Multi-Agent Clinical Curation Framework

• Multi-Agent Clinical Curation Framework is a modular AI system that distributes evidence gathering and curation among specialized agents for dynamic clinical applications.
• It leverages explicit protocols and modular workflows, such as consultation flows and directed acyclic graphs, to integrate heterogeneous clinical data into unified representations.
• The framework aims to enhance process fidelity and transparency while addressing challenges in inter-agent communication and validation in clinical settings.

Searching arXiv for the cited frameworks to ground the synthesis in current literature. Multi-agent clinical curation frameworks are clinical AI systems that distribute evidence gathering, interpretation, critique, and synthesis across specialized agents rather than assigning the entire task to a single model. In the recent literature, these frameworks appear in dynamic diagnosis, ICU decision support, patient-level EHR question answering, multidisciplinary oncology deliberation, qualitative analysis of clinical interviews, clinical trial matching, radiology report generation, biomarker discovery, clinical coding, and personalized nutrition planning. Their shared premise is that heterogeneous clinical evidence and multi-stage workflows are better handled by role-specific components coordinated through explicit protocols such as consultation flow, directed acyclic graphs, argumentation schemes, retrieval pipelines, consensus procedures, or reinforcement learning controllers (Wang et al., 19 Mar 2025, Chen et al., 25 Mar 2025, Qian et al., 15 Jan 2026, Zhang et al., 9 Dec 2025, Wu et al., 20 Feb 2026).

1. Conceptual foundations

An important precursor is the ontologies-based clinical decision support literature, in which a Supervised Multi-Agent system based on Ontologies integrates heterogeneous clinical knowledge bases, including stochastic models, expert heuristics, deductive reasoning, and case-based reasoning. In that formulation, a Supervisory Agent (TAS) orchestrates workflow transitions, a General Clinical Agent Type (TACG) provides common agent functions, Knowledge Model Agents (TAMC) specialize by reasoning model, Ontology Agents maintain domain semantics, and workflow control is encoded through a finite state automaton (AEFCG) and per-step AEFS structures. The explicit objective is semantic interoperability and modularity across diagnosis, prognosis, therapy, and follow-up (Shen et al., 2020).

Recent LLM-centered systems inherit this modular logic but shift the emphasis from ontology-mediated interoperability to workflow-grounded orchestration. The diagnosis framework in "Empowering Medical Multi-Agents with Clinical Consultation Flow for Dynamic Diagnosis" organizes interaction around Inquiry Phase, Examination Phase, and Diagnosis Phase, each with a sub-goal $g_k$. MedCollab emulates the hierarchical consultation workflow of modern hospitals through a GP agent, dynamic specialist recruitment, IBIS-Structured Argumentation Protocol, and a Hierarchical Disease Causal Chain. The ICU clinical decision support system in "Reinforcing Clinical Decision Support through Multi-Agent Systems and Ethical AI Governance" uses a modular agent stack linked by an asynchronous Directed Acyclic Graph (DAG). These formulations collectively place curation inside an explicitly staged clinical process rather than a single inference pass (Wang et al., 19 Mar 2025, Zhan et al., 1 Mar 2026, Chen et al., 25 Mar 2025).

2. Agent specialization and organizational forms

The literature reports several recurring organizational forms: consultation teams, modality-specific processing chains, supervisor–worker hierarchies, and human-in-the-loop review loops. In each case, the agent boundary is not merely computational; it corresponds to a clinical role, an evidence type, or a governance function.

Framework	Specialized agents	Coordination form
Dynamic diagnosis	Doctor, Patient, Examiner	Consultation flow with inquiry, examination, diagnosis
ICU decision support	Lab, Vitals, Context, Integration, Prediction, Validation, Transparency	Asynchronous DAG
ColaCare	DoctorAgents, MetaAgent	MDT-style consultation and debate
EHRNavigator	Table Reviewer, SQL Writer, Note Retriever, Answer Synthesizer	Module orchestration over structured and unstructured EHR
TAMA	Generation, Evaluation, Refinement, Human/Cardiac Expert	Iterative human-in-the-loop thematic analysis
TriAgent	Scoping, Data Analysis, Research Supervisor, Research Sub-Agents, Reporting	Graph-based multi-agent deep research
MedCollab	GP, clinical specialists, examination agents	Dynamic recruitment and multi-round consensus

In diagnosis-oriented systems, the role structure frequently mirrors clinical labor division. The consultation-flow framework assigns decision authority to the Doctor Agent, subjective symptom reporting to the Patient Agent, and objective multi-modal results to the Examiner Agent. MedCollab similarly separates a GP agent from specialist and examination agents, while ClinicalAgents defines an Agent Pool with a Triage Nurse, Specialist Doctor, Radiologist, and Attending Physician under a central Clinical Orchestrator. In radiology, MARCH reproduces a departmental hierarchy through a Resident Agent, multiple Fellow Agents, and an Attending Agent (Wang et al., 19 Mar 2025, Zhan et al., 1 Mar 2026, Ge et al., 27 Mar 2026, Lin et al., 17 Apr 2026).

