Abstract

Cardiovascular disease (CVD) modeling has become highly accurate, yet most models remain “black boxes” that prioritize predictive scores over reliable and reusable clinical knowledge. We propose Evidence-Grounded Learning (EGL), which repeatedly tests candidate variable sets under a fixed evaluation budget and records auditable evidence about which factors and pairwise relationships are consistently supported. EGL–LLM adds an interpretable large language model (LLM) control plane that uses a semantic agenda to decide what to test next, while keeping the evidence definitions and budget unchanged. On the COPDGene cohort, EGL shows reliable convergence of core evidence and recovers clinically plausible dominant factors and cohort-relevant interaction patterns. Semantic steering broadens the exploration path while preserving stable dominant factors and redirecting evidence acquisition toward alternative interaction neighborhoods, supporting evidence-grounded discovery that yields auditable cohort-level insights. Code is available at https://github.com/liangyuxuan1/EGL.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4195_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/liangyuxuan1/EGL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LiaYux_LLMSteered_MICCAI2026,
        author = { Liang, Yuxuan AND Xu, Xuanang AND Wang, Ge AND Yan, Pingkun},
        title = { { LLM-Steered Clinical Knowledge Discovery for CVD Outcome Modeling } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses the problem of explainable knowledge discovery in cardiovascular disease (CVD) modeling, moving beyond conventional black-box prediction toward clinically reusable insights. The main contributions are as follows, 1.The authors propose an EGL framework that reformulates clinical modeling as a process of evidence accumulation rather than single-model optimization. It constructs a graph-structured representation of knowledge, where nodes and edges encode feature importance and interactions derived from repeated empirical validation. 2.The method models feature subset distributions using an Ising-based formulation and performs optimization via Gibbs sampling combined with REINFORCE, enabling structured exploration of high-order feature combinations. 3.A paired probing strategy is introduced to quantify both individual feature contributions (∆f) and pairwise interaction effects (Γ_uv) through a difference-in-differences scheme, resulting in an auditable evidence ledger. 4.The framework incorporates a large language model (LLM) as a semantic controller to guide feature subset exploration under a constrained budget. Through structured prompts (e. g. , must_include, prefer_exclude, frontier_edges), the LLM aims to improve the efficiency of interaction discovery.

  • 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.

    Major Strengths 1.The paper shifts the focus from predictive performance to reusable clinical knowledge discovery, aligning well with current demands for interpretability and reliability in medical AI. 2.The proposed method integrates feature subset search, statistical validation (paired probing), reinforcement learning, and graph-based representation into a unified pipeline, demonstrating strong internal logical consistency. 3.Unlike traditional feature importance methods that focus on marginal effects, the work quantifies interaction effects via a difference-in-differences formulation, offering a more expressive and informative representation of feature relationships.

  • 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.

    Major Weaknesses 1.Experimental results suggest that the LLM primarily alters the search trajectory without significantly improving the final discovered features. In some cases, it even slows convergence. Moreover, the paper lacks thorough ablation studies (e. g. , different prompts, alternative LLMs, or removal of semantic guidance) to justify its inclusion. 2.The evaluation is restricted to a single dataset (COPDGene), with no cross-dataset or cross-population validation, limiting confidence in the generalizability of the proposed approach. 3.The paper does not include comprehensive comparisons with established feature selection and knowledge discovery methods (e. g. , LASSO, Boruta, SHAP-based selection, or causal discovery approaches), weakening the empirical support for its claimed advantages. 4.Several key design choices are insufficiently motivated, including the use of the Ising model for feature subset modeling, the adoption of REINFORCE instead of more stable alternatives, and the specific design of the LLM agenda and model selection. 5.The method involves extensive subset sampling, K-fold evaluation, and paired probing, leading to substantial computational cost. However, the paper does not provide a detailed analysis of efficiency or scalability. 6.The method section is dense, with heavy notation and limited intuitive explanation. Important components such as the reward design and evidence ledger are not sufficiently clarified, affecting accessibility. 7.There are concerns regarding reference formatting and overall writing redundancy, which detract from the paper’s presentation quality.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 addresses an important and timely problem—moving from black-box prediction toward interpretable and reusable clinical knowledge discovery in CVD modeling. The proposed Evidence-Grounded Learning (EGL) framework is conceptually novel, particularly in its formulation of knowledge as an evidence accumulation process and its integration of feature subset search, statistical probing, reinforcement learning, and graph-based representation. The explicit modeling of feature interactions via paired probing is also a meaningful contribution beyond standard feature attribution methods. However, several limitations prevent a stronger recommendation. The empirical validation is limited to a single dataset, and comparisons with established feature selection or causal discovery baselines are insufficient. The contribution of the LLM module is not fully convincing, as it mainly alters the search trajectory without clear gains in final outcomes. In addition, some design choices lack sufficient justification, and the overall method introduces notable computational complexity. Despite these weaknesses, the paper presents a novel perspective with potential clinical impact, and its core framework could inspire future work on interpretable knowledge discovery. Therefore, I recommend a weak accept.

