Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Ventilator decision support requires sequential decisions that track evolving physiology and disease trajectories while respecting safety boundaries and clinician specific tuning styles. Rule based approaches rarely generalize personalization, and end to end reinforcement learning or single large language model systems remain difficult to control and audit. We propose the Ventilator Decision Support System (VDSS), a human in the loop multi agent framework that coordinates modular decision components through contract driven structured interfaces and produces traceable evidence for review. VDSS performs online preference adaptation with a contextual bandit, updating clinician specific preferences from the final accepted decision at each adjustment cycle and using them to guide subsequent recommendations. Structured rejection feedback triggers targeted replanning to reduce unproductive iterations and improve interaction stability. Retrospective ICU trajectory replay with expert review indicates higher recommendation acceptability and fewer interaction rounds to reach an acceptable plan, supporting clinically deployable human AI collaboration.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0552_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{LiSij_HumanInTheLoop_MICCAI2026,
        author = { Li, Sijia AND Tan, Xiaoyu AND Wang, Qixing AND Zhao, Weiyi AND Zhan, Chen AND Hao, Teqi AND Wang, Xuemin AND Gu, Lei AND Eils, Roland AND Qiu, Xihe},
        title = { { Human-In-The-Loop Multi-agent Ventilator Decision Support with Contextual Bandit Preference Learning } },
        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

    This paper presents VDSS, a human-in-the-loop multi-agent ventilator decision support system that decomposes ventilator titration into modular large language model-based agents connected via contract-driven structured interfaces, producing safety-checked recommendations with an auditable evidence trail. The system is designed in alignment with established clinical workflows and incorporates a cycle-based accept-or-reject interaction protocol with targeted replanning following rejection. Clinician-specific preference adaptation is performed online using a contextual bandit updated at cycle closure from the final accepted decision. The framework is evaluated through retrospective ICU trajectory replay and expert clinician review, demonstrating improved replay metrics and higher clinician-rated acceptability, safety, and clarity compared with single-model end-to-end generation; ablation analyses indicate contributions from both waveform evidence and preference context.

  • 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) Strong clinical translation positioning for a high-risk clinical AI task: VDSS explicitly prioritizes safety boundaries, auditability, and clinician oversight through a cycle-based human-in-the-loop protocol that mirrors bedside titration practice, incorporating accept-or-reject decisions, bounded interaction rounds, and structured clinical notes. This positioning aligns well with MICCAI Clinical Translation and Clinical AI expectations for workflow impact and user interaction.

    2) Principled system design: The modular multi-agent decomposition, encompassing detection, phase and goal inference, hold-or-adjust gating, planning, safety checking, reflection and replanning, and note generation, constitutes a practical alternative to opaque end-to-end generation and supports traceable reasoning through structured interfaces.

    3) Conservative and appropriate preference adaptation: The contextual bandit is updated exclusively from the final accepted plan in each adjustment cycle, reducing sensitivity to noisy intermediate feedback and maintaining alignment between learned preferences and clinician-confirmed outcomes.

    4) Evaluation encompasses clinically meaningful endpoints beyond prediction error: In addition to replay metrics (MSE, MAE, and R²), the study reports expert clinician ratings across acceptability, safety, clarity, and overall performance, as well as interaction efficiency measures including rejection dynamics and regret. Ablation analyses suggest that both waveform evidence and preference context contribute to improved stability and acceptability.

    5) Cross-backbone consistency: Performance gains observed when running identical backbone models within VDSS relative to single-model baselines suggest that improvements are attributable to the workflow architecture rather than to any specific model selection.

  • 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) Evidence maturity remains retrospective (must address): Validation is based on retrospective trajectory replay and expert review rather than prospective silent-mode deployment or real-time bedside evaluation. This limits claims regarding clinical effectiveness, safety in deployment, and operational reliability in live clinical workflows. The authors should add explicit clarification regarding the current stage of deployment readiness. Suggested wording for considerations: Abstract: soften “clinically deployable” to “potential for clinical translation” or “pre-clinical feasibility. ” Conclusion: replace “effective in practice” with “promising under retrospective evaluation and warranting prospective validation. ” Introduction, Methods, or Experiments: add one explicit sentence clarifying that the evaluation is retrospective and does not constitute a deployment study.

