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

Abstract

Multimodal medical data captures complementary aspects of disease. It provides a comprehensive basis for diagnosis and the identification of biological associations linked to progression or treatment response. Established contrastive learning approaches exploit shared information across modalities, but fall short in exploiting complementary, modality-specific signals relevant for a task. Here, we propose a synergistic multimodal variational autoencoder (VAE) framework that separates task-relevant unimodal and multimodal information from instance-specific variation task-irrelevant, allowing complementary signals across modalities to be leveraged for prediction. The method yields benefits even if modalities are missing at test time. We evaluate the method on a treatment response prediction task using medical imaging and RNA data. Empirical comparison with alignment-based and multimodal VAE state-of-the-art approaches demonstrates improved multimodal prediction performance, and consistent benefits for prediction from a single modality after multimodal training. The code will be made available upon acceptance.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2011_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/cirmuw/SYNPRED

Link to the Dataset(s)

https://github.com/naomifridman/BreastDCEDL

BibTex

@InProceedings{JanIva_SYNPRED_MICCAI2026,
        author = { Janíčková, Ivana AND Tan, Yen Y. AND Helbich, Thomas H. AND Seeböck, Philipp AND Di Marco, Julieta AND Bago-Horvath, Zsuzsanna AND Heber, Ulrike AND Spiegel, Thomas AND Langs, Georg},
        title = { { SYNPRED: A Synergistic Approach to Multimodal Learning for Clinical Prediction } },
        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

    The paper proposes SYNPRED, a multimodal variational autoencoder (VAE) framework that factorises information into task-relevant multimodal, task-relevant unimodal, and task-irrelevant instance-specific components. The framework can be used at test time for downstream prediction even with missing modