The Big Picture
Using multiple specialist agents that each turn parts of a patient record into a structured evidence graph, then linking those graphs and checking them against expert rules, yields clearer and safer referral decisions than a single language model handling all data at once.
ON THIS PAGE
The Evidence
A multi-agent approach that builds local evidence graphs for outpatient notes, lab results, radiology reports, and pathology reports, then connects them with explicit logical links (cause, support, conflict, context), reduces the chance that a single strong cue will drown out urgent but subtle signals. An expert arbitration step that checks completeness, coherence, and compliance with a curated guideline lets the system prioritize urgent needs (for example, hypoxemia) even when a disease label suggests another department. On a real-world inpatient dataset, this structured pipeline outperformed baseline large language models and prior multi-agent dialogue systems in referral accuracy and interpretability. Chain of Thought Pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1Dataset: 952 inpatient cases drawn from one tertiary hospital
2Scope: referrals across 37 hospital departments
3Design: 4 domain agents (outpatient, lab, radiology, pathology) using 4 relation types (cascade, syndrome, exclusion, context) to build reasoning graphs
What This Means
Engineers building clinical decision support or referral-assistant systems will find a practical architecture for splitting tasks across specialist agents and producing auditable reasoning. Clinical informatics leads and hospital product managers can use the expert arbitration idea to encode operational rules so urgent clinical signals override less critical diagnoses. Orchestrator-Worker Pattern
Key Figures
Fig 1: Figure 1: A comparison of our framework with existing decision-making paradigms for a patient with lung cancer complicated by severe hypoxemia.
Fig 3: 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.
Fig 5: 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 green ✓ \checkmark indicates that a specific relationship type is incorporated and red × \times denotes its exclusion.
Fig 6: Figure 6: Ablation study on different mechanism or module. “w/o” means our framework without a certain part.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
The dataset comes from a single tertiary hospital, so transfer to other hospitals or outpatient settings may require re-tuning the expert guideline and retraining domain prompts. The system works on structured reports rather than raw image pixels, so performance depends on the quality of upstream reporting and extraction. Computational and integration costs are higher than single-model prompts because multiple agents, graph assembly, and rule checking are required. Semantic Capability Matching Pattern
Methodology & More
The approach treats referral as a structured reasoning problem instead of a single-label classification. Four specialist agents convert different parts of a patient record (clinic notes, labs, radiology text, pathology text) into local reasoning graphs of evidence nodes. Agents then co-build a global graph by explicitly linking nodes with defined logical relations—causal chains, mutual reinforcement, conflict/exclusion, and context—so cross-domain discordances are visible rather than lost in freeform dialogue. An expert arbitration module evaluates candidate graphs against a curated guideline for evidence completeness, logical coherence, and operational compliance, allowing rules (for example, a low oxygen pressure threshold) to override disease-focused signals when triage demands it. The system was tested on 952 inpatient cases spanning 37 departments; experiments and ablations show the structured relations and expert arbitration materially improve referral decisions and make the chain of reasoning inspectable. Practical benefits include fewer risky mis-referrals caused by attention getting cues and better auditability, though adoption will need hospital-specific guideline curations and attention to integration costs. Agent Registry Pattern
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
ArXiv preprint with no affiliations, citation metrics, or recognizable top-venue placement. Lacking strong author or institutional signals, so rated as emerging/limited info.