List of Papers Browse by Subject Areas Author List
Abstract
Risk stratification for pulmonary embolism (PE) is critical for clinical decision-making. Stratification guidelines are based on patient medical records, parameters measured from computed tomography pulmonary angiography (CTPA), and blood tests. However, blood tests are often missing in routine practice. This work studies whether state-of-the-art models can accurately classify risk stratification from only medical records and biomarkers extracted from CTPA images. We benchmark different approaches to combine medical records and cardiac biomarkers with rich pulmonary vascular information; we add vascular biomarkers to tabular models and apply graph neural networks (GNNs) on the vascular tree’s intrinsic graph representation. We use a private dataset (n=353) with uniquely complete data for PE risk stratification. Our results show that, among global features, medical records and cardiac biomarkers are the most significant predictors, while vascular biomarkers do not further improve stratification. Even more surprising, even GNNs on vascular graphs fail to outperform strong tabular baseline on global features. We consider hypotheses, on both models and data, that could explain this suboptimal performance. Our investigation suggests that, counter-intuitively, vascular graphs may not help predict PE guidelines risk better than simpler clinical biomarkers. Code is available from https://github.com/creatis-myriad/GENESIS.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5113_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/creatis-myriad/GENESIS
Link to the Dataset(s)
N/A
BibTex
@InProceedings{PaiNat_Pulmonary_MICCAI2026,
author = { Painchaud, Nathan AND Habémont, Tristan AND des Ligneris, Morgane AND Serva, Allan AND Croisille, Pierre AND Bertoletti, Laurent AND Lampert, Thomas AND Lutzeyer, Johannes F. AND Merveille, Odyssée},
title = { { Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need } },
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 presents a multimodal pipeline for pulmonary embolism (PE) risk stratification using medical records and CTPA-derived features without relying on blood tests. The authors construct pulmonary vascular graphs from segmented CTPA images and benchmark multiple approaches to combine global clinical features with graph-based representations, including tabular models and graph neural networks (GNNs). The study provides a comprehensive empirical comparison of fusion strategies and reports that vascular graph representations do not improve performance over strong tabular baselines based on global features.
- 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-Clinically relevant problem: The paper addresses pulmonary embolism (PE) risk stratification, which is highly relevant for clinical decision-making, especially given the frequent absence of blood tests in practice. 2-Well-motivated multimodal framework: The work is well-positioned at the intersection of medical imaging and clinical data, combining structured medical records with CTPA-derived biomarkers. 3-Comprehensive benchmarking: The authors perform an extensive comparison between strong tabular models (e. g. , TabPFN, XGBoost) and multiple GNN architectures (GCN, GIN, GPS), including different fusion strategies (early, late, and proposed intermediate fusion methods). 4-Exploration of graph-based representations: The construction of patient-specific pulmonary vascular graphs and their use in GNNs is interesting and technically relevant. 5-Insightful negative results: The finding that vascular graph representations do not improve over strong tabular baselines is valuable and contributes to a better understanding of when graph-based modeling is beneficial. 6-Effort toward interpretation: The authors attempt to investigate several hypotheses (overfitting, model limitations, label noise) to explain the observed results, which strengthens the empirical contribution. 7-End-to-end pipeline: The proposed pipeline, including segmentation, graph extraction, and multimodal fusion, is comprehensive and demonstrates significant engineering effort.
- 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-Overly strong conclusions: The claim that vascular graphs “do not provide discriminative information” appears too strong given the limited dataset size, lack of external validation, and potential noise in graph construction. A more cautious interpretation would be appropriate. 2-Limited dataset size and generalization: The study is conducted on a relatively small private dataset (n=353) from a single source, which raises concerns about generalizability. No external validation is provided. 3-Reproducibility concerns: The dataset is private and no clear commitment to releasing data or code is provided, which limits reproducibility. 4-Dependence on preprocessing quality: The pipeline relies heavily on segmentation (ventricles, vessels, thrombi) and graph extraction, yet there is no quantitative evaluation of segmentation quality or analysis of error propagation. 5-Insufficient analysis of graph representation: While the paper concludes that vascular graphs are not useful, alternative explanations such as suboptimal graph construction, feature design, or insufficient data scale are not fully explored. 6-Potential bias in model comparison: The comparison between pretrained tabular models (TabPFN) and GNNs trained from scratch may be unfavorable to GNNs, especially in the low-data regime. 7-Limited novelty in methodology: While the benchmarking is extensive, most components (segmentation models, GNN backbones, fusion strategies) are based on existing methods, with relatively incremental methodological contributions.