    2) Preference learning dynamics are not directly characterized (strongly recommended): The paper introduces a contextual bandit updated at cycle closure (Eq. 2) and reports fewer interaction rounds and improved clinician ratings with preference context enabled. However, the learning dynamics are not directly analyzed, for example, whether the clinician preference state is stable or convergent, susceptible to drift, or remains clinically appropriate over time. The work would be strengthened if the authors clarified how θd concretely modulates candidate ranking and selection across the pipeline, and provided lightweight diagnostics such as a per-cycle evolution of preference-arm weights for the clinician illustrated in Figure 2, or qualitative examples demonstrating how learned preferences influenced subsequent recommendations. A brief discussion of safeguards against preference drift and the transferability of learned preferences across clinicians and handovers would further improve translation readiness.

    3) Limited statistical support for clinician-rated improvements (strongly recommended): Table 1 reports substantial gains in clinician ratings across acceptability, safety, clarity, and overall performance, but the manuscript does not provide measures of uncertainty such as standard deviation or confidence intervals, inter-rater reliability, or formal statistical testing. This limits the ability to assess the robustness of the observed differences, even where effect sizes appear large. The evaluation would be strengthened if the authors reported dispersion measures (mean ± SD or median with IQR), clarified the rating protocol including the number of clinicians, the rating rubric, and the degree of overlap across cycles, and, where applicable, provided a paired significance test or inter-rater reliability metric. If only a single rater was used, this should be stated explicitly and acknowledged as a limitation.

    4) Runtime and latency concerns (must address): The paper reports an average runtime of approximately 305 seconds per cycle under local inference and identifies waveform analysis as a primary computational bottleneck. This raises practical questions regarding bedside usability and intended workflow integration. The clinical translation framing would be strengthened if the authors clarified the target use scenario, distinguishing real-time bedside use from offline review or training contexts, provided a module-level latency breakdown, and discussed plausible optimization strategies such as output caching, asynchronous waveform processing, or a lightweight waveform module, alongside acceptable latency targets for ICU ventilator titration.

    5) Novelty is primarily systemic rather than algorithmic (optional enhancement): The principal contribution of this paper is a clinically grounded integration of multi-agent orchestration, layered memory, safety checking, and preference-updated human-in-the-loop interaction for ventilator titration. Given the existence of prior work on multi-agent large language model frameworks such as AutoGen, and on long-term memory management for language model agents such as MemGPT, the novelty of VDSS would be more clearly articulated if the authors positioned it explicitly as a ventilator-specific, safety-gated, contract-structured decision-support workflow and directly contrasted it with these general-purpose frameworks. Additionally, clarifying whether “contract-driven” refers to structured software interfaces or to formal assume–guarantee contracts would further sharpen the technical contribution.

  • 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?

    The paper uses retrospective multi-center ICU data. The submission should explicitly state ethics approval or waiver, or the data source, or patient consent procedures where applicable, data de-identification methods, and data governance arrangements, to confirm compliance with human-subjects research and data privacy requirements.

  • 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

    Positioning: The contribution is strongest as a Clinical Translation/CAI system paper. It may help to explicitly clarify the boundaries of responsibility (clinician override, safety constraints).

    Preference learning: Please clarify how the 12 preference categories are instantiated operationally (how they modulate ranking/selection) and provide any available insight into stability (e. g. , qualitative examples of preference state evolution).

    Clinician expert review protocol: Please describe the expert review setup more fully (number of clinicians, expertise, blinding, rating rubric, and whether multiple raters reviewed the same cycles), this can be provided as a supplementary material to provide more context and rigor.

    Reporting: Adding uncertainty measures for ratings and replay metrics (where feasible within space) and clarifying failure cases would strengthen the evidence.

    Latency: Please add brief discussion on realistic bedside latency targets and any planned optimizations (particularly waveform analysis) for future deployment-readiness.

  • 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 recommendation is supported on the basis that the paper addresses a clinically significant and safety-critical problem, ventilator titration in the ICU, through a thoughtfully designed human-in-the-loop architecture that prioritizes auditability, safety gating, and workflow-aligned interaction. The modular multi-agent design is well motivated for high-risk clinical decision support, and cycle-level preference adaptation via a contextual bandit represents a conservative and principled approach to personalizing recommendations while maintaining clinician oversight. The retrospective trajectory replay and expert review indicate meaningful improvements over single-model end-to-end baselines, and ablation analyses confirm that both waveform evidence and preference context contribute to improved recommendation acceptability and stability.

