List of Papers Browse by Subject Areas Author List
Abstract
Identifying robust associations between cardiac imaging phenotypes and clinical diseases is fundamental to population-scale cardiovascular research and reliable risk stratification. However, current phenome-wide association studies rely on pre-defined, single-variable phenotypes or expert-crafted features, which limits their ability to capture clinically meaningful non-linear effects and cross-phenotype interactions. To address this, we propose CPAgent, an iterative phenotype-Composition framework for cardiovascular Phenome-wide association study (PheWAS) that automatically constructs and validates interpretable composite phenotypes (e.g., polynomial, ratio, and interaction forms) from base imaging features. Specifically, our system coordinates three agents: (i) an Analyst that identifies statistical pathologies and nominates candidate transformations; (ii) a Proposer that generates constrained, medically and statistically motivated expressions under numerical safety rules; and (iii) a Verifier that evaluates candidates using multi-stage criteria and produces transparent evidence trails for accepted phenotypes. Evaluated on a population-scale cardiac imaging cohort, the discovered composite phenotypes markedly improve disease discrimination: across 72 classifier–disease–metric combinations, our variants achieve the top rank in 56 cases versus 18 for baselines, with gains observed across all nine clinical disease categories. Our framework yields compact, clinically interpretable phenotype formulas with transparent evidence trails, enabling scalable discovery of stronger phenotype-disease associations beyond expert-driven feature selection.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0092_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/LumaLabAI/CPAgents
Link to the Dataset(s)
UK Biobank: http://www.ukbiobank.ac.uk/register-apply
BibTex
@InProceedings{LiZuo_CPAgents_MICCAI2026,
author = { Li, Zuoou AND Zhao, Wenlong AND Yu, Kelly AND Zhang, Weitong AND Matthews, Paul M. AND Bai, Wenjia AND Kainz, Bernhard AND Qiao, Mengyun},
title = { { CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces CPAgents, an agentic framework for automatically constructing interpretable composite phenotypes from cardiac imaging data. By integrating statistical profiling, LLM-guided feature generation, and multi-stage validation, it discovers clinically meaningful feature combinations that improve disease association analysis and classification performance compared to expert-defined and prior automated approaches in large-scale cohorts.
- 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’s key strengths are its structured agentic framework for feature discovery, combining statistical analysis with LLM-guided generation, and its focus on interpretable, closed-form phenotypes. It includes a robust multi-stage validation process and a comprehensive evaluation across diseases and models, demonstrating consistent performance gains and practical relevance on large-scale clinical data.
- 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.
A key weakness is limited conceptual novelty. The overall pipeline closely resembles prior work on LLM-driven feature engineering and symbolic composition, such as LLM-FE (Abhyankar et al. , 2025) and agent-based systems like MESHAgents (Zhang et al. , 2025). The Analyst–Proposer–Verifier loop is intuitive but not clearly differentiated from existing multi-agent or symbolic regression frameworks.
- 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 has provided an anonymized link to the source code, dataset, or any other dependencies.
- 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
CPAgents presents a compelling agentic framework for cardiac phenotype discovery with strong empirical results. The three-agent design is well-motivated, though broader dataset validation and computational cost analysis would strengthen the contribution.
- 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?
CPAgents uses three AI agents to automatically discover interpretable composite cardiac phenotypes, outperforming expert-defined features across nine diseases.
- 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 proposes CPAgents, a three-agent LLM framework for automated composite phenotype discovery in cardiovascular PheWAS. An Analyst profiles statistical properties of base features, a Proposer generates closed-form candidate transformations over a constrained AST operator set, and a Verifier filters them via ΔAUC, ElasticNet stability, and SHAP ranking. Candidates are accepted through a cross-classifier ensemble rule with early stopping.