  • 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 #2

  • Please describe the contribution of the paper

    The paper introduces a novel framework for clinical knowledge discovery in cardiovascular disease (CVD) modeling by combining Evidence-Grounded Learning (EGL) with LLM-driven semantic control (EGL–LLM). The key contributions are: EGL: A method that learns a distribution over feature subsets to identify reproducible node and edge effects (marginal and interaction evidence) via an Ising-form policy and paired probing. EGL–LLM: An extension that integrates a large language model (LLM) to generate interpretable, semantically guided agendas (e.g., must_include, frontier_edges) for subset sampling and probing, prioritizing clinically meaningful exploration. Auditable Evidence Ledger: A graph-based ledger that accumulates reproducible evidence (node/edge effects) through repeated cross-validation, enabling transparent and reusable clinical insights.

  • 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.

    Novel Integration of LLMs in Clinical Discovery. EGL–LLM is the first to use an LLM for semantic steering of feature exploration in CVD modeling. The LLM generates agendas based on real-time evidence (e.g., uncertainty, stability) rather than purely statistical criteria, addressing the “black-box” problem in predictive models.

    Structured Methodology for Interaction Discovery The Ising-form distribution over feature subsets, combined with counterfactual probing (e.g., comparing subsets with/without a feature), provides a principled way to test marginal and interaction effects.

    Empirical Validation of Clinical Plausibility Both EGL and EGL–LLM recover clinically plausible dominant factors (e.g., Agatston score, hypercholesterolemia) and interactions (e.g., Agatston score with comorbidities). Negative interactions (sub-additive effects) highlight redundancy in overlapping CAC-related markers, aligning with known clinical mechanisms

    Efficiency in Interaction Evidence Consolidation EGL–LLM achieves faster convergence of interaction evidence in later iterations, demonstrating the value of semantic steering in prioritizing meaningful neighborhoods.

  • 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.

    Limited Generalizability: The framework is evaluated only on the COPDGene cohort (smokers/COPD patients). No analysis is provided for non-COPD populations or other diseases, limiting the applicability of the findings.

    • No comparisons are made to existing feature selection/interaction discovery methods (e.g., SHAP, Bayesian networks), making it unclear whether EGL–LLM offers unique advantages.

    Ambiguity in Clinical Plausibility: While factors like Agatston score are labeled as “clinically plausible,” the paper lacks references to prior studies or statistical validation of these associations in COPDGene. The LLM’s role in avoiding domain-specific biases is not addressed; there is no evidence that generated agendas align with validated clinical knowledge (e.g., clinician review).

    Methodological Limitations Fixed Evaluation Budget: The identical budget for EGL and EGL–LLM ensures fair comparison but may not reflect real-world scenarios where resources vary. Reward Function Design: The scalar reward R(S) combines confirmation and acquisition terms, but the weighting of AUC/AuPRC in the metric summary m(.) is not explicitly defined, leaving room for arbitrary tuning. CT Pipeline Sensitivity: The CT-derived biomarkers are processed via a specific pipeline, but no comparison is made to alternative methods, risking overfit to this pipeline.

    Risk of Missing Data Bias: Clinical variables are filtered by a 70% missingness threshold, but the paper does not address how missing data in the remaining 30% might bias the evidence ledger.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 current evaluation is limited to the COPDGene cohort, which makes it difficult to assess how well the proposed framework would generalize to other populations, datasets, or disease settings. Several discovered factors and interactions are described as clinically plausible, but this plausibility is not clearly validated against prior literature, external evidence, or clinician review. The manuscript would benefit from a clearer discussion of the LLM component, particularly regarding possible bias inherited from pretraining data and whether any domain-specific tuning, adaptation, or validation of the generated agendas was performed. The paper does not compare EGL–LLM with existing feature selection or interaction discovery methods, which makes it difficult to evaluate the specific added value of the proposed approach. It would also be helpful to clarify whether the absence of external validation and baseline comparisons is due to scope limitations or is left for future work.

  • 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 novel framework (EGL–LLM) that addresses critical gaps in clinical knowledge discovery for CVD modeling. The empirical results are promising, and the evidence ledger concept is innovative. The lack of external validation, clinical plausibility checks, and methodological comparisons limits the confidence in the framework’s broad utility. However, the work provides a strong foundation for future research and has the potential to advance interpretable CVD risk modeling.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 have satisfactorily addressed my comments and questions.



Review #3

  • Please describe the contribution of the paper

    The paper proposes an evidence-grounded learning (EGL) framework for clinical knowledge discovery in cardiovascular disease (CVD) outcome modeling. Instead of focusing on a single predictive model, the paper aims to identify reproducible and interpretable feature level and interaction level evidence by learning a distribution over feature subsets.

    The framework combines an Ising-form probabilistic sampling policy with a paired probing strategy to estimate marginal and interaction effects, which are accumulated in an evidence ledger. A key contribution is the integration of a large language model (LLM) as a semantic control plane that guides subset sampling and probe selection through structured agendas, with the final goal of improving exploration efficiency and clinical plausibility of discovered relationships in CVD modelling.

  • 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) Interesting and timely research direction: The paper shifts the focus from pure predictive performance to clinical knowledge discovery, which is highly relevant for medical AI where interpretability and reproducibility are critical, minimizing the black box label associated with AI-driven diagnosis in medicine.

    2) Novel integration of LLMs for semantic guidance: The use of an LLM as a control mechanism to guide feature exploration based on clinical semantics is a creative idea. Rather than using LLMs for prediction, the work leverages them to evaluate and steer the discovery of clinically meaningful feature subsets and interactions.

    3) Effort toward interpretability and transparency: The concept of an evidence ledger that accumulates marginal and interaction effects provides a more interpretable alternative to black-box models, potentially helping mitigate the black-box nature of standard ML approaches.

    4) Methodological coherence: The framework (sampling, probing, evidence accumulation, and control) is logically structured and consistent, even if complex.

  • 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) Limited theoretical grounding of the LLM control component: While the idea of semantic steering is interesting, the role of the LLM remains heuristic. There is no formal objective or theoretical justification explaining why or how the LLM-guided agenda improves convergence or guarantees better discovery of clinically relevant interactions.

    2) Insufficient analysis of the LLM contribution: The paper would benefit from deeper analysis or ablation of the LLM component. For example, isolating the effects of agenda shaping versus probe prioritization, or comparing against simpler heuristic strategies, would strengthen the claims.

    3) Clarity issues in method description: The paper is conceptually dense and sometimes difficult to follow, especially in the description of the sampling policy, reward formulation, and interaction between EGL and the LLM control plane. Several components (e.g., “semantic agenda”, “control paths”) are introduced at a high level without sufficient intuitive explanation.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    This is an interesting and promising direction toward bridging machine learning and clinical knowledge discovery. The idea of combining probabilistic exploration with LLM-based semantic guidance is compelling and could inspire future work.

    To further strengthen the contribution, the authors are encouraged to improve the clarity of the methodological description, and to provide a more rigorous analysis of the role and impact of the LLM component. In particular, clearer justification and more detailed evaluation of the semantic control mechanism would significantly enhance the paper.

  • 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 offers an interesting and relevant research direction with the focus on evidence-grounded clinical knowledge discovery rather than predictive modeling. Moreover, the introduction of the LLM as a semantic control mechanism is interesting and demonstrates good results in the exploration of feature interaction.

    However, the research has some limitations in terms of the absence of theoretical foundations and analysis of the LLM component mentioned in the paper , which is the core of the research. Moreover, the clarity of the methodology can be improved in the paper.

    Overall, the conference paper is above the acceptance threshold in terms of originality and potential impact.

  • 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 rebuttal clarified several important aspects of the proposed framework, particularly the intended role of the LLM component as a semantic scheduler rather than a direct predictive or optimization module. The authors provided a clearer explanation of how the LLM influences exploration of interaction neighborhoods while leaving the evidence estimation process unchanged.

    The response also clarified the rationale behind the Ising-form distribution, Gibbs sampling, and REINFORCE-style optimization, which improved the overall interpretability of the framework.

    While concerns regarding the lack of strong theoretical grounding and limited experimental isolation of the LLM contribution remain partially unresolved, the rebuttal sufficiently addressed the main conceptual misunderstandings and strengthened the overall coherence of the work.

    Overall, I maintain my positive assessment of the paper. The work presents an interesting and timely direction for interpretable and evidence-grounded clinical knowledge discovery, with meaningful potential impact for medical AI.



Author Feedback

We thank the reviewers for recognizing the novelty of evidence-grounded clinical knowledge discovery. Reviewers raised several related comments regarding the role of LLM, comparison with existing methods, cross-dataset validation, and the justification of key design choices. We therefore provide a unified response organized around these shared themes. 1.Role of LLM. Our proposed Evidence-Grounded Learning (EGL) follows numerical scores and can over-test locally good but clinically redundant regions. LLM in our enhanced EGL-LLM serves as a semantic scheduler to explore clinically meaningful features and consolidate interaction evidence. For example, vanilla EGL may stick to interaction between one strong node (like Agaston score) and others, but LLM can explore edges between other variables. Thus, LLM only suggests what features to test, while EGL still verifies them. As shown in Fig.2a, EGL quickly settles on the main nodes, while EGL-LLM keeps small substitutions around it, which helps avoid repeatedly testing the same locally strong variables and keeps the search open to nearby clinically related alternatives. The wider MDS footprint in Fig.2c shows expanded exploration without reducing edge acquisition (Fig.2b); identical top nodes confirm stable individual-variable effects (Fig.2d). The main difference between EGL and EGL-LLM is in the interaction set (Fig.2e), which is the intended effect: semantic scheduling redirects tests toward different neighborhoods of interaction. 2.Comparison with other methods. We agree that LASSO, Boruta, SHAP-based selection, Bayesian networks, and causal inference methods are important. However, their objectives and outputs are not the same as those of EGL. Feature selection methods mainly return selected or ranked predictive variables, while knowledge discovery methods often estimate model-specific interactions, dependencies, or causal structures. EGL instead evaluates the reproducibility of marginal and pairwise evidence through repeated matched subset comparisons and outputs a stable evidence graph. Thus, a fair comparison would require a unified protocol to translate different outputs into comparable node- and edge-level evidence and to assess stability and biological plausibility consistently. We will clarify this limitation, construct fair baselines and perform a comparative study in the future. 3.Cross-dataset validation. COPDGene is a strong primary testbed rather than a narrow single-site dataset: it is a large multi-site cohort and combines CT-derived cardiac/CAC phenotypes with risk factors, comorbidities, medications, and CVD labels. Its scale, site diversity, and multimodal coverage provide a meaningful setting for evaluating stable evidence discovery. We agree that external validation across multiple cohorts would further strengthen the findings, but such validation requires careful harmonization of imaging pipelines, clinical variables, and outcome definitions. We therefore regard this as an important direction for future work beyond the scope of the current study. 4.Justification of key design choices. The Ising-form distribution models correlated feature subsets in binary states (sampled vs not sampled), allowing candidate variables to be sampled jointly when interactions exist. Gibbs sampling was used as a standard sampler for Ising-type distributions. We used a REINFORCE-style algorithm because the search involves stochastic sampling, making it infeasible to backpropagate gradients through the full sampling–evaluation process. The reward captures two types of evidence: the marginal gain from adding an individual variable and the excess joint gain from adding a pair of variables beyond their separate contributions. Practical hyperparameters, including the missing-rate filter, reward weights, and the magnitude of the LLM-induced bias, were tuned empirically but omitted due to space limits.




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.

    Although three positive scores, several critical concerns have been raised by all the reviewers about the limited ablation studies, the lack of cross-dataset validation, insufficient comparison, and insufficient justification. The authors are invited to submit a rebuttal to carefully respond and clarify all the concerns raised by the reviewers.

  • 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.

    This paper was initially invited for rebuttal following the primary review period, where reviewers raised several concerns. After carefully evaluating the authors’ rebuttal, the reviewers’ revised feedback, and re-examining the paper, I conclude that the authors have effectively addressed the key concerns raised, and the work now meets MICCAI’s publication standards.



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 work was rated positively by the three reviewers, recognizing the novelty and the contribution of the work. The possiblity of address technical concerns in the rebuttal was taken and appreciated by 1+ reviewer(s).



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.

    Although there are initial concerns regarding limited ablation, cross-dataset validation, and comparison to baselines, authors have adequately addressed reviewers’ concerns in their rebuttal. R2 and R3 now recommend acceptance, acknowledging the novelty and potential impact of the work. The rebuttal clarified the role of the LLM as a semantic scheduler, justified key design choices, and explained the scope of the current study regarding dataset limitations. While the paper would benefit from further quantitative validation, it already presents a valuable contribution to interpretable clinical knowledge discovery. Hence, AC recommends acceptance.



back to top