    The principal factors precluding a stronger recommendation concern evidence maturity and reporting clarity rather than the validity of the core approach. First, the evaluation is retrospective, based on trajectory replay and expert review, and the framing in several places implies deployment readiness beyond what the evidence supports. Specifically, moderating the phrase “clinically deployable” in the abstract and “effective in practice” in the conclusion, and adding a brief statement clarifying that this is a pre-clinical rather than a prospective or silent-mode deployment study, would bring the claims into full alignment with the presented evidence.

    Second, the clinician rating analysis would benefit from clearer reporting of uncertainty, including means with standard deviations or confidence intervals, and a more complete description of the expert review protocol, specifying the number of raters, the rating rubric, whether ratings were paired across methods, and any inter-rater agreement measures where multiple raters were involved. These additions would materially strengthen the credibility of the reported superiority claims.

    Third, while the preference-learning mechanism is well motivated, the manuscript provides limited direct characterization of the bandit’s learning dynamics, including the stability or potential drift of clinician preference states and how preference context influences candidate selection over time. A brief diagnostic analysis, such as qualitative examples or a figure illustrating how preference weights evolve across cycles for a representative clinician, together with a short discussion of safeguards against preference drift, would strengthen the translational credibility of the system.

    Finally, the reported runtime of approximately 305 seconds per cycle, with waveform analysis identified as the primary bottleneck, raises questions about bedside usability. A clearer discussion of the intended deployment context, acceptable latency targets, and practical optimization pathways, such as asynchronous waveform processing, output caching, or a computationally lighter waveform module, would enable readers to assess deployment feasibility more concretely.

    In summary, as a Clinical AI and Clinical Translation contribution, the system design and evaluation are above the acceptance threshold. Strengthening the manuscript requires tighter alignment between claims and the retrospective evidence base, clearer reporting of the expert review methodology and associated uncertainty, more explicit characterization of preference adaptation behavior, and a more detailed discussion of latency and deployment constraints.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper introduces the Ventilator Decision Support System (VDSS), a human-in-the-loop multi-agent framework designed for dynamic bedside ventilator titration. Continuous ventilator management requires complex, multi-step adjustments that contribute to clinician workload. While existing rule-based systems struggle to personalize care, monolithic AI models lack the transparency and auditability needed for clinical use. To address this, the authors propose a modular architecture of LLM agents that stores patient history in long-term memory, enables targeted human correction, and learns clinician preferences online via a contextual bandit. VDSS achieves higher recommendation acceptability, requires fewer interaction rounds, and improves trajectory replay accuracy compared to single-LLM 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.

    • Agentic AI is valuable and carries high potential for multi-step complex medical workflows. This paper connects the latest developments in LLMs with a relevant medical domain.

    • Decomposing a monolithic LLM task into a modular multi-agent pipeline reduces brittle dependencies and allows for targeted replanning when a user rejects a proposal.

    • The multi-agent design offers an auditable system, tracing each action and decision.

    • Human-in-the-loop correction is important. Also, the framework analyzes human feedback and determines which stage to continue from, avoiding recomputation of all intermediate results.

    • The evaluation is strong, incorporating both objective trajectory replay metrics (MSE, MAE, R²) and subjective expert clinician reviews (Acceptability, Safety, Clarity).

    • The methodology is robustly validated by comparing the multi-agent framework against single-agent baselines across multiple state-of-the-art LLMs, clearly demonstrating the performance benefits of the proposed system.

  • 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.

    Presentation and Paper Organization

    • The paper uses the term “agent” too liberally. It should explicitly define which components are LLM-based and provide more details on specialized modules (e.g., is the Waveform Analyzer also LLM-based? Is it fine-tuned for waveform images?).

    • Section 2.1 is dense and would benefit from being restructured into a bulleted list for improved readability.

    • Methodological details are improperly mixed into the Results (Section 3). The first paragraph, along with “NoImg”, “Regret”, and a proper introduction of evaluated LLMs, should be moved to the Methods section.

    • The single-agent baseline is not described. The authors must clarify in Methods the input-output structure (e.g., whether all data is concatenated and fed into one LLM for a single-shot decision).

    • The Methods should further include implementation details like chosen agent framework (e.g., LangChain, OpenAI SDK) and LLM hyperparameters like temperature.

    • Section 4 (Case Study) adds little value, and Figure 3 is unreadable. This space could be better utilized to include these methodological details.

    Experiment Design and Clinical Viability

    • The introduction motivates reducing clinical variability, yet the contextual bandit inherently adapts to individual preferences. Doesn’t this risk overfitting to specific clinicians? Could universal ICU guidelines be injected to enforce a little stricter clinical boundaries? Like a soft spot between rules and personalization.

    • The reported runtime is 305 seconds per cycle. How long does a clinician typically require? Including this comparison would improve the paper’s added value.

    • Also, could a human clinician be included as a baseline in Table 1 to better ground the framework?

    • Regarding the contextual bandit: was the model actively learning during the evaluation cycles, or were there prior “warm-up” training cycles to establish baseline preferences?

    • The system has 7% failure rate. This could be discussed more in the paper. What were the pitfalls?

    • The pipeline relies on commercial LLM APIs (GPT, Gemini). Does this expose sensitive clinical data and waveform images to third-party vendors? Are there no privacy concerns?

    • Could the authors use the public MIMIC dataset for better reproducibility?

  • Please rate the clarity and organization of this paper

    Poor

  • 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?

    The paper introduces a highly relevant multi-agent framework for ICU ventilator management and successfully demonstrates the value of human-in-the-loop agentic AI in a complex clinical workflow. While the core system design shows strong potential and yields compelling results, the manuscript currently suffers from organizational issues and lacks critical implementation details. These gaps can be clarified by the authors during the rebuttal. I think the emerging conceptual questions (e.g., clinician overfitting, data privacy) offer great opportunities for constructive discussion within the MICCAI community.

  • Reviewer confidence

    Very confident (4)

  • [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 VDSS, a multi-agent framework using LLMs for ventilator titration. Instead of relying on a single end-to-end model, it divides the clinical workflow into specialized agents (detection, goal setting, mode selection, parameter planning). A key feature is the contextual bandit module that learns individual clinician preferences from accepted/rejected proposals. It also incorporates a Reflect Agent to revisit specific decision stages when a clinician rejects a recommendation, keeping the human firmly in the loop.

  • 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.This paper breaks down the decision-making process into independent agents, an approach that closely aligns with the actual clinical reasoning patterns of ICU teams. For this high-risk clinical environment, a Human-in-the-Loop verification mechanism is undoubtedly a practical and essential design choice. 2.This paper employs the Contextual Bandit to dynamically adapt to individual decision-making preferences based on feedback from clinicians. This ingenious strategy effectively addresses the common problem of inter-rater variability in clinical assessment, which is often ignored in similar studies. 3.The Reflect Agent introduced in this paper possesses the capability to precisely pinpoint the specific stages requiring revision. When a proposed solution is rejected, this module eliminates the need to recompute the entire decision-making process from scratch. This design constitutes a significant contribution at the system level and yields substantial time savings in practical applications. 4.The performance improvements over baseline LLMs are substantial. For instance, adding VDSS on top of GPT-5.2 drops the MSE from 0.343 to 0.102, and the expert ratings for safety and acceptability show massive improvements.

  • 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.This assessment relies entirely on reenacting past retrospective ICU treatment trajectories. For a system explicitly designed to support real-time human-computer interaction, this paper lacks real-world validation results such as pilot studies or simulations of real-time feedback, and therefore cannot demonstrate its effectiveness in practice.

