical data, including patient information, laboratory results, radiology findings, and pathology data. The lower section displays the referral decisions and reasoning processes of GPT-5, CoT, MMA, and our method. Specifically, the bottom-most part illustrates the clinical reasoning graph constructed by our method to support its decision. 6.4. Impact of Multi-Agent Discussion Rounds We analyze the impact of the number of discussion rounds on the reasoning performance. In Figure 7, the horizontal axis represents the number of discussion rounds (ranging from 1 to 5), and the vertical axis denotes the F1 score. We evaluate four representative lightweight foundation models as backbones within our framework: DeepSeek- V3, Qwen-Turbo, GPT-4o-mini, and Gemini-2.0-Flash. All models exhibit a performance increase as the discussion deepens from Round 1 to Round 3, demonstrating the effectiveness of the multi-agent collaboration mechanism. Among the evaluated models, DeepSeek-V3 achieves superior performance. It reaches a peak F1 score of 95.3% at Round 3. Extending the discussion beyond this point (to Round 4 and 5) yields negligible gains or slight declines. This suggests that the reasoning graph converges effectively at Round 3, making it the optimal hyperparameter for efficiency and accuracy. 6.5. Case Study To provide a concrete understanding of our method, we visualize two representative cases in Figure 8. The left part displays the first case, a patient presenting with recurrent cough, sputum, and shortness of breath. Methods like GPT-5, CoT, and MMA are misled by the prominent respiratory symptoms. They focus heavily on the βcoughβ and βlung consolidationβ shown in the radiology data, leading them to incorrectly refer the patient to the Department of Pulmonology. In contrast, our method successfully identifies the underlying cause. By constructing a reasoning graph, our model captures critical non-respiratory evidence, specifically the β10-year history of hypertensionβ and the βunstable plaque in the carotid bulbβ. The graph connects these nodes to identify a high risk of systemic vascular failure. Consequently, our method correctly recommends the Department of Cardiology. The right part shows Case 2, which highlights a limitation of current agents. The patient presents with cough and fever, and pathology data shows βMasson bodiesβ (often a sign of inflammation) alongside epithelial cells with a βhobnail appearanceβ. The correct referral is the Department of Oncology. However, GPT-5, CoT, MMA, and our method all incorrectly select the Department of Pulmonology. The models focus on the βMasson bodiesβ and inflammatory markers, interpreting : Preprint submitted to ElsevierPage 16 of 18 the condition as simple pneumonia. They fail to recognize that the βhobnail appearanceβ is a high-risk feature for malignancy. In this case, the inflammation is likely a secondary change caused by tumor obstruction. Our agent lacks the specific expert knowledge to distinguish this subtle pathological feature. This suggests a need for future improvements, such as training more specialized sub-agents to detect fine-grained malignancy cues. 7. Conclusion In this paper, we propose Multi-Agent Structured Graph Reasoning (MASGR), a multi-agent framework for complex medical referral task. In this framework, we first utilize four doctor agents to construct a reasoning graph which links the clinical evidences across different types of medical. Then, we set an expert agent combined with a pre- defined referral guideline to select the most appropriate department for the patient. Experiments show that MASGR outperforms other baseline methods, demonstrating stronger capability in cross-modal evidence interpretation and prioritizing referral decisions. Acknowledgments This research is supported by the National Natural Science Foundation of China (62476097), the Fundamental Re- search Funds for the Central Universities, South China University of Technology (x2rjD2250190), Guangdong Provin- cial Fund for Basic and Applied Basic ResearchβRegional Joint Fund Project (Key Project) (2023B1515120078), Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (2024B1515040010), the