Abstract

Multi-page polysomnography (PSG) reports contain dense plots, summary indices, and annotations that clinicians must integrate with free-text patient narratives (CC/HPI) to produce diagnoses and treatment plans. However, current multimodal LLMs often fail to provide auditable clinical outputs: they struggle with evidence grounding, may hallucinate unsupported claims, and must abstain when evidence is missing or indeterminate. We propose a constraint-driven two-stage framework for auditable structured clinical decision support from multimodal sleep documents. Stage 1 aligns report reading via rule-scored metric extraction using reinforcement learning with group-relative policy optimization (GRPO). Stage 2 warm-starts from the aligned checkpoint and performs supervised fine-tuning to generate structured clinical JSON with diagnosis, evidence atoms, notes items, and treatment fields. Outputs are grounded through a restricted Pointer Bank and a bounded evidence-atom budget, enabling deterministic provenance checks. We evaluate on a retrospective tertiary-center dataset with 657 studies; the train/val/test split is 479/78/100 instances. Our model achieves 100% JSON parseability and strong evidence alignment (Grounding Pass 88%; metric-pointer F1 0.8910, 95% CI 0.8676–0.9113), with strong insomnia subtype prediction (F1 0.8821). Under our evaluation protocol, it avoids forced OSA diagnosis under indeterminate evidence (0% OverDx) and produces no detected safety-rule violations. We additionally conduct a physician rubric evaluation on 100 cases to assess long-form clinical recommendations. Ablations show Stage 1 alignment is critical for grounding and diagnostic reliability under limited supervision.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5410_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{WuSip_TRACEPSG_MICCAI2026,
        author = { Wu, Sipeng AND Tang, Shiqin AND Liu, Chenxi AND Wu, Jingwei AND Jiang, Wenjing AND Xiao, Xiaoyu AND Liu, Hongbin AND Meng, Gaofeng},
        title = { { TRACE-PSG: Traceable Sleep Diagnosis Support with Pointer-Grounded Evidence } },
        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

    The paper proposes TRACE-PSG, a two-stage multimodal framework for auditable clinical decision support from multi-page PSG reports together with CC/HPI text. In the first stage, the model is trained to read reports and extract structured metrics using rule-scored supervision. In the second stage, it generates structured JSON outputs containing diagnosis, evidence atoms, notes items, and treatment recommendations. A key contribution is the use of a restricted Pointer Bank and bounded evidence budget, which makes diagnostic claims explicitly traceable to report evidence and allows deterministic verification of provenance. The work is evaluated on a retrospective sleep-study dataset and emphasizes grounding, safety, and auditability in addition to task performance.

  • 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 addresses an important and practical problem. Many recent multimodal medical LLM papers focus on general question answering or report generation, but fewer address auditable decision support on dense clinical documents where explicit evidence localization matters. The focus on grounding and traceability is therefore timely and clinically relevant.

    A major strength is the task formulation. Instead of relying on unconstrained free-text generation, the paper uses structured JSON outputs with pointerized evidence atoms. This makes the system more checkable and clinically accountable, and the distinction between syntactic validity and factual grounding is well made.

    The two-stage design is also sensible. Teaching the model to read and extract report content before full structured generation appears principled, and the ablations suggest that this alignment stage contributes meaningfully to parsing, grounding, and diagnostic behavior.

    I also appreciate the evaluation design. The paper reports not only end-task prediction performance, but also parseability, pointer validity, grounding, safety-oriented overdiagnosis behavior, and clinician rubric scores. This broader evaluation is better aligned with deployment concerns than using a single accuracy metric alone.

    Overall, the paper feels reasonably novel in the way it combines multimodal document reading, explicit provenance constraints, bounded evidence, and abstention-aware evaluation for sleep-report decision support.

  • 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 baseline section is somewhat narrow. The paper compares mainly against Qwen3-VL variants, which are useful foundation-model references, but not necessarily the strongest comparison set for a paper centered on grounding and auditable decision support. Stronger grounding-oriented or verification-oriented baselines would better contextualize the contribution.

    A second issue is that some of the strongest gains are on verifiability metrics that are partly induced by the design itself. The Pointer Bank and bounded evidence budget naturally improve parseability and grounding-pass behavior. That is still valuable, but these improvements should be interpreted primarily as gains in controllability and auditability rather than broad evidence of superior clinical reasoning across all output types.