    1. Several key implementation details are missing, making it hard to judge reproducibility. While the paper lists 12 preference categories, it fails to disclose how these are mathematically encoded. Furthermore, the exact mechanics of the BanditUpdate function (e.g., reward calculation, state encoding, the specific bandit algorithm used), prompt templates, and specific rules for the Safety Check are not sufficiently detailed. 3.Weak baselines: The introduction heavily criticizes rigid rule-based protocols, yet the experiments only compare VDSS against standalone LLMs. To prove its superiority over standard care, the authors should have included common clinical protocols or traditional rule-based methods as baselines.
    2. Insufficient description of certain modules. The Waveform Analyzer and the specific parameters/thresholds used in the safety-check mechanism are described only briefly.
  • 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 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
  • 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 framework is highly relevant to clinical practice, and I appreciate how the system architecture mirrors actual ICU workflows (especially the human-in-the-loop and preference learning aspects). The experimental results also show clear and substantial improvements over standalone LLMs. However, the contribution is primarily at the system-integration level rather than introducing fundamental algorithmic novelty. Moreover, the evaluation remains limited to retrospective studies, lacks rule-based clinical baselines, and misses critical mathematical details regarding the contextual bandit implementation. Despite current limitations, the framework’s system design and high clinical relevance make it a valuable practical contribution. Recommend Weak Accept.

  • 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 the meta-reviewer and all three reviewers for their careful assessment and positive evaluation. We are pleased that the reviewers recognized the clinical significance of dynamic bedside ventilator titration, as well as the novelty, clinical relevance, and system-level value of VDSS. The core contribution of VDSS is an auditable, interactive, and safety-constrained workflow for high-risk ICU decision support, rather than a simple application of LLMs to ventilator recommendation. By decomposing ventilator titration into clinically meaningful modules and integrating clinician feedback, Reflect Agent-based targeted replanning, and contextual-bandit preference adaptation, VDSS improves recommendation acceptability, clarity, and interaction efficiency while preserving clinician oversight and safety boundaries. R1-W1, R3-W1: Evaluation Boundary As stated in the manuscript, VDSS is evaluated through retrospective ICU trajectory replay and expert review. This setting is intentionally chosen to assess a high-risk ventilator decision-support workflow in a controlled, auditable, and replayable environment before prospective clinical use. Within this setting, VDSS demonstrates that a multi-agent, safety-gated, human-in-the-loop framework can improve recommendation quality, interaction efficiency, and clinical interpretability while preserving clinician oversight. Future validation will extend VDSS toward prospective silent-mode studies, clinician-in-the-loop simulation, bedside workflow evaluation, and institution-specific safety review. We will further clarify this staged evaluation boundary and intended use scenario in the revised manuscript. R1-W4, R2-W2: Runtime and Use Scenario Under the current local inference setting, VDSS requires approximately 305 seconds per cycle on average, with waveform analysis as the main computational bottleneck. At this stage, VDSS is intended for retrospective review, simulation-based evaluation, training, and pre-clinical workflow assessment rather than unsupervised real-time closed-loop control. Its modular design provides clear paths to reduce latency, including stable-context caching, asynchronous waveform preprocessing, lightweight waveform modules, and localized replanning after rejection. R1-W2/W3, R2-W1/W4, R3-W2/W4: Clarity and Reproducibility VDSS is a workflow-level clinical AI system built on modular clinical reasoning, structured interfaces, safety constraints, and human-AI interaction. The system distinguishes LLM-based agents from deterministic routing and validation modules, while the ablations and baselines isolate the contribution of waveform evidence, preference context, targeted replanning, and the overall workflow design. The single-agent baseline verifies that the gains of VDSS arise from the multi-agent workflow and clinician-in-the-loop interaction, rather than differences in input information. For preference learning, VDSS uses twelve clinically interpretable categories as contextual-bandit arms. For example, prefer_conservative_step captures a clinician’s preference for conservative, small-step, and reversible parameter adjustments. The clinician preference state is updated only after cycle closure based on the final accepted decision, reducing intermediate rejection noise and preference drift. Preference learning affects candidate ranking and adjustment style, but does not override safety constraints, device compatibility, or clinician judgment. The Waveform Analyzer provides structured evidence, and the Safety Check verifies mode compatibility and parameter consistency before clinician review. We will add clearer module-level descriptions to the final version to improve reproducibility and transparency. We again thank the meta-reviewer and reviewers for their constructive comments and positive evaluation of our work.



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.

    All reviewers like the manuscript, which proposed a human-in-the-loop multi-agent framework to address the problem of dynamic bedside ventilator titration. The study is timely and the method is novel. However, there are raised issues regarding the runtime and the clarity of the paper. Authors shall address these problems in the final camera-ready submission.



back to top