List of Papers Browse by Subject Areas Author List
Abstract
Real-world clinical care is a long-horizon decision-making process in which patient state and evidence evolve over multiple turns. In contrast, static, single-turn medical VQA is insufficient for training and evaluating interactive medical agents. To bridge this gap, we introduce DiagWorld, a scalable multimodal virtual medical environment based on large-scale real-world clinical interactions—an EHR-grounded simulator that supports multi-turn interviewing and tool-augmented evidence acquisition under realistic clinical workflows. Built from real patient records in MIMIC-IV, our pipeline transforms longitudinal EHR data—admissions, laboratory tests, imaging, and clinical notes—into 500K+ interactive trajectories, enabling large-scale training and evaluation. Unlike prior EHR-only simulators, we integrate external disease-level medical databases for consistency-aware completion and constraint validation, reducing diagnostic confounding from missing or noisy EHR signals and enabling scaling in patient volume and interaction horizon without amplifying artifacts. Leveraging this environment, we train a diagnostic agent with confidence-based process supervision: rather than sparse terminal rewards or manually annotated key steps, we provide intermediate rewards tied to increases in confidence toward the correct diagnosis, reinforcing informative questioning, retrieval, test ordering, and CXR analysis. This shaping improves sample efficiency and stabilizes learning in large-scale interactive training.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5084_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{FanZhi_MedEnv_MICCAI2026,
author = { Fan, Zhiting AND Ren, Xuxiang AND Wang, Yuan AND Chen, Ruizhe AND Meng, Zijie AND Hu, Keli AND Liu, Jiaxiang AND Wu, Jian AND Zhai, Weiqi AND Xu, Hongxia AND Liu, Zuozhu},
title = { { MedEnv: Scaling Multimodal Virtual Medical Environments for Long-Horizon Diagnosis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16886},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This is a paper about building a virtual medical environment containing long-term electronic health records of 496K patients, for help in training and diagnosis. It seems to model a physicians approach where the expert looks at the long history, current data, and develops a diagnosis over a period. The main accomplishment here seems to be take long-term data (EHR) and use them in an interactive environment. From the table, it looks like one of the largest environment for multimodal data analysis. It incorporates external medical databases to provide consistency and validation, reducing issues from missing or noisy EHR signals and enabling longer interaction horizons. It thus provides a large-scale dataset for training and evaluating medical agents.
- 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.
In my opinion, these are the four main contributions: (1) use of multimodal environment, (2) interactive environment using EHR, current data, and multiple events, (3) linking to external disease database to reduce noise and clutter, and (4) large scale – it seems to be the largest medical environment of its kind.
- 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.
This is not my field so I cannot think of a major weakness. The paper is not developing any machine learning or AI methodology for diagnosis. Rather it is developing an environment for incorporating as much data as possible.
- 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
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?
This paper is far outside of my expertise. If any other reviewer has a strong opinion, one way or another, I am happy to go with their evaluation.
- Reviewer confidence
Not confident (1)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
This paper is far outside my areas of expertise. Based on my minimal knowlegde I recommend accepting the paper.
Review #2
- Please describe the contribution of the paper
The paper introduce, MedEnv, a large-scale multimodal multi-turn medical environment constructed from MIMIC-IV EHR and MIMIC-CXR, augmented with a disease-level knowledge base to mitigate missingness and support long-horizon, tool-augmented diagnostic interactions. The authors further propose a reinforcement learning approach with a confidence-based process reward (Diagnosis-R1) in addition to ORM, to encourage informative questioning, reporting 41.95% diagnostic accuracy and improved interaction efficiency versus strong 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 introduces MedEnv, a large-scale (496K+ trajectories) multimodal, multi-turn medical environment built on real-world EHR data (MIMIC-IV), which is a significant scale-up compared to closest alternative. 2) I think using process reward rather than the ORM based reward is an apt choice.
- 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.
Weakness and Questions:
1) How do trained models fare on other benchmarks - specifically AgentClinic? Also how does MedEnv compare with AgentClinic? 2) How do you balance correctness and information to gain rewards i.e. what is the value of various lambda you have used ? 3) It is not clear how LLM as a judge is used? How does it take into account synonyms and abbreviations? Can this be replace this with a deterministic ICD mapping. 4) Implementation details are missing - Can you explain hyper-parameters for the GRPO like group size and lr for Qwen. 5) Authors claim that “Overall, adding images yields a modest gain in accuracy and more consistently improves interaction efficiency, typically increasing Efficient Turns while reducing Avg. Turns, indicating that visual evidence helps models ask more targeted questions and converge with fewer uninformative turns.” However from table 2, for Diagnosis-R1 with image accuracy is 41.15 as compared to without image - 41.95, and a similar pattern is shown for Effective turns reducing from 80.99 → 77.87 (assuming higher is better ) and 10.58 → 11.23 (assuming lower is better). Can you explain what the reason is? 6) Is there a category based analysis to identify which modality/disease are easy to solve and which are difficult both in terms of accuracy and number of turns. 7) Why is a doctor agent trained on only ~3K trajectories? Also what is train-val-test split?
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I really liked the problem and moving forward this will be an important problem to solve. However, I don’t think paper in its current form is publishable. It would require major changes to clearly state and validate the contributions.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Thank you for the detailed rebuttal. The construction and scale of the MedEnv dataset alone represent a highly valuable contribution to the community. For the final manuscript, I strongly advise the authors to recalibrate their claims regarding the imaging modality, ensuring they accurately reflect the results discussed in the rebuttal. Additionally, please ensure that all training configurations and GRPO hyperparameters are thoroughly documented to guarantee experimental reproducibility.
Review #3
- Please describe the contribution of the paper
This paper proposes MedEnv, a large-scale virtual medical environment that simulates realistic, multi-step clinical diagnosis using real patient data and multiple data types (e.g., text and images) . It allows AI agents to interact with patients, ask questions, and use tools over multiple turns, which better reflects real clinical workflows. The authors also introduce a new training method that rewards the agent for making useful steps toward the correct diagnosis. Together, the environment and training approach improve diagnostic accuracy and efficiency, and provide a useful benchmark for developing interactive medical AI systems.
- 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 introduces MedEnv, a novel multimodal, multi-turn medical environment that better reflects real clinical diagnosis compared to prior static or single-turn benchmarks
2.It constructs a large-scale dataset (496K+ trajectories) from real-world EHR data, which is significantly larger and more realistic than existing medical agent benchmarks.
3.The environment integrates EHR data with a disease-level knowledge base, providing an original way to handle missing or incomplete clinical information.
3.It supports tool-augmented interactions (e.g., retrieval, imaging), enabling agents to simulate realistic clinical workflows such as ordering tests and interpreting results.
4.The proposed confidence-based process reward is a novel training strategy that gives dense feedback by rewarding informative steps during multi-turn diagnosis.
- 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 process reward based on log-probability of the correct diagnosis depends on exact label matching, but the paper does not provide enough details on how to deal with synonym normalization or ICD mapping.
2.The evaluation relies on an LLM-as-judge instead of deterministic label matching, which introduces subjectivity and is weaker than standard clinical evaluation practices used in prior work.
3.There is no external validation or clinician evaluation, limiting evidence for real-world clinical feasibility.
4.The imaging component is relatively weak and not clearly novel, as it relies on generic grounding tools (e.g., Grounding DINO) rather than specialized medical vision models used in recent radiology studies.
- 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
The paper presents an interesting framework, but several aspects would benefit from further clarification and validation. First, the proposed process reward relies on the log-probability of the correct diagnosis, yet the paper does not provide sufficient details on how diagnosis labels are normalized (e.g., handling synonyms or mapping to ICD codes). Second, it is unclear why an LLM-as-judge is used to evaluate diagnostic correctness when ground-truth labels are available. Finally, the study lacks external validation or clinician-based evaluation, which limits the evidence for real-world clinical feasibility and applicability of the proposed system.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses an important gap in medical AI by introducing a scalable, multimodal environment for long-horizon clinical diagnosis. The proposed MedEnv framework is well-motivated and enables realistic multi-turn interactions with tool use, and the confidence-based process reward provides a reasonable approach to improving step-by-step diagnostic reasoning. The experimental results show consistent gains in both accuracy and interaction efficiency over strong baselines, and the large-scale dataset has potential value for the community. Despite some concerns regarding evaluation and implementation details, the overall contribution is meaningful and timely, and the work provides a solid foundation for future research in interactive, agent-based medical AI systems.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the rebuttal, I still support acceptance, although some concerns remain partially unresolved. The authors clarified several important implementation details, including diagnosis canonicalization, reward weighting, GRPO hyperparameters, and the role of the LLM-as-judge, which improves the transparency of the framework. I also appreciate the clearer positioning of the imaging component as an environmental design contribution rather than a novel radiology model. While external clinical validation and stronger deterministic evaluation would further strengthen the work, the rebuttal sufficiently addresses the main concerns for a benchmark/environment paper, and the overall contribution remains meaningful and timely for long-horizon medical agent research.
