Source-linked AI summary
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
TL;DR
Complex medical referral requires synthesizing multimodal records while avoiding overlooked urgency cues and unstructured collaboration. MASGR constructs local and global reasoning graphs with specialized agents, then applies expert knowledge arbitration; it outperforms existing LLM and multi-agent baselines, though a case study reveals missed fine-grained malignancy cues.
Problem
Medical referral requires multimodal evidence synthesis, but LLMs can overlook subtle urgency indicators and existing multi-agent systems rely on unstructured collaboration.
Method
MASGR uses specialized agents to construct local graphs, collaboratively links cross-modal evidence in global graphs, and applies expert knowledge arbitration for department selection.
Results
MASGR outperforms existing LLM and multi-agent methods, achieving an F1 score of 95.3% versus 91.4% for MMA and the highest Evidence Completeness and Reasoning Coherence scores.
Takeaways & Limitations
MASGR supports cross-modal evidence interpretation and prioritizes referral decisions through structured reasoning and clinical referral knowledge.
Takeaways & Limitations
A case study shows the system can miss fine-grained malignancy cues such as the pathological “hobnail appearance,” motivating more specialized sub-agents.
Abstract
from arXiv · showhide
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.
1. Introduction
Medical referral requires integrating heterogeneous clinical data while avoiding missed urgency signals and unstructured reasoning. MASGR addresses these challenges with modality-specific evidence extraction, structured graph relations, and expert arbitration.
- Medical referral synthesizes outpatient narratives, laboratory indicators, radiology imaging, and pathology findings to direct patients to appropriate departments.Accurate referral supports resource allocation and timely treatment for complex conditions.
- General-purpose LLMs may fixate on frequent disease terms and overlook subtle urgency indicators, producing clinically unsafe referrals despite correct disease identification.The paper illustrates respiratory failure being referred to Oncology instead of the ICU when a blood-oxygen decline is missed.
- MASGR reframes referral from classification as structured graph construction inspired by multidisciplinary specialist workflows.The framework explicitly represents evidence and reasoning rather than relying on a single overwhelmed model.
- Specialized outpatient, laboratory, radiology, and pathology agents build local reasoning graphs before global integration.This decomposition is intended to preserve subtle modality-specific cues such as nodule texture and critical blood-gas changes.
- Agents co-build a global graph using cascade, syndrome, exclusion, and context relations to connect, reinforce, or distinguish conflicting clinical evidence.The graph makes diagnostic discordance, such as chronic cancer alongside acute respiratory crisis, explicit for further arbitration.
- Experiments on real-world inpatient records show MASGR significantly outperforms state-of-the-art LLMs and medical multi-agent systems.Ablations identify logical-relation modeling and expert arbitration as crucial for complex cases.
2. Related Work
Related work situates MASGR within broader LLM and multi-agent research, including medical collaboration systems and graph-based clinical reasoning. Existing datasets are compared as part of this context.
- LLMs have been applied broadly in software development, fundamental sciences, social simulation, and medical tasks such as question answering.
- Multi-agent collaboration has improved complex tasks including code development, strategic games, and multi-robot coordination before extending into medicine.
- Table 1 compares existing datasets and indicates that the proposed dataset covers all types of medical data.
- Graph-based clinical reasoning models causal relationships among symptoms, examinations, and diagnoses to improve inference accuracy and interpretability.
3. Data Construction
The paper constructs a real-world inpatient referral dataset and a structured referral knowledge base to support evaluation and expert arbitration. Clinicians then assess the knowledge base for accuracy, completeness, and utility.
- Dataset construction: The dataset contains pre-transfer inpatient information spanning demographics, illness history, laboratory indicators, radiology, pathology, and department-transfer records.
- Dataset construction: Cases are structured as JSON records and evaluated as four-option multi-class referral decisions containing the ground-truth department and distractor departments.
- Dataset statistics: 952 patients across 37 clinical departments and 4,823 disease types comprise the final dataset.The most frequent departments are oncology, gastrointestinal surgery, and thyroid surgery.
- Referral knowledge base: The referral knowledge base organizes department-level reasoning around clinical scope, diagnostic boundaries, transfer constraints, unsuitable cases, and urgency guidance.
- Referral knowledge base: GPT-5 generates knowledge-base content from department-specific prompts, while MedRAG retrieves clinical materials from PubMed, StatPearls, and Wikipedia for factual grounding.
- Knowledge-base validation: Clinician evaluation of a random 10% sample produced average scores of 4.8 for accuracy, 4.7 for completeness, and 4.7 for utility, with Fleiss’ Kappa of 0.76.