    The clinical evaluation is also still limited. The clinician rubric is useful, but it is based on 100 cases and appears to show only modest separation among systems, especially for the longer-form recommendation content. This makes it harder to judge how large the practical downstream benefit is beyond the structured, verifiable fields.

    Another limitation is generalization. The dataset comes from a single tertiary sleep center, is insomnia-dominant, and may reflect site-specific reporting conventions. The paper acknowledges this, but the question of portability to other institutions and report templates remains open.

    Finally, some methodological details would benefit from stronger justification. In particular, the choice of beta equals zero in the GRPO stage raises questions about stability and reward exploitation, and the paper would be stronger with a clearer explanation of why this training setup was reliable in practice.

  • 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.

    (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?

    I lean accept because the paper addresses a meaningful clinical problem and proposes a coherent, practically motivated framework for grounded and auditable multimodal decision support. The strongest part of the work is the structured generation setup with explicit provenance constraints, which makes the output more verifiable and clinically accountable. I also appreciate that the evaluation includes grounding, safety, and clinician-oriented criteria rather than relying only on task accuracy.

  • 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.

    I initially assigned an accept because I found the paper to make a relevant and technically sound contribution, while still having some concerns that limited my confidence. After reading the authors’ rebuttal, I believe the response helped clarify several points raised in my review and provided reasonable justification for the methodological choices. The rebuttal does not substantially change my overall assessment, but it reinforces my view that the paper is above the acceptance threshold. Therefore, I maintain my initial recommendation.



Review #2

  • Please describe the contribution of the paper

    This paper proposes a two-stage framework for auditable clinical decision support from multimodal polysomnography (PSG) reports and clinical text. The method first performs report-reading alignment using GRPO with rule-based rewards to extract sleep metrics, and then performs supervised fine-tuning to generate structured clinical outputs in JSON format. A key contribution is the use of a restricted pointer-based evidence representation, where each clinical claim is linked to predefined evidence sources, enabling deterministic provenance checks and improved audits.

  • 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 addresses an important and under-explored setting: multi-page clinical documents, which are significantly harder for LLMs compared to single image or text-only tasks.
    • The pointer-based evidence grounding mechanism, which enforces that all claims are referenced back to a “pointer back” of evidence sources, is a well-motivated and practically meaningful design for clinical settings.
    • The two-stage framework (GRPO followed by SFT) is also sensible and motivated by the ablation studies.
    • The authors provide a comprehensive evaluation suite, including validity and grounding metrics, safety, and clinician preference.
  • 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.
    • There is a crucial disconnect between quantitative metrics and clinician evaluation. In the clinician rubric (Table 4), the Stage2-only model matches or slightly outperforms the full method across multiple dimensions (clinical quality, usability, and overall impression), despite the full method achieving substantially better grounding and diagnostic metrics. This discrepancy is not adequately analyzed. Given the importance of clinician judgment in this domain, it raises questions about whether the proposed improvements on the quantitative metrics translate to meaningful gains in real-world clinical utility.
    • Relatedly, the paper does not clearly discuss the tradeoff between auditability and usability. The pointer-based constraints and evidence budgeting likely make outputs more conservative or rigid, which may improve verifiability but reduce perceived usefulness or fluency.
    • While the pointer mechanism is novel and compelling, the overall modeling approach is still somewhat incremental, relying on a combination of GRPO and SFT. The novelty lies more in the problem formulation and constraints than in fundamentally new techniques.
    • Additionally, the dataset is relatively small (657 studies) and drawn from a single center, which raises some concerns about generalization. There is no external validation or cross-institution evaluation.
    • Some design choices, like the fixed evidence atom budget of K=6, why some information like the treatment label are left unpointerized, and sensitivity of the results to the specific constraints or reward design, are not fully justified. Further, a published codebase would help support reproducibility.
  • 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?

    The paper addresses an important problem of grounded clinical decision making based on multimodal medical documents, and introduces a compelling pointer-based evidence framework. The formulation and evaluation are well-aligned with real-world clinical needs, and the results demonstrate improvements in grounding and report validity. However, the paper has some limitations that aren’t adequately discussed, particularly the lack of alignment between quantitative improvements and clinician evaluation, which raises questions about practical impact. Additionally, while the formulation is novel, the underlying modeling techniques are somewhat incremental.

  • 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.