Author Feedback
We thank the reviewers and AC for the constructive comments. We clarify that MedEnv is not a deployable diagnostic system, but a scalable research environment for training/evaluating long-horizon medical agents. Its contributions are EHR-grounded multi-turn interaction, multimodal/tool-augmented evidence acquisition, KB-based completion/validation to reduce unsupported responses, and confidence-based process supervision. We also use MedEnv to train Diagnosis-R1 with RL, encouraging informative questioning and evidence acquisition.
Compared with AgentClinic, MedEnv is much larger (496K+ trajectories vs. 1.6K cases) and provides a large RL source. Since the two benchmarks differ in case construction, APIs, and label space, direct substitution is not our claim. We will clarify this comparison and include cross-benchmark evaluation if permitted.
Each case is anchored to the principal diagnosis in the clinical record. When available, labels are linked to ICD/disease entries; cues, aliases, and abbreviations are used for disease-level KB matching. It measures whether new evidence increases the model’s log-probability of a fixed canonical target diagnosis. The final reward combines terminal correctness and information gain with weights 1.0 and 0.5.We will better specify canonicalization, synonym/abbreviation handling, ICD mapping, and reward-weight analysis.
We use LLM-as-judge only for final free-text diagnosis evaluation, since correct diagnoses may differ in wording, abbreviation, or granularity. Exact matching would under-credit such cases, while deterministic ICD mapping can also be ambiguous for generated text and may collapse clinically meaningful distinctions. We therefore use a constrained Qwen3-8B judge for binary YES/NO consistency, explicitly considering synonyms, abbreviations, and equivalent clinical terms. We will include the judge prompt, manual-inspection results, and describe ICD matching as complementary validation, not a full replacement for semantic clinical equivalence.
We agree that clinician and external validation are required before clinical translation. Our claims concern simulation realism and scalable agent learning, not clinical deployability. MedEnv is grounded in MIMIC-IV/MIMIC-CXR, uses a pre-treatment observation window with de-leakage, follows an EHR-first -> KB-augmentation -> no-evidence fallback policy, and supports verifiable tool calls. These reduce unsupported patient information, but do not establish clinical safety; we will state this explicitly.
For imaging, we agree the module is not a new radiology model. Its novelty is environmental: CXR is exposed as a callable evidence source in long-horizon diagnosis, so agents must decide when to request, localize, and integrate visual evidence. We will revise the overgeneralized claim about image benefits. Table 2 shows model-dependent effects: images improve several models, but for Diagnosis-R1 they slightly reduce accuracy and efficiency, likely because it already exploits EHR/retrieval strongly, while current CXR-only evidence can introduce noisy or low-yield steps. Grounding DINO is used only for localization, not diagnosis generation; the module is replaceable by specialized radiology models.
Regarding implementation, the ~3K trajectories are only for SFT cold start to learn the interaction format; the main improvement comes from RL in MedEnv after difficulty-aware filtering. The held-out benchmark contains 2K cases sampled across 17 disease categories and is separate from RL training cases. For GRPO on Qwen3-VL-8B, we use group size 8, learning rate 1e-6, batch size 64, PPO mini-batch 32, and max horizon 15.We will add these details, the train/test split, benchmark statistics, and category-wise analysis covering both accuracy and interaction turns.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- 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.
This paper proposes MedEnv, a large-scale multimodal virtual medical environment for long-horizon diagnostic interactions, integrating EHR, imaging, and a confidence-based process reward. Reviewers acknowledge the relevance of the problem, the realistic multi-turn simulation, and the potential value of the dataset and framework.
The authors should focus their response on clarifying and justifying the following points: explain how diagnosis labels are normalized (handling synonyms or ICD mapping), clarify why an LLM-as-judge is used instead of deterministic evaluation, provide more evidence for real-world clinical feasibility including clinician evaluation or external validation, and address questions regarding the contribution and novelty of the imaging component.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper introduces a large-scale multimodal virtual medical environment for long-horizon diagnosis, which reviewers view as a valuable contribution. The rebuttal appears to have resolved key concerns, and all post-rebuttal opinions available are positive. The final version should recalibrate claims about imaging modality use and document training details and hyperparameters for reproducibility.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have clearly clarified the issues raised by the reviewers. The explanations in the rebuttal look reasonable and correct to me.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
MedEnv is ambitious in scale, and the reviewers are convinced by the methodology and the rebuttal. One potential weakness is that the evaluation relies on simulated interactions and LLM-as-judge outcomes, so clinical validity depends on simulator fidelity and scoring quality. The authors shall pay attention to these concerns if they want to publish at a higher platform. Best luck with the future work.