- 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
This is a well-executed and carefully designed empirical study addressing an important clinical problem. The benchmarking effort is thorough and the negative findings are potentially valuable for the community. However, the main conclusions appear stronger than what is supported by the experimental evidence. In particular, the claim regarding the lack of discriminative power of vascular graphs should be toned down or better justified. Additional validation on external datasets, as well as deeper analysis of the graph construction pipeline, would significantly strengthen the 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.
(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 paper addresses an important and clinically relevant problem and presents a comprehensive empirical study comparing tabular and graph-based approaches. The experimental design is generally solid, and the negative findings are interesting. However, the main limitation lies in the strength of the conclusions, which appear disproportionate to the available evidence. The study is conducted on a relatively small private dataset without external validation, and the impact of preprocessing steps (e.g., segmentation and graph construction) is not sufficiently analyzed. These limitations make it difficult to support the claim that vascular graphs do not provide discriminative information in general. While the work has merit, I believe it falls slightly below the acceptance threshold in its current form.
- 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 addressed several of the major concerns raised in the reviews in a clear and constructive manner. In particular, they acknowledged that some of the conclusions were overstated and clarified that their findings are limited to the specific task setting and feature configuration rather than claiming that vascular graphs are generally uninformative. The rebuttal also clarified the rationale behind the chosen endpoint, provided additional details regarding segmentation and graph extraction quality, and addressed reproducibility concerns through code and annotation release plans. While limitations related to dataset scale, lack of external validation, and endpoint formulation still remain, I believe the work provides a valuable and carefully conducted empirical study on an important clinical problem. The negative findings are informative for the community and supported by extensive benchmarking across multiple fusion strategies and graph-based models.
Overall, the rebuttal satisfactorily addressed my primary concerns, and I therefore support acceptance of the paper.
Review #2
- Please describe the contribution of the paper
Instead of proposing new methods, authors try to dig deeper and find out vascular biomarkers do not further improve risk stratification for pulmonary embolism task. They do lots of benchmark experiments to prove this counterintuitive laws. From my perspective, this paper is useful, especially for following work and raising a question that: When we are designing algorithms, is it better to have more features?
- 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 key question authors want to discuss is quite interesting and reasonable. For PE task, how to design features / data for input is critical but might be thought as limitation of models. In this paper, authors put forward a bold hypothesis: Could there be some features that are actually unnecessary? 2.Authors benchmark different methodologies under different hyperparameters to reduce random bias. From my perspective, results are solid. 3.All settings of experiments are clear. Authors also provide code link for reproducibility.
- 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 still have some confused part. They are attached as below: 1.Although authors try lots of basic models and different combination with global features (Sec 2.2, Fig.2), there exist some SOTA models. Do authors think about use SOTA models with vascular biomarkers feature / medical records / cardiac biomarkers as feature resources for validation? For example, A model with feature a / b / c / a+b / … But A is not just basic GNN, using more SOTA model instead. 2.The “Vascular graphs are not all you need” might not be accurate. Do authors want to discuss about vascular biomarker feature, right? It still necessary to get vascular graphs (with nodes and edges) for GNN models.
- 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
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 think authors try to discuss with more basic and more significant question: How we choose our useful features for methods design. Within PE task, they benchmark lots of experiments and verify that some biomarkers, somehow, are not as useful as we think. If it can have some downstream tasks to validate this laws, it will become a new angle and make us rethink our following model design.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After checking all reviewers’ comments, I think this paper does have too strong assumptions. For example, the key question and key expression they propose in title and introduction may mislead readers that vascular information has nothing to do but just hinder models. Also, maybe limited by pages, authors don’t do in-depth analysis on this phenomenon and discuss why it shows so counterintuitive. But although the paper has such flaws, the key findings and experiment results still interest me. Besides, authors do lots of experiments to verify the results are not random bias. From my perspective, if authors will correct their expressions in CR version, it would still be a inspiring paper to read. (But it needs to be very carefully refined and corrected all over assumptions) Besides, the concern about dataset is also a big problem. For example, whether all of the data you use are from same region, ethnics, etc., and whether it will be the reason makes the vascular information hinder the performance of models.