    The authors’ rebuttal clarified aspects of the method and helped better contextualize the contributions. I still believe some concerns remain. Most notably, the discrepancy between the strong improvements on grounding/verification metrics and the relatively similar clinician rubric scores across methods is important and somewhat limits the strength of the practical clinical claims. In addition, the dataset size and single-center setting continue to leave open questions regarding generalization and robustness across institutions and report formats. That said, I continue to view the paper positively overall. The formulation is clinically relevant, the evaluation suite is broad, and the pointerized evidence representation provides a meaningful step toward more interpretable/auditable models. While the methodological novelty is more incremental than fundamentally algorithmic, I think the paper offers sufficient practical and conceptual value for acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper proposes TRACE-PSG, a two-stage multimodal framework for structured sleep-diagnosis support from multi-page polysomnography reports plus CC/HPI text. Stage 1 uses rule-scored GRPO to align report reading through metric extraction, and Stage 2 uses supervised fine-tuning to generate structured clinical JSON containing diagnosis, evidence atoms, notes, and treatment. The main technical idea is to restrict diagnosis-level and note-level evidence to a predefined Pointer Bank so that outputs can be audited through deterministic provenance checks rather than free-form rationales.

  • 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 tackles a real clinical-document problem that is more realistic than standard medical VQA. Multi-page PSG reports with mixed plots, indices, and narrative context are genuinely difficult inputs for current multimodal models. 2.The pointer-grounded output representation is directionally sensible. Restricting evidence citations to a controlled pointer space is a practical way to improve auditability relative to unconstrained justification text. 3.The paper evaluates more than just task accuracy. It includes parseability, pointer validity, over-diagnosis under indeterminate evidence, and a clinician-facing rubric, which is better than reporting one headline metric. 4.The ablation that removes Stage 1 does show that the alignment stage improves parsing and grounding under the authors’ setup, so the two-stage design is not entirely vacuous.

  • 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 empirical validation is much weaker than the presentation suggests. The dataset has 657 instances but only 136 patients from a single tertiary center, and the test set is just 100 instances. For a clinically heterogeneous sleep-diagnosis task, this is a very small basis for broad claims about reliable structured clinical decision support. 2.The key claim that the full method is better is not consistently supported by the authors’ own results. In Table 3, the ablation without the Pointer Bank actually has higher insomnia subtype F1 than the full method, and in Table 4 the Stage2-only model has the best aggregate clinician scores. This directly undermines the implied narrative that the full TRACE-PSG stack is clearly superior overall. 3.The paper overclaims on safety and hallucination. All three models receive 0.0 percent critical error and 0.0 percent hallucination in the 100-case clinician review, and all three have 0 percent safety violation in the rule-based evaluation. When even weaker ablations look equally perfect on safety, the metric is probably not sensitive enough, the rubric is too coarse, or the task has been defined too conservatively. In any case, these numbers are not persuasive evidence of real clinical safety. 4.The method’s novelty is limited. Pointer-restricted evidence citation, schema-constrained generation, warm-starting from extraction alignment, and reward-based extraction tuning are all recognizable ingredients rather than a fundamentally new MICCAI-level method. The paper packages them cleanly for PSG reports, but the algorithmic advance itself is modest. 5.The paper’s auditability story is incomplete. Diagnosis and notes are pointerized, but treatment is intentionally left unpointerized free-form text. That is exactly the portion where clinically harmful recommendations would matter most. As a result, the system is only partially auditable, while the title and framing suggest broader traceability than the implementation actually delivers.

  • 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?

    First, the dataset and evaluation are too limited for the strength of the claims. A single-center cohort with 136 patients and a 100-instance test set is not enough to support claims of reliable structured clinical decision support for multimodal PSG interpretation.

    Second, the full method does not clearly win on the paper’s own results. The Pointer Bank improves grounding metrics, but not the most clinically relevant downstream metrics in a consistent way. In particular, the no-Pointer-Bank model outperforms the full system on insomnia subtype F1, and Stage2-only has the best clinician-rated aggregate scores.

    Third, the novelty is incremental. The work is a task-specific assembly of constrained generation, extraction alignment, and provenance pointers rather than a substantial methodological advance.

    Fourth, the safety and hallucination evaluation is not convincing. Perfect or near-perfect safety numbers across all compared methods suggest that the evaluation protocol is not sufficiently discriminative. This makes the strongest claims in the abstract difficult to trust.

    Overall, I see a reasonable workshop-level or early systems paper, but not a sufficiently rigorous MICCAI paper at a high standard.

  • 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 thank reviewers and meta-reviewer for their comments. [Meta/R2: limited baselines; Pointer Bank/budget may favor verifiability metrics] We agree these gains indicate controllability/auditability, not broad clinical-reasoning superiority. Our comparisons ask whether scale/instruction following suffices for auditable PSG diagnosis and whether report-reading alignment improves grounding/abstention. Qwen3-VL-8B has high Grounding Pass but lower Pointer F1 and insomnia F1, showing format compliance does not ensure correct evidence/diagnosis. We will clarify in Sec. 4.[R3/R4: TRACE-PSG combines known components, not a new architecture] It is not a generic architecture contribution but a PSG-specific auditable formulation integrating report reading, pointerized provenance, bounded evidence, and abstention-aware evaluation. We will revise the contribution statement. [Meta/R2/R3/R4: full model improves automatic metrics but not clinician aggregates; ablations are slightly higher on some scores] TRACE-PSG is not claimed to dominate every metric. Its advantage is auditable/controllable structured support, measured by parseability, pointer grounding, evidence bounding, and abstention: Stage2-only shows 13.95% OSA OverDx, while TRACE-PSG reduces this to 0%. Clinician aggregates include less-atomized treatment/readability/usability aspects; the small Table 4 differences suggest these auditability gains do not cost clinician-facing quality. Thus, Table 4 marks the system boundary rather than contradicting the main claim. We will make this explicit. [R3: pointer constraints may reduce usability] The clinician rubric shows no usability/readability collapse; scores remain high. With improved pointer validity/grounding, this suggests better auditability without clear usability loss. We will clarify this. [R4: treatment is unpointerized despite the auditability framing] TRACE-PSG provides field-level auditability for diagnosis/evidence/notes, not every generated word or treatment recommendation. We will state pointer-grounded treatment as future work. [Meta/R3/R4: single-center small cohort/no external validation; R2: generalization concern] We agree the 657-study/136-patient single-center cohort with 100 test cases does not establish deployment/external reliability. We will temper the claim: the evidence supports retrospective internal feasibility of an auditable PSG formulation, not external validation of a deployable system. Physician-derived labels and patient-level splitting prevent multi-night leakage. [R2: insomnia-dominant data and site-specific conventions may limit portability] We agree this limits external generalization. TRACE-PSG uses predefined metric/note pointers and structured labels, making evidence explicit and auditable, but portability across templates still requires multi-center validation. We will state this. [R4: 0% safety/critical-error/hallucination rates do not prove real-world safety] These zero rates are minimum safety screens, not the sole safety evidence. More discriminative is abstention: Stage2-only has 13.95% OSA OverDx under indeterminate evidence, while TRACE-PSG reduces it to 0%, suggesting fewer unsafe forced diagnoses. We will distinguish limited checks from deployment-level safety validation. [R2: beta=0 may affect GRPO stability/reward exploitation] Stage 1 is constrained metric extraction with format-gated rule rewards, short training, and parsing/extraction monitoring, reducing reward-exploitation risk. We will clarify this. [R3: K=6 and reward/constraint sensitivity need justification] K=6 forces a small set of key, non-redundant evidence atoms clinicians can audit efficiently, rather than long exhaustive rationales. We will clarify this and discuss sensitivity to budget/reward design. [R3/R4: reproducibility and ethics details are insufficient] Data are de-identified; public release may be privacy-restricted. We will provide details and release code/prompts where permitted.




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.

    The reviewers appreciated very much the motivation/importance of the study, the task formulation, the two-stage design and the comprehensive evaluation.

    However, there are still some concerns raised by the reviewers and I would like to invite the authors for a rebuttal to see if the following major concerns can be addressed: 1.The limited baseline comparison 2.The disconnection between the quantitative metrics and the clinician evaluation 3.The model’s generalizability due to the dependence of a single data source and relative small dataset 4.Clarity about the auditability and usability

  • 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 provided a high-quality rebuttal and the concerns seem to be all resolved. Therefore, I recommend acceptance of this paper.



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 majority of reviewers recommend acceptation and appreciate the clarification done in the rebuttal.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The paper makes a relevant contribution to traceable sleep diagnosis support with pointer-grounded evidence, and two reviewers support acceptance after rebuttal. However, the review record still includes an unwithdrawn Reject from a reviewer who argues that the dataset and evaluation are too limited for the strength of the clinical decision-support claims, that downstream clinical metrics are mixed, and that the novelty is incremental. Since a substantive negative opinion remains unresolved



back to top