4. Problem Formulation
The referral task maps heterogeneous patient records to a target department through three sequential stages: local graph generation, global graph co-building, and expert arbitration. The formulation represents evidence, logical relations, candidate graphs, and department selection explicitly.
- 4.1. Task Input and Objective: Each case consists of heterogeneous patient sources, including laboratory results, imaging reports, and prior medical history, and must be assigned to an appropriate department.
- 4.2. Task Formulation: Reasoning Graph Construction: The task is reformulated as a multi-stage structured graph generation problem rather than a direct referral classification.
- Stage 1: Local Reasoning Graph Construction: In Stage 1, each domain agent independently analyzes its designated source and constructs a preliminary local reasoning graph.
- Stage 1: Local Reasoning Graph Construction: A local graph G_local,i consists of evidence nodes V_i and their internal logical relationships E_i, with all domain graphs retained as the stage output.
- Stage 2: Global Reasoning Graph Co-Building: In Stage 2, agents collaboratively merge the preliminary graphs into candidate global graphs that combine evidence from multiple sources coherently.
- Stage 3: Expert-Knowledge Arbitration: In Stage 3, an expert agent with an expert knowledge base filters candidate graphs for logical validity and selects the most reasonable reasoning chain.
- Stage 3: Expert-Knowledge Arbitration: The final department recommendation is derived from the diagnostic path in the selected optimal reasoning chain and mapped to a department in the candidate set.
- 4.2. Task Formulation: Reasoning Graph Construction: Figure 3 summarizes the framework’s upper local-graph, middle multi-agent integration, and lower expert-selection processes.
5. Methodology
The framework constructs referral decisions through modality-specific local graphs, structured cross-agent relations, graph aggregation, and expert arbitration. This pipeline converts heterogeneous clinical evidence into verified referral reasoning paths.
- Framework overview: The framework models referral as multi-stage graph construction rather than classification, spanning local reasoning, global co-building, and expert arbitration.The stages dynamically construct and aggregate reasoning structures across heterogeneous medical domains.
- Local reasoning graph construction: Four domain agents interpret outpatient, laboratory, radiology, and pathology data, extracting evidence nodes and logical relations into local reasoning graphs.Each agent is guided to abstract raw inputs into evidence nodes, intermediate inferences, and referral-oriented graph conclusions.
- Global graph co-building: Agents iteratively exchange and update graphs through structured interactions that preserve explicit reasoning relations and synchronize cross-domain evidence.The update operator incorporates new evidence and feedback from other agents while maintaining interpretable graph structure.
- Cross-Domain Linking: Cross-domain linking encodes cascade, syndrome, exclusion, and context relations to represent causal chains, mutual support, conflicts, and conditional validity.These relation types align conflicting evidence, including discordance between chronic diagnoses and acute crises.
- Graph aggregation: After convergence, graph aggregation produces candidate global diagnostic pathways that become inputs to expert knowledge arbitration.Each candidate represents a logically connected pathway derived from multi-agent consensus.
- Expert Knowledge Arbitration: Expert arbitration selects the most clinically valid path by checking evidence completeness, reasoning coherence, and knowledge compliance.The arbitration prompt directs semantic validation against expert knowledge before producing the final referral outcome.
6.1. Experimental Settings
The evaluation uses a newly constructed inpatient referral dataset, compares LLMs, prompt-based methods, and multi-agent systems, and assesses both objective referral performance and clinician-rated reasoning quality.
- Dataset: The study constructs a real-world inpatient dataset containing patient information, laboratory indicators, radiology findings, pathology findings, and pre-transfer referral information.The dataset addresses the absence of an existing dataset for complex medical referral.
- Baselines: The comparison includes general-purpose LLMs, prompt-based methods, and medical multi-agent systems evaluated in a zero-shot setting.DeepSeek-V3 is used as the base model for prompt-based and multi-agent comparisons to support a fair comparison.
- Objective metrics: Objective evaluation treats referral as multi-class department selection and reports precision, recall, and F1 score.The metrics assess whether methods select appropriate departments from all available medical examination data.
- Scoring criteria: The scoring guidelines judge whether graphs incorporate essential clinical evidence and connect evidence to diagnoses through medically coherent reasoning.Higher scores require complete evidence representation, explicit causal logic, and alignment with standard clinical guidelines.
- Clinical reasoning metrics: Subjective evaluation rates reasoning graphs on Evidence Completeness and Reasoning Coherence using clinician assessments on a 0-5 scale.Three experienced clinicians independently score each generated graph, with final dimension scores averaged across evaluators.
6.2. Main Results
MASGR outperforms LLM, prompt-based, and existing multi-agent approaches on medical referral, with stronger evidence completeness and reasoning coherence. The results support the value of structured clinical reasoning and expert arbitration for multimodal referral decisions.
