List of Papers Browse by Subject Areas Author List
Abstract
Emergency-room (ER) decision-making is characterized by high clinical risk, strict time constraints, and rapidly evolving yet incomplete evidence. Recently, multi-agent systems (MAS) have shown promise for interactive clinical reasoning. However, we identify a key limitation in ER-style workflows: multi-turn dialogues often terminate with insufficient evidence coverage, and similar coverage ratios can still lead to markedly different outcomes when a small set of decision-critical evidence is missing. To make evidence a first-class optimization target, we propose a closed-loop Quantify, Collect, and Correct (QCC) multi-agent framework for emergency reasoning. In the Quantify stage, we introduce the Weighted Evidence Convergent Index (WECI), which importance-weights collected evidence to measure how well MAS capture decision-critical signals during interaction. In the Collect stage, an Evidence-Driven Planning Module leverages current evidence and retrieval-augmented medical knowledge to infer plausible diagnostic directions, plan a small set of prioritized missing evidence targets for subsequent turns, while the Doctor autonomously determines when the accumulated evidence is sufficient to conclude the consultation. In the Correct stage, we perform evidence-driven inference and apply evidence–diagnosis cross-checking to detect inconsistencies and revise conclusions when necessary. Experiments on two ER datasets and one general diagnosis dataset demonstrate that QCC improves emergency decision quality and robustness under noisy and insufficient evidence, consistently outperforming strong MAS baselines across multiple backbone LLMs and evaluation settings.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4233_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HaoyuCao/QCC_MAS
Link to the Dataset(s)
AgentClinic dataset: https://agentclinic.github.io/
ER-Reason dataset: https://physionet.org/content/er-reason/1.0.0/
MIMIC-IV-ED dataset: https://physionet.org/content/mimic-iv-ed/2.2/
BibTex
@InProceedings{CaoHao_Quantify_MICCAI2026,
author = { Cao, Haoyu AND Fu, Yueze AND Gao, Wenbo AND Lin, Yuxin AND Li, Xiangyu AND Wang, Wei},
title = { { Quantify, Collect, and Correct: Evidence-Driven Multi-agent Framework for Critical Evidence Refinement in Emergency Rooms } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Propose the QCC closed-loop multi-agent framework that takes evidence as a first-class optimization target for ER clinical reasoning, addressing insufficient evidence collection and missing critical evidence in existing MAS.
Invent the Weighted Evidence Convergent Index (WECI) to measure the collection quality of decision-critical evidence, outperforming equal-weight coverage metrics.
Design an Evidence-Driven Planning Module (EDPM) to guide targeted collection of Prospective Primary Evidence (PPE) via specialist agents.
Build an Evidence-Based Verifier Module (EBVM) to cross-check evidence-diagnosis consistency and revise biased conclusions.
Validate on 1 general and 2 real-world ER datasets, outperforming mainstream MAS baselines under noisy/insufficient evidence.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Prioritizes evidence optimization for ER reasoning, and WECI solves the flaw that traditional metrics ignore evidence importance, matching clinical reality.
The Quantify-Collect-Correct cycle fixes unguided evidence accumulation and diagnostic inconsistency in ReAct/REFINE.
Specialist-guided planning conforms to real ER diagnostic processes, adapting to time-limited, fragmented evidence scenarios.
Validated on diverse datasets with multiple LLMs; ablation confirms EDPM and EBVM are core to performance gains.
Directly addresses key bottlenecks of existing MAS in emergency decision-making.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Only tested on public datasets, no real-world ER deployment or clinician evaluation.
Vague descriptions of evidence weight calculation, PPE generation, and specialist agent assignment reduce reproducibility.
Does not compare with clinical-specific MAS like MedAgents .
Only tested on diagnosis tasks, no verification in other clinical scenarios.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper targets a critical ER clinical pain point, proposes a practically useful QCC framework with valid empirical results, and aligns with real clinical needs. However, it lacks disruptive innovation, has incomplete clinical validation, and ambiguous technical details. It is marginally acceptable, and acceptance depends on rebuttal supplements of technical details, real clinical tests, and extended baselines.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes QCC (Quantify, Collect, Correct) for multi-agent ER diagnosis. The main idea is to focus on getting the right evidence, not only on giving a final answer. The framework has three parts: WECI to score how important the collected evidence is, a planning module to guide what to ask next, and a verifier to check whether the final diagnosis is really supported by the evidence. The method is evaluated on AgentClinic, MIMIC-IV-ED, and ER-REASON.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper studies an important problem. In interactive diagnosis, especially in emergency settings, collecting the right evidence is critical. This is also consistent with recent work such as AMIE, which emphasizes diagnostic dialogue and information gathering, and AgentClinic, which shows that sequential clinical decision-making is much harder than static QA.