Hong Kong Polytechnic University under the Postdoc Matching Fund Scheme (Project No. P0049003), the Hong Kong Research Grants Council under the Theme-based Research Scheme (project no. T22-501/23-R). References Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al., 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 . Anyanwu, E., Abedi Harrison, O., Onohwakpor Efe, A., 2015. The practice of medical referral: Ethical concerns. Am J Public Health Res 3, 31β5. Falotico, R., Quatto, P., 2015. Fleissβ kappa statistic without paradoxes. Quality & Quantity 49, 463β470. Fan, A., Gokkaya, B., Harman, M., Lyubarskiy, M., Sengupta, S., Yoo, S., Zhang, J.M., 2023. Large language models for software engineering: Survey and open problems, in: 2023 IEEE/ACM International Conference on Software Engineering: Future of Software Engineering (ICSE- FoSE), IEEE. p. 31β53. Fan, Y., Sun, H., Xue, K., Zhang, X., Zhang, S., Ruan, T., 2025. Medodyssey: A medical domain benchmark for long context evaluation up to 200k tokens, in: Findings of the Association for Computational Linguistics: NAACL 2025, p. 32β56. GLM, T., Zeng, A., Xu, B., Wang, B., Zhang, C., Yin, D., Zhang, D., Rojas, D., Feng, G., Zhao, H., et al., 2024. Chatglm: A family of large language models from glm-130b to glm-4 ateam2024qwen2l tools. arXiv preprint arXiv:2406.12793 . Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al., 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 . Heuer, L., Palmieri, L., Mannucci, A., Koenig, S., Magnusson, M., 2024. Benchmarking multi-robot coordination in realistic, unstructured human- shared environments, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE. p. 14541β14547. Hong, S., Zheng, X., Chen, J., Cheng, Y., et al., 2023. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.00352 3, 6. Hou, X., Zhao, Y., Liu, Y., Yang, Z., Wang, K., Li, L., Luo, X., Lo, D., Grundy, J., Wang, H., 2024. Large language models for software engineering: A systematic literature review. ACM Transactions on Software Engineering and Methodology 33, 1β79. Kim, Y., Park, C., Jeong, H., Chan, Y.S., Xu, X., McDuff, D., Lee, H., Ghassemi, M., Breazeal, C., Park, H., et al., 2024. Mdagents: An adaptive collaboration of llms for medical decision-making. Advances in Neural Information Processing Systems 37, 79410β79452. Kline, A., Wang, H., Li, Y., Dennis, S., Hutch, M., Xu, Z., Wang, F., Cheng, F., Luo, Y., 2022. Multimodal machine learning in precision health: A scoping review. NPJ digital medicine 5, 171. Kumari, M., Chauhan, R., Jain, R., Garg, P., 2025. A novel context-aware retrieval framework for biomedical knowledge integration with large language models. Information Fusion , 103902. Liang, D., Wei, X.Y., Zheng, C., 2026. Multi-agent undercover gaming: Hallucination removal through counterfactual test for multimodal reasoning, in: Proceedings of the AAAI Conference on Artificial Intelligence, p. 6807β6815. Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al., 2024a. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 . Liu, C., Tian, Y., Chen, W., Song, Y., Zhang, Y., 2024b. Bootstrapping large language models for radiology report generation, in: Proceedings of the AAAI Conference on Artificial Intelligence, p. 18635β18643. Liu, S., McCoy, A.B., Wright, A., 2025. Improving large language model applications in biomedicine with retrieval-augmented generation: a systematic review, meta-analysis, and clinical development guidelines. Journal of the American Medical Informatics Association 32, 605β615. : Preprint submitted to ElsevierPage 17 of 18 Lucas, M.M., Yang, J., Pomeroy, J.K., Yang, C.C., 2024. Reasoning with large language models for medical question answering. Journal of the American Medical Informatics Association 31, 1964β1975. Ma, S., Huang, J., Yang, B., Zhang, F., Wu, J., Shen, Y., Fan, G., Zhang, Z., Zang, Z., 2025. Medla: A logic-driven multi-agent framework for complex medical reasoning with large language models. arXiv preprint arXiv:2509.23725 . Nori, H., King, N., McKinney, S.M., Carignan, D., Horvitz, E., 2023. Capabilities of gpt-4 on medical challenge problems. arXiv preprint arXiv:2303.13375 . Park, J.S., OβBrien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S., 2023. Generative agents: Interactive simulacra of human behavior, in: Proceedings of the 36th annual acm symposium on user interface software and technology, p. 1β22. Peng, Q., Cai, Y., Liu, J., Zou, Q., Chen, X., Zhong, Z., Wang, Z., Xie, J., Li, Q., 2025a. Integration of multi-source medical data for medical diagnosis question answering. IEEE Transactions on Medical Imaging 44, 1373β1385. Peng, Q., Cui, J., Xie, J., Cai, Y., Li, Q., 2025b. Tree-of-reasoning: Towards complex medical diagnosis via multi-agent reasoning with evidence tree, in: Proceedings of the 33rd ACM International Conference on Multimedia, p. 1744β1753. Shaik, T., Tao, X., Li, L., Xie, H., VelΓ‘squez, J.D., 2024. A survey of multimodal information fusion for smart healthcare: Mapping the journey from data to wisdom. Information Fusion 102, 102040. Singhal, K., Tu, T., Gottweis, J., Sayres, R., Wulczyn, E., Amin, M., Hou, L., Clark, K., Pfohl, S.R., Cole-Lewis, H., et al., 2025. Toward expert-level medical question answering with large language models. Nature Medicine , 1β8. Taberna, M., Moncayo, F.G., JanΓ©-Salas, E., Antonio, M., Arribas, L., Vilajosana, E., Torres, E.P., MesΓa, R., 2020. The multidisciplinary team (mdt) approach and quality of care. Frontiers in oncology 10, 85. Tang, X., Zou, A., Zhang, Z., Li, Z., Zhao, Y., Zhang, X., Cohan, A., Gerstein, M., 2024. Medagents: Large language models as collaborators for zero-shot medical reasoning, in: Findings of the Association for Computational Linguistics ACL 2024, p. 599β621. Team, Q., 2024. Qwen2 technical report. arXiv preprint arXiv:2407.10671 . Toma, A., Lawler, P.R., Ba, J., Krishnan, R.G., Rubin, B.B., Wang, B., 2023. Clinical camel: An open expert-level medical language model with dialogue-based knowledge encoding. arXiv preprint arXiv:2305.12031 . Vishwanath, K., Alyakin, A., Alber, D.A., Lee, J.V., Kondziolka, D., Oermann, E.K., 2025. Medical large language models are easily distracted. arXiv preprint arXiv:2504.01201 . Wang, S., Zhao, Z., Ouyang, X., Wang, Q., Shen, D., 2023. Chatcad: Interactive computer-aided diagnosis on medical image using large language models. arXiv preprint arXiv:2302.07257 . Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., Zhou, D., 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 . Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al., 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, 24824β24837. Wen, Y., Wang, Z., Sun, J., 2024. Mindmap: Knowledge graph prompting sparks graph of thoughts in large language models, in: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 10370β10388. Xiong, G., Jin, Q., Lu, Z., Zhang, A., 2024. Benchmarking retrieval-augmented generation for medicine, in: Findings of the Association for Computational Linguistics ACL 2024, p. 6233β6251. Xu, D., Chen, Y., Chai, Z., Xiao, Y., Yan, Y., Ding, W., Wang, H., Jin, Z., Jiao, W., Yue, W., et al., 2026. Knowledge fusion in deep learning-based medical vision-language models: A review. Information Fusion 125, 103455. Yan, W., Liu, H., Wu, T., Chen, Q., Wang, W., Chai, H., Wang, J., Zhao, W., Zhang, Y., Zhang, R., et al., 2024. Clinicallab: Aligning agents for multi-departmental clinical diagnostics in the real world. arXiv preprint arXiv:2406.13890 . Yang, D., Wei, J., Li, M., Liu, J., Liu, L., Hu, M., He, J., Ju, Y., Zhou, W., Liu, Y., et al., 2025. Medaide: Information fusion and anatomy of medical intents via llm-based agent collaboration. Information Fusion , 103743. Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T., Cao, Y., Narasimhan, K., 2023. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems 36, 11809β11822. Zhang, Y., Wei, C., He, Z., Yu, W., 2024. Geogpt: An assistant for understanding and processing geospatial tasks. International Journal of Applied Earth Observation and Geoinformation 131, 103976. Zheng, C., Liang, D., Zhang, W., Wei, X.Y., Chua, T.S., Li, Q., 2024. A picture is worth a graph: A blueprint debate paradigm for multimodal reasoning, in: Proceedings of the 32nd ACM International Conference on Multimedia, p. 419β428. : Preprint submitted to ElsevierPage 18 of 18