- Main comparisons: Multi-agent methods outperform most LLMs and prompt-based methods on complex referral tasks involving multiple medical data types.MedAgents achieves an F1 score of 90.3%, compared with 81.9% for DeepSeek-R1 and 90.0% for CoT.
- Main comparisons: MASGR achieves an F1 score of 95.3%, exceeding MMA’s 91.4% among existing multi-agent systems.The framework links clinical information across modalities through a reasoning graph and adds an expert agent with referral knowledge.
- Ablation visualization: Figure 5 compares F1 scores across four reasoning-graph relationship types using inclusion and exclusion indicators.Green ✓ marks an incorporated relationship type, while red × marks its exclusion.
- Ablation visualization: Figure 6 presents ablation results for different mechanisms or modules, with “w/o” denoting removal of a framework component.The figure is intended to compare performance when specific parts are excluded.
- Subjective evaluation: MASGR achieves the highest scores for Evidence Completeness and Reasoning Coherence.Its EC score is 4.57 and its RC score is 4.20, exceeding GPT-5 by 0.74 and 0.80, respectively.
6.3. Ablation Study
Ablation studies show that performance depends on specialized agents, structured graph reasoning, interagent collaboration, referral knowledge, and the selected graph relations. Cascade contributes especially strongly, while three discussion rounds provide the reported backbone-model optimum.
- Discussion-round analysis: Figure 7 compares F1 trends for four backbone models across one to five multi-agent discussion rounds.The evaluated backbones are DeepSeek-V3, Qwen-Turbo, GPT-4o-mini, and Gemini-2.0-Flash.
- Agent ablations: Adding laboratory, radiology, pathology, and expert agents progressively improves F1 performance.The reported increases are 0.7% with the laboratory agent, 4.3% with the radiology agent, 0.1% with the pathology agent, and 2.2% with the expert agent.
- Agent ablations: Removing the laboratory, radiology, pathology, and expert agents reduces the framework to a single LLM with an F1 score of 88.0%.The passage attributes the decrease to simultaneous processing of multiple medical data types and omission of key information.
- Module ablations: Removing the reasoning graph decreases F1 by 5.1%, while removing multi-agent collaboration decreases F1 by 6.1%.Without collaboration, agents independently construct graphs and merge results without interagent updates.
- Relationship ablations: Incorporating more of the four relationship types consistently increases overall F1, with Cascade producing the largest contribution.When one relationship is removed, excluding Cascade lowers F1 from 95.3% to 90.6%.
- Case studies: Figure 8 qualitatively compares two cases using multimodal medical data, referral decisions, and reasoning processes from several systems.Its lower section includes the clinical reasoning graph constructed by the proposed method.
6.4. Impact of Multi-Agent Discussion Rounds
Increasing discussion from one to three rounds improves reasoning performance across the evaluated backbone models, after which gains become negligible or slightly decline. DeepSeek-V3 performs best and reaches the framework’s peak F1 at round three.
- Discussion-round effects: All four backbone models improve from Round 1 to Round 3, while further discussion yields negligible gains or slight declines.The evaluated models are DeepSeek-V3, Qwen-Turbo, GPT-4o-mini, and Gemini-2.0-Flash.
- Discussion-round effects: DeepSeek-V3 achieves the highest performance, reaching a peak F1 score of 95.3% at Round 3.The authors identify Round 3 as the optimal hyperparameter for efficiency and accuracy because the reasoning graph converges effectively there.
6.5. Case Study
Figure 8 illustrates MASGR’s ability to integrate non-respiratory evidence for correct referral, while also exposing a limitation in recognizing subtle malignancy cues.
- MASGR correctly refers a patient with respiratory symptoms to Cardiology by linking hypertension history with an unstable carotid plaque.Other methods focus on cough and lung consolidation and incorrectly recommend Pulmonology.
- In a second case, the correct Oncology referral is obscured by inflammatory findings and a subtle hobnail epithelial appearance.The agents interpret the case as simple pneumonia rather than recognizing the malignancy risk.
- The agent lacks expert knowledge to distinguish the hobnail appearance as a fine-grained malignancy cue.The paper suggests developing more specialized sub-agents for such pathological features.
7. Conclusion
MASGR uses doctor agents to build a cross-modal reasoning graph and an expert agent with referral guidelines to select departments. Experiments report stronger cross-modal interpretation and referral prioritization than baseline methods.
- MASGR uses four doctor agents to link clinical evidence across medical data types in a reasoning graph.
- An expert agent combined with predefined referral guidelines selects the most appropriate department.
- Experiments show MASGR outperforms baseline methods in cross-modal evidence interpretation and prioritizing referral decisions.