Another strength is that the paper has a clear and reasonable message: missing a few key findings can matter more than collecting many less useful facts. The experiments are also fairly solid for a conference paper. The method is tested on both a general benchmark and two ER datasets, and the ablations support the value of the planning and verifier modules.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
My main concern is that the novelty is limited. The paper is well organized, but many of the ideas are not fully new. Recent work like REFINE already argues that strong reasoning alone is not enough, uses evidence-based evaluation, and adds diagnostic verification to guide further evidence collection. Because of that, QCC feels more like a good combination and adaptation of existing ideas than a clearly new technical direction.
A second issue is the evaluation setup. The paper says specialist clinicians curate the evidence, but it also relies on external LLMs to split records into atomic evidence and to assign importance weights. Since the proposed metric depends on these steps, this part should be validated more carefully. Finally, some method details are still unclear, especially how the verifier improves evidence-collection metrics if it is only applied after the consultation ends.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please refer to my previous comments.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes QCC, an evidence-driven multi-agent framework for emergency-room reasoning that explicitly treats evidence as a first-class optimization target. The framework consists of three stages: Quantify, which introduces the Weighted Evidence Convergent Index (WECI) to measure how well collected evidence covers decision-critical signals; Collect, which uses an Evidence-Driven Planning Module (EDPM) to guide subsequent information gathering toward prioritized missing evidence; and Correct, which applies an Evidence-Based Verifier Module (EBVM) to check evidence–diagnosis consistency and revise conclusions when necessary. The paper evaluates QCC on one general diagnosis benchmark and two emergency-room datasets and reports improved evidence quality and diagnostic accuracy over several interactive multi-agent baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper addresses an important and clinically appropriate problem. In emergency settings, evidence is incomplete, time-sensitive, and highly unequal in diagnostic value, so shifting the focus from generic multi-agent dialogue to evidence-efficient reasoning is well motivated. 2.The paper is conceptually coherent. The Quantify–Collect–Correct decomposition is intuitive, and the three modules are meaningfully connected rather than being a loose collection of ideas. 3.WECI is an interesting contribution. Standard evidence coverage metrics often treat all evidence equally, whereas this paper attempts to distinguish decision-critical evidence from low-value evidence, which is particularly relevant in emergency medicine. 4.The Evidence-Driven Planning Module is practically motivated. The idea of periodically steering questioning toward prospective primary evidence targets is sensible and aligns with how clinicians refine hypotheses under time pressure. 5.The verifier module adds a useful consistency layer. It is valuable that the framework does not stop at collecting more evidence, but also checks whether the final diagnosis is consistent with the collected evidence and planning history. 6.The experimental evaluation is reasonably broad. The paper includes one general diagnosis benchmark and two emergency-related datasets, which helps show that the framework is not restricted to one narrow setting. 7.The ablation study is informative and supports the contribution of both planning and verification. 8.The manuscript is generally clear, and the framework figures are helpful.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The novelty should be positioned more carefully. Multi-agent clinical reasoning, interactive evidence elicitation, and self-refinement or verifier-style correction have already been explored in prior work such as MedAgents [19], conversational diagnostic AI [21], Self-Refine [12], ReAct [27], multi-agent debate [3], and recent evidence-elicitation evaluation work [10]. The current contribution is best understood as an evidence-centric integration and refinement of these directions for emergency reasoning, rather than a fundamentally new paradigm. 2.The reliance on external LLMs is substantial. In addition to the main backbone models, the framework uses extra LLMs for evidence structuring, evidence weighting, and verification-related components. This makes it harder to separate the contribution of the proposed framework design from the contribution of stronger auxiliary model support. 3.The real-world evaluation is still relatively limited in scale. The paper samples 200 cases from each data source to balance coverage and cost, which is understandable, but this remains a modest sample size for heterogeneous emergency reasoning settings. 4.The comparison to CoT as an upper bound is useful, but the interpretation should be cautious. Since CoT observes all evidence at once, it is not directly comparable to interactive systems in terms of deployment assumptions; this comparison is best viewed as a loose ceiling rather than a practically attainable reference. 5.The framework claims improved robustness under noisy and insufficient evidence, but robustness is mainly inferred from benchmark performance rather than being stress-tested through explicit perturbation analyses or controlled evidence removal experiments. 6.Some of the gains may come from increased system complexity rather than the specific QCC design. Because QCC adds both a planning module and a verifier, a stronger decomposition against simpler evidence-guided single-agent or single-pass alternatives would help isolate what is essential.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
I appreciate the authors for focusing on evidence as the central object of optimization in interactive clinical reasoning. This is an important shift in perspective. In emergency medicine, simply producing the right answer is often less informative than knowing whether the system collected the right evidence efficiently and coherently, and I think the paper makes this point well. I also found the framework design thoughtful. The three-stage Quantify–Collect–Correct decomposition is clinically intuitive, and the idea of periodically planning toward prospective primary evidence targets is a meaningful attempt to model how focused diagnostic questioning actually evolves in practice.