In non-diagnostic curation tasks, specialization is often modality- or function-centered. EHRNavigator divides work among structured query, note retrieval, and multimodal synthesis components. ColaCare binds each DoctorAgent to a domain-specific expert model such as AdaCare, ConCare, or SAFARI, with a MetaAgent synthesizing deliberation. TAMA assigns one agent to theme generation, one to evaluation, and one to refinement, while retaining a domain expert as the stopping authority. TriAgent separates scoping, data analysis, literature-grounded validation, and reporting, thereby treating curation as a coupled analytic and evidence-review process rather than as direct prediction alone (Qian et al., 15 Jan 2026, Wang et al., 2024, Xu et al., 26 Mar 2025, Delikoyun et al., 17 Oct 2025).

3. Evidence curation, multimodality, and memory

A central property of these frameworks is the transformation of heterogeneous inputs into curated intermediate representations that can be shared, audited, or reused. In the consultation-flow diagnosis framework, all multi-modal information flows are eventually made into text, so that LLMs can operate over symptoms, laboratory data, imaging findings, and ECG results in a single interpretable channel. The Examiner Agent is explicitly responsible for transforming objective results into textual form for the doctor. EHRNavigator uses a different decomposition: table descriptions are generated by a Table Reviewer Agent, relevant tables are selected through semantic similarity, notes are chunked and indexed, and an Answer Synthesizer Agent combines $E_{tab}$ and $E_{note}$ under temporal constraints (Wang et al., 19 Mar 2025, Qian et al., 15 Jan 2026).

Several systems add explicit provenance structures. MedMMV constructs an evidence graph whose nodes are atomic facts and whose edges are type-annotated relations with provenance. Claims in candidate reasoning paths must be grounded in evidence subgraphs, and the graph is used for claim verification and conflict detection. MAKAR treats augmentation itself as a curation stage: a Knowledge Probing Agent identifies insufficiency in trial criteria, a Navigation Agent chooses an augmentation path, augmentation agents enrich the criterion, and a Supervision Agent checks alignment with the original. Hybrid-Code restricts accepted outputs to a 257-code knowledge base and requires evidence validation from source text before code acceptance, making curation inseparable from symbolic verification (Liu et al., 29 Sep 2025, Shi et al., 2024, Yu, 26 Dec 2025).

Memory design is another recurring mechanism. ClinicalAgents separates mutable Working Memory $$\mathcal{M}_\text{work} = \langle \mathcal{E}_t,\mathcal{H}_t,\tau \rangle$$ from static Experience Memory $\mathcal{M}_\text{exp}$, which stores guidelines and historical cases retrievable via hybrid RAG. MedChain-Agent similarly introduces MedCase-RAG, storing cases as 12-dimensional feature vectors and retrieving top-3 similar cases from a department database. MedAgentSim maintains a Medical Records Buffer for correctly diagnosed cases and an Experience Records Buffer for corrected misdiagnoses, using KNN retrieval and experience replay. This suggests that curation is increasingly treated as a longitudinal process over both current patient evidence and archived clinical experience (Ge et al., 27 Mar 2026, Liu et al., 2024, Almansoori et al., 28 Mar 2025).

4. Control, supervision, and learning mechanisms

The most technically distinctive feature of recent frameworks is the move from static prompting to explicit control policies. In dynamic diagnosis, the multi-agent interaction is modeled as a Partially Observable Markov Decision Game (POMG),

$(n,S,A,O,T,G,R),$

with $n=3$, hierarchical phase-dependent action space $A$, observations $O$, sub-goals $G=\{g_1,\ldots,g_k\}$, and reward $E_{tab}$0. The doctor selects actions to maximize expected cumulative reward,

$E_{tab}$1

while a hierarchical action set constrains decisions through progressive actions and retrospective actions, thereby reducing premature closure and regularizing stage transitions (Wang et al., 19 Mar 2025).

Other systems implement control through orchestration and repair. ClinicalAgents models non-linear reasoning as a Monte Carlo Tree Search (MCTS) process with actions for agent activation, retrieval, backtracking, and termination, and uses dense reward shaping based on information gain and confidence gain. MedMMV stabilizes reasoning through diversified short rollouts, a Hallucination Detector (HD Supervisor), stepwise AutoRepair, and a Combined Uncertainty (CU) Scorer,

$E_{tab}$2

The counterfactual diagnostic framework introduces Counterfactual Case Editing and the Counterfactual Probability Gap,

$E_{tab}$3

to quantify how strongly an individual finding supports a diagnosis and to guide multi-round specialist discussion (Ge et al., 27 Mar 2026, Liu et al., 29 Sep 2025, You et al., 29 Mar 2026).