Review #3
- Please describe the contribution of the paper
The paper studies whether patient-specific pulmonary vascular graphs extracted from CTPA can improve pulmonary embolism (PE) risk stratification beyond structured clinical records and CTPA-derived cardiac biomarkers. To address this question, the authors benchmark a range of tabular and graph-based multimodal fusion approaches and report a largely negative result: neither vascular biomarkers nor vascular graph representations provide a measurable improvement over strong tabular baselines built on global clinical features. In addition, the submission describes, with substantial detail provided in the supplementary material, an automated pipeline for pulmonary arterial tree graph extraction, artifact correction, and computation of clinically relevant obstruction scores such as Mastora and Qanadli.
- 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 this submission is that it tackles an important and clinically relevant question in PE management. Risk stratification is central to treatment decisions, and studying whether CTPA-derived vascular structure can compensate for missing blood biomarkers is a meaningful problem.
The paper is also commendable for reporting a negative result. Such studies are valuable for the field when the experimental setting is carefully analyzed, and the authors do make a serious effort to benchmark multiple families of models rather than relying on a single architecture.
Another clear strength is the technical and engineering effort behind the vascular graph construction pipeline. The automated extraction of a patient-specific pulmonary arterial tree representation, together with quality control, artifact handling, and automatic computation of clinical thrombus burden scores, is highly impressive and likely of independent methodological value to the community. In my view, this is one of the most substantial contributions of the submission.
Finally, the empirical study is broad. The authors compare tabular models, several GNN variants, and multiple fusion strategies, including proposed mechanisms for integrating global features into graph models. This breadth makes the paper informative from a benchmarking perspective.
- 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 the problem formulation. The prediction target is ESC guideline-based risk stratification, while the strongest input features already substantially overlap with the criteria underlying that stratification, including sPESI-related clinical variables and RV/LV-derived cardiac information. As a result, the task setup strongly favors global clinical features and makes it difficult to draw a strong conclusion about the intrinsic value of vascular graphs. In other words, the current benchmark is useful for testing whether vascular graphs add value on top of already highly predictive clinical variables, but it is less well suited to supporting the broader conclusion that vascular graphs may not contain discriminative information for PE risk stratification.
Related to this, I do not think the negative result is interpreted with sufficient caution. The experiments support the narrower statement that vascular biomarkers and graph representations did not improve performance in this specific dataset and task setup. However, parts of the manuscript frame this finding more broadly, as if vascular graphs may simply not be informative for PE severity. Given the current endpoint definition and feature set, that conclusion seems stronger than the presented evidence supports.
A second weakness is the presentation of the contributions. The main manuscript is framed primarily as a multimodal fusion benchmark, whereas the most substantial and distinctive technical contribution appears to be the vascular graph extraction and correction pipeline, which is largely deferred to the supplementary material. This creates an imbalance in emphasis: the main text focuses on a negative multimodal ablation study, while the strongest methodological component is under-presented.
A third concern is the discussion of generalization and overfitting in Section 3.2.The argument that the observed generalization gap is acceptable because it is comparable to gaps reported on graph benchmarks such as ZINC or ogbg-molhiv is not convincing. Molecular graphs and patient-specific anatomical vascular trees differ fundamentally in structure, variability, label semantics, and inductive bias. Therefore, cross-dataset comparison of generalization gaps does not provide a strong justification for sample efficiency or robustness in this setting.
More broadly, the paper would benefit from a stronger clinical endpoint or a more targeted analysis of feature dependence. ESC risk category is a guideline-derived surrogate target, and the current feature set already captures much of that guideline logic. If the goal is to demonstrate that vascular morphology and thrombus spatial distribution add clinically meaningful information, endpoints such as ICU admission, hemodynamic deterioration, or continuous biomarker prediction would be more compelling where feasible. If such endpoints are not practical in the current cohort, an analysis excluding features that directly overlap with the guideline definition would still be important to test whether vascular graphs contribute independent signal.
- 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 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
I think the paper would become substantially stronger with a reformulation of the prediction target. The current endpoint is heavily tied to clinical variables already present in the baseline features, which limits the room for vascular graph information to demonstrate incremental value. A more clinically meaningful and less circular evaluation would be to predict a hard downstream outcome such as 30-day mortality, ICU admission, or hemodynamic deterioration, or alternatively to predict continuous biomarkers such as troponin or BNP from imaging and clinical inputs.