- 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 analyse–propose–verify decomposition is clean and better scoped than most LLM-FE work. Grounding the Proposer with Analyst statistics (ΔAIC, MI, ICC) is a meaningful design choice. 2.The evaluation is large-scale and leakage-aware. A 6:2:2 split with discovery restricted to train+val, five seeds, and bootstrap stability selection show care. 3.The 72-configuration ranking analysis avoids cherry-picking. The silhouette score, association heatmap, and agent-level ablation triangulate the core claim well. Which deserves credit. 4.The sex-stratified allometric LVM Z-score example is a convincing demonstration that LLM medical priors can surface genuinely clinical constants. This is really intuitive.
- 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.
I would like to have few clarifications from the authors regarding- 1.Gains are small and often within overlapping standard deviations. SVM Hypertension is 0.759±0.006 vs 0.755±0.007.No paired significance tests (DeLong, bootstrap, Wilcoxon) are reported to back the 56 vs 18 ranking claim. Can you do the paired-significance tests and report the results in your rebuttal? 2.I think a compariosn with non-LLM AutoFE methods (OpenFE, AutoFeat, featuretools) or a random search over the same operator set O could be really intuitive to observe. Without these, the LLM’s contribution cannot be isolated from the AST pipeline itself. The ablation already removes agents but never the LLM. 3.LLM reproducibility is unclear. Results use DeepSeek-V3.2 and GPT-5.1 at default temperature. The five-seed averaging does not isolate LLM stochasticity from classifier seed variance. A temperature-0 run would help. While working on LLMs, I think this is necessary to report in the results table.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Method is strong and well written but the absolute gains are often within one standard deviation, the baseline set omits standard non-LLM AutoFE methods needed to credit the LLM specifically, and no paired significance tests back the ranking claim. The iterative accept rule is a multiple-testing procedure that is not characterized. If authors resolve the queries mentioned in the weakness section, I will be happy to make this an ‘accept’. Good luck.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal satisfactorily addresses my main concern on statistical validation and partially addresses the AutoFE baseline gap through an OpenFE comparison. However, the LLM-specific contribution and reproducibility are still not fully isolated, so I lean toward a weak accept rather than a strong accept.
Review #3
- Please describe the contribution of the paper
This paper proposes a novel agent‑based framework for cardiovascular phenome‑wide analysis that automatically constructs and validates interpretable composite phenotypes from base cardiac imaging features. The key contribution lies in decomposing phenome construction into three interacting agents—an Analyst that identifies statistical pathologies, a generator of statistically motivated phenotype expressions, and a Verifier that evaluates candidates and provides transparent evidence trails. Together, these components aim to produce robust, interpretable, and reproducible phenotypes suitable for large‑scale cardiovascular studies and downstream association analyses.
- 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.
A major strength of the paper is its emphasis on interpretable composite phenotypes, moving beyond single imaging measurements toward higher‑level constructs that better reflect cardiovascular disease complexity. This aligns well with clinical reasoning and phenome‑wide study objectives. A second strength is the agent‑based formulation, where analysis, synthesis, and verification are explicitly separated, providing built‑in quality control and reducing manual, ad‑hoc feature engineering.
- 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.
The is the use of the term “phenome‑wide association study” without performing any actual association analysis, particularly genetic associations, which may lead to conceptual ambiguity under traditional PheWAS definitions. Second, the framework operates exclusively on pre‑extracted imaging features, inheriting any biases or limitations of upstream steps. Third, while statistically well motivated, the proposed criteria for phenotype construction are not explicitly grounded in pathophysiological or clinical priors, leaving the clinical meaningfulness of some composite phenotypes to be inferred rather than demonstrated.
- 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.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The work reflects a solid and novel methodological contribution with clear strengths in interpretability. While the paper does not fully realize a traditional PheWAS and has limitations related to scope, these issues are largely clarifications rather than fundamental flaws.
- 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.
N/A
Author Feedback
We thank the reviewers for their constructive feedback. We are pleased that the reviewers recognised CPAgents as a compelling framework for clinical reasoning, highlighting its “clean analyse-propose-verify decomposition” (R1, R2), “interpretable composite phenotypes” (R1, R3), “robust leakage-aware validation” (R1, R2), and “comprehensive evaluation” (R1, R2).
< Statistical validation of performance gains (R2-W1, Meta) > To statistically validate the reported gains, we performed paired Wilcoxon signed-rank tests for the overall matched AUC comparisons and paired DeLong tests for disease-level AUC comparisons against MESHAgents under matched evaluation settings. Both our DeepSeek-V3.2-based and GPT-5.1-based agents show significant improvements over MESHAgents, with p less than 0.05 in all tests. We will explicitly clarify the statistical significance of these comparisons in the revised version.
< Additional baselines against AutoFE and novelty (R1, R2-W2, Meta) > We agree AutoFE methods such as OpenFE [Zhang et al., 2023] are useful references. However, CPAgents are designed with a different objective from conventional AutoFE pipelines, which typically search shallow algebraic combinations to maximize predictive accuracy on a fixed target. CPAgents are motivated by the construction of clinically interpretable composite phenotypes under explicit validity constraints. Candidate phenotypes are generated within a constrained closed-form operator space and are then filtered through verifier-based checks for physiological plausibility, statistical stability, and cross-classifier consistency. In our evaluation, OpenFE underperformed both baselines, indicating that unconstrained feature generation alone is insufficient for this setting. Directly applying unconstrained algebraic feature search may require additional filtering to ensure unit consistency and valid value ranges for downstream classification. We will make this distinction clearer in the revised version and position AutoFE methods as useful comparative references.
< LLM reproducibility (R2-W3)> We agree that unconstrained LLM generation can raise reproducibility concerns. In our framework, however, the LLM is not used as a raw stochastic generator whose outputs are directly evaluated. Instead, it proposes candidate features within a predefined feature space, and each candidate must pass deterministic structural, unit-consistency, and value-range checks before downstream evaluation. Under this fixed validation and evaluation protocol, we report results averaged over five independent seeds, covering both LLM-generation stochasticity and classifier seed variance. The non-zero temperature is used only to diversify candidate proposals, while the retained candidates are governed by the fixed constraints and validation protocol. We will emphasize this constrained generation-and-validation design in the revised version.
< Clarification of PheWAS scope, feature dependency, and clinical grounding. (R3-W1/W2/W3) > First, we agree that “PheWAS study” may be ambiguous under the traditional genetic PheWAS definition. Our work does not perform genetic association analysis but a PheWAS-inspired, disease-wide association analysis between cardiovascular imaging-derived phenotypes and clinical outcomes. To avoid confusion, we will revise the terminology to “phenome-wide imaging phenotype association analysis”. Second, we acknowledge that CPAgents operates on pre-extracted imaging-derived features, which is an intentional design choice to ensure interpretability, unit-level validity and fair comparison under a shared feature space. Third, we agree that clinical grounding should be made more explicit. The LLM component is intended to support exploratory hypothesis generation rather than to serve as a standalone source of clinical evidence. We will clarify the current scope and discuss external evidence grounding through curated biomedical literature or clinical guidelines as a future work.
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 paper received a mixed review. The key strengths of the paper include a well-structured agentic framework, strong clinical interpretability and translational potential, rigorous and large-scale evaluation design, as well as clinical relevance to cardiac imaging. At the same time, following weakness has been identified by the reviewers, including limited conceptual novelty relative to existing pipeline, insufficient statistical validation of performance gain, and missing critical baseline comparisons. I would recommend an invite for rebuttal such that the authors can address concerns raised by 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.
The authors have satisfactorily addressed All Reviewers’ criticisms, suggestions and observations. It is recommended, however, that the authors incorporate the statistical analysis performed, in the final version of the paper, as well as the clarifications needed, which authors acknowledged.
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 rebuttal addressed the non-LLM AutoFE baseline difference and speculated about the stat significance testing. This addresses some of the serious critiques from the review round. I think the actuality of this paper and the sota agentic methods place this one above the threshold and recommend accept.
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.
The authors have satisfactorily addressed the concerns raised in the original reviews. Though some of the comments could have used additional justification (as raised by R2), the contributions are strong, and the paper deserves to be published in the current form.