Reinforcement learning appears in both dialogue and process-supervised curation. DoctorAgent-RL formalizes consultation as an MDP in which the doctor chooses between $E_{tab}$4 and $E_{tab}$5, while a Consultation Evaluator returns a composite reward

$E_{tab}$6

The gene–disease validity curation framework uses Group Relative Policy Optimization (GRPO) to train a supervisor agent under outcome rewards or hybrid process-plus-outcome rewards. The reported distinction is substantive: outcome-only optimization can raise final accuracy while degrading process alignment, whereas process-level supervision improves process fidelity without sacrificing final classification performance (Feng et al., 26 May 2025, Lee et al., 15 Feb 2026).

5. Evaluation regimes and empirical record

Evaluation protocols vary with task, but the literature consistently measures not only final correctness but also intermediate process quality. On the MVME benchmark of 506 multimodal diagnostic cases, the consultation-flow framework reports that the mode of interaction turns increases from 4-5 to ~8, and that Ours-Qwen-Max reaches F1 40.25 against a baseline of 29.56 in entity matching. The same paper reports higher and more consistent scores across symptoms, examinations, diagnostic results, diagnostic rationales, and treatment plan (Wang et al., 19 Mar 2025).

Process-centered evaluations frequently expose benefits not visible in end metrics alone. In TAMA, end-to-end thematic analysis is completed in under 10 minutes rather than 30 hours manually, with HIT rate improving from 83% to 92% and Jaccard decreasing from 42% to 29%, indicating more distinct themes after refinement. In the process-supervised gene–disease curation system, a GRPO-trained supervisor with outcome-only rewards raises outcome accuracy from 0.195 to 0.732 but leaves process fidelity at 0.392 F1, whereas process + outcome rewards yield 0.750 outcome accuracy and 0.520 F1 process fidelity. These results make process metrics an explicit part of system quality rather than a secondary explanation layer (Xu et al., 26 Mar 2025, Lee et al., 15 Feb 2026).

Clinical deployment studies use broader operational criteria. EHRNavigator reports 86% accuracy on real-world cases with Median clinical response time: ~12.16 seconds on YNHHQA, including Labs 87.5%, Drugs 80%, and Combination 95%. MedCollab reports 76.9% diagnosis accuracy on ClinicalBench versus ClinicalAgent’s 68.7%, and an ablation in which disabling logic auditing drops accuracy from 76.9% to 49.7%. NutriOrion reports a 12.1 percent drug-food interaction violation rate, 97.8% actionability, 81.4% overall clinical acceptability, and biomarker–nutrient correlations from -0.26 to -0.35. Hybrid-Code reports no hallucinated codes among accepted outputs within the knowledge base, 24.47% verification rate, 34.11% coverage (95% CI: 31.2%--37.0%), and 86%+ LLM utilization (Qian et al., 15 Jan 2026, Zhan et al., 1 Mar 2026, Wu et al., 20 Feb 2026, Yu, 26 Dec 2025).

6. Limitations, tensions, and future directions

A recurrent misconception is that multi-agent decomposition automatically improves transparency or predictive quality. The ICU decision support study explicitly reports that prediction performance between MAS and single-agent are nearly identical, with Mortality Accuracy (%) of 85.54 versus 85.52 and LOS Mean Error (days) of 0.7491 for both, while Average Transparency Score (%) is lower for the multi-agent system (60.01) than for the single-agent system (75.04). The paper attributes this in part to the fact that, without explicit inter-agent communication, internal decision-making may become more distributed and opaque (Chen et al., 25 Mar 2025).

Other limitations arise from knowledge coverage, validation scope, and workflow granularity. Hybrid-Code’s precision depends on a restricted 257-code knowledge base, which the paper identifies as a coverage bottleneck. TriAgent reports improvements in biomarker justification and literature-grounded novelty assessment but states that it is not yet performing prospective clinical validation. The gastrointestinal oncology MDT framework identifies future plans for direct image processing for radiology and dynamic agent querying. The gene–disease curation work notes that the current MAS is single-turn, that full multi-agent rollouts are expensive, and that sub-agents are frozen in the main pipeline (Yu, 26 Dec 2025, Delikoyun et al., 17 Oct 2025, Zhang et al., 9 Dec 2025, Lee et al., 15 Feb 2026).

The trajectory of the field, as suggested by these systems, is toward tighter coupling of specialization, process supervision, and clinical interoperability. NutriOrion already maps outputs into ADIME and FHIR R4 resources; MedMMV and MedCollab foreground auditable reasoning traces; ClinicalAgents and MedChain-Agent make backtracking and feedback explicit; and counterfactual multi-agent reasoning turns evidence verification into a first-class operation. A plausible implication is that future clinical curation frameworks will be judged less by isolated end-task accuracy than by their ability to expose evidence provenance, preserve process fidelity, and remain compatible with institutional workflows, chart review, and regulatory scrutiny (Wu et al., 20 Feb 2026, Liu et al., 29 Sep 2025, Ge et al., 27 Mar 2026, Liu et al., 2024, You et al., 29 Mar 2026).