My main suggestions are intended to strengthen what is already a promising submission. First, I encourage the authors to position the work more explicitly relative to prior evidence-elicitation and multi-agent refinement literature, emphasizing that the novelty lies in the evidence-centric formulation and its emergency-room adaptation. Second, the paper would benefit from more discussion of the limitations of WECI, especially its dependence on an external LLM judge and ground-truth-conditioned weighting. Third, stronger robustness analysis—such as controlled evidence omission or noise injection—would make the evidence-oriented claims even more convincing. Finally, a more detailed qualitative error analysis would help readers better understand where the planner and verifier contribute most. Overall, I think this is a well-motivated and relatively mature paper, with a clear practical angle and solid empirical evidence.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on the paper’s strong problem formulation, coherent framework design, and consistent empirical gains, balanced against some limitations in evaluation methodology and novelty positioning. I think the paper addresses an important gap in current multi-agent clinical reasoning work. Rather than treating evidence as generic context, it explicitly models the quality and directional usefulness of collected evidence, which is especially relevant in emergency settings. The proposed QCC framework is logically structured, the ablation results support the contribution of both planning and verification, and the improvements across the reported datasets are meaningful. My main reservations concern the evaluation setup and scope of claims. In particular, the WECI metric depends on an external LLM-based weighting process, the framework relies on multiple auxiliary LLM components, and the real-world emergency evaluation remains relatively limited in scale. I also think the novelty should be described more carefully relative to existing multi-agent reasoning, evidence elicitation, and self-correction literature. Despite these concerns, I believe the paper makes a useful and timely contribution. It is more clinically grounded than many generic medical agent papers, and the evidence-centric perspective is valuable. For these reasons, I place it slightly above the acceptance threshold.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all reviewers and the meta-reviewer for their positive and constructive assessment, as well as for their time and valuable suggestions. We are encouraged that the reviewers consistently recognized the clinical importance of evidence-efficient ER reasoning, the coherence of the Quantify–Collect–Correct (QCC) design, the usefulness of WECI, and the empirical gains across AgentClinic, MIMIC-IV-ED, and ER-REASON.
We particularly appreciate the meta-reviewer’s suggestion to clarify the novelty positioning, which we believe is accurate and constructive. We fully agree that the contribution of QCC should be positioned carefully, and we will more clearly articulate the novelty positioning of the QCC framework in the camera-ready version. We would like to emphasize that the core contribution of QCC is not merely adding another multi-agent workflow, but shifting the optimization target of interactive clinical reasoning. Existing systems are often optimized or evaluated mainly by final diagnostic accuracy, while evidence is treated as passive dialogue context. In contrast, QCC moves the target from answer-only optimization to evidence-oriented optimization, and further to decision-critical evidence acquisition and evidence-chain consistency. Therefore, driven by an evidence-oriented optimization objective, QCC is not designed to indiscriminately increase the amount of collected evidence, but to prioritize the acquisition of more decision-critical and clinically relevant evidence under the same evidence budget.
We also thank the reviewers for their comments on the evaluation scale. We agree that sampling 200 cases from each data source cannot cover the full heterogeneity of real-world emergency medicine. This design was chosen as a practical trade-off among evaluation fairness, dataset coverage, and experimental cost. More importantly, we believe that the proposed optimization targets, namely decision-critical evidence acquisition and evidence-chain consistency, are testable and extensible beyond the current evaluation samples and datasets. As a future direction, we aim to extend QCC from a static diagnostic framework toward a self-evolving clinical reasoning system based on long-term memory and timely feedback, where evidence weights, planning policies, and verification criteria can be continuously updated with incoming cases and clinician feedback to better adapt to evolving clinical distributions and real-world emergency scenarios.
Overall, we believe that the evidence-centric perspective of QCC is timely and highly relevant to the MICCAI community. We hope this perspective can encourage future medical AI systems to move beyond final-answer accuracy toward safer, more interpretable, and more clinically aligned decision-making.
We once again thank all reviewers and the meta-reviewer for their valuable suggestions, which will help us further refine the work, and we sincerely appreciate their consistent recognition of the clinical relevance and potential value of QCC.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The reviewers’ comments are positive and quite aligned. They agreed on the topic’s importance, the validation pipeline and the paper clarity. However, they also raised some concerns about the novelty positioning. It would be nice if the authors could state the novelty of the paper more clearly in the camera-ready version.