A second recommendation is to reposition the paper. The vascular graph extraction, artifact correction, and automated obstruction scoring pipeline appears to be a major contribution and, in my opinion, deserves much more prominence in the main manuscript. Even if the negative result is retained, the paper would read more convincingly if the engineering and methodological innovation were centered rather than treated as background to the benchmark.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I recommend rejection in its current form. The paper addresses an important clinical question and includes a technically impressive vascular graph extraction pipeline, but I am not convinced by the current experimental framing or by the strength of the main conclusion. Because the target label is guideline-derived and the strongest baseline inputs substantially overlap with the criteria underlying that label, the benchmark is biased toward global clinical features. This makes the reported negative result difficult to interpret as evidence against the usefulness of vascular graphs more broadly.
In addition, the manuscript under-emphasizes what appears to be its most original contribution—the automated graph construction and scoring pipeline—while placing most of the narrative weight on a multimodal benchmark whose conceptual setup is, in my view, not yet sufficiently compelling. Overall, I believe the work contains valuable technical components, but the paper would need a clearer problem formulation, a more appropriate endpoint, and a rebalanced presentation before it is competitive for acceptance.
- 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 reviewers and meta-reviewer for their valuable comments and the time they have given our work. We appreciate that negative results, like ours, are difficult but valuable to publish, and feel confident in doing so given our careful, exhaustive analysis and experimentation. Regarding the raised concerns, our point-by-point replies clarify our claims and detail the small adjustments to fix missing context in the camera ready (CR) version. Overall, we believe our paper would interest the community, as it compiles a great amount of methodological and empirical work comparing a broad range of SOTA and novel methods on a clinically relevant problem, and its unexpected conclusions can inform the design of future studies on similar applications.
“Overly strong conclusions” (All reviewers) Most of the underlying concerns with the conclusions are addressed individually below. Still, we agree that our abstract and conclusions phrase the claims too strongly. Our intent was not to suggest “vascular graphs are not useful“ in general (R1 #5), but rather that they do not “contribute discriminative information for PE severity on top of simpler clinical biomarkers” (Conclusion, p.8). We will rephrase these claims in the CR version.
“Problem formulation: overlap between input features and prediction target” (R3) We carefully considered the prediction target when designing the study. We looked at 30-days mortality & ICU admission as a “more clinically meaningful” target, but their prevalence was too low (~4%) for relevant analysis. We also tested hemodynamic deterioration, and observed the same ranking between methods, confirming the conclusions drawn from the guidelines risk target. Because of space constraints, we did not include these results in favor of investigating possible explanations for negative results. We will add a mention of alternative targets when discussing discrepancy between target and real risk (Sec. 3.2, p.8) in the CR version.
“Insufficient data scale / no external validation” (R1 #2,5, R3) While the PERSEVERE dataset’s size may appear modest, it is substantial given its rich data (CTPA with quality-controlled segmentations, clinical records, and reference risk stratification). In fact, PERSEVERE contains 50% more cases than popular CT/CTPA datasets (cf. Limitations, p.8). External validation, while ideal, is left for future work as no other dataset—public or private—currently provides the comprehensive data required for such a comparison.
“Presentation of the contributions: under-emphasis / insufficient analysis of the graph extraction pipeline” (R1 #4-5 & R3) The graph extraction pipeline comes from prior work, and is the subject of another publication currently under review and cited in our submission. Our submission studies multimodal fusion on top of the extracted graphs, and evaluates the graphs’ relevance for downstream clinical applications. Potentially significant errors in graph construction are limited, considering the segmentations’ excellent accuracy and anatomical consistency (0.943/0.817 Dice on arteries and thrombi, respectively, and ~1 MAE on segmental arteries count, cf. Tables VIII and III in supplementary file). In this context, we prioritized analysing likelier causes of suboptimal results in the limited space available.
“Comparisons with more advanced graph models” (R2) The benchmark was designed to include SOTA models for each task, i.e., TabPFN for tabular data, GPS for graph-level tasks. Without specific suggestions of more appropriate SOTA models, we do not think adding more models would change our findings.
“Reproducibility concerns: no clear commitment to releasing data or code” (R1 #3) Unfortunately, the hospital’s data policy does not allow us to publish the dataset. However, to promote reproducibility, we have already released the code (to be published with the paper) and plan to publish the vascular annotations and graphs to help research on downstream tasks.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper studies an important and clinically relevant question in pulmonary embolism risk stratification and presents a comprehensive multimodal benchmark combining CTPA-derived vascular graphs, biomarkers, and structured clinical data. Reviewers appreciated the significance of the problem, the breadth of the empirical comparison, and the substantial engineering effort behind the segmentation and vascular graph construction pipeline. The paper was also positively viewed for reporting a carefully analyzed negative result, namely that graph-based vascular representations did not outperform strong tabular baselines in the current setting.
At the same time, the reviewers raised several important concerns that should be addressed in the rebuttal. The main issue is the strength of the paper’s conclusions. Multiple reviewers felt that the current wording overstates what can be inferred from the experiments: the results support the narrower claim that vascular graph representations did not add value in this specific dataset and task setup, but not the broader conclusion that vascular graphs lack discriminative information more generally. This concern is closely related to the problem formulation, since the prediction target and input features already overlap substantially with guideline-based risk stratification criteria, which may inherently favor global clinical variables over graph-derived representations.
Additional concerns relate to generalization and evaluation scope. The study is based on a relatively small single-center private cohort without external validation, and the heavy dependence on preprocessing quality is not fully analyzed. Reviewers also suggested stronger comparisons with more advanced graph models and a clearer presentation of the graph extraction pipeline, which appears to be one of the paper’s most substantial contributions but is currently under-emphasized in the main text.
Overall, the paper addresses a meaningful question and contains a valuable empirical study, but the interpretation of the negative findings and the framing of the contribution require clarification. I encourage the authors to use the rebuttal to address these concerns carefully.
- 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 presents a multimodal benchmark for pulmonary embolism risk stratification combining CTPA-derived vascular graphs, biomarkers, and structured clinical data. The study is clinically well motivated, the empirical comparison across multiple fusion strategies and graph models is extensive, and the automated vascular graph extraction pipeline represents a substantial engineering contribution.
The rebuttal has meaningfully addressed the reviewers’ primary concerns. Reviewer #1 now explicitly supports acceptance, noting that the authors appropriately softened the scope of their conclusions and clarified that findings are limited to the specific task and feature configuration. Reviewer #2 finds the key results interesting and the experimental verification thorough, and supports acceptance subject to careful refinement of the wording and overstatements. Reviewer #3 remains cautious about the broader claims but acknowledges the technical value of the graph extraction pipeline presented in the supplementary material.
The recurring issue is the overstatement of the negative finding: the results do not support a general conclusion that vascular graphs lack value, but rather a narrower, task-specific finding. This can be corrected in the camera-ready version. The dataset size, single-center limitation, and overlap between prediction target and input features are acknowledged as limitations by all parties.
I recommend accept. For the camera-ready version, the authors must: (1) thoroughly revise the title, abstract, and introduction to accurately reflect the narrow scope of the conclusions, explicitly stating that findings are specific to this task formulation and dataset; (2) add a dedicated discussion on why the overlap between guideline-derived targets and clinical features may favor tabular baselines; (3) bring the vascular graph extraction pipeline from the supplementary material into the main text, as it represents one of the paper’s strongest technical contributions; and (4) explicitly discuss dataset limitations including geographic and ethnic homogeneity as potential confounding factors.
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.
Authors addressed the main concerns raised by the reviewers in the rebuttal. In particular, they clarified and softened the scope of their conclusions. The explanations regarding endpoint selection, graph extraction quality, and reproducibility are reasonable.
Despite limitations remain regarding dataset size and external validation, all reviewers agree that the work still represents a valuable contribution, and that the extensive supplementary material provided is very valuable.
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.
There are 2 Accepts and 1 Reject after rebuttal. Reviewers agreed that the paper addresses an important clinical problem and provides a comprehensive benchmark of tabular, graph, and multimodal approaches for PE risk stratification. The main concern during review was that the conclusions were stated too broadly. The rebuttal clarified that the findings are limited to the current task setting and do not imply that vascular graphs are generally uninformative. This clarification addresses a major concern raised by multiple reviewers. Some limitations remain, including the lack of external validation and the overlap between the prediction target and available clinical features. However, the empirical study is carefully conducted, the negative findings are informative, and the work provides useful evidence for the community. I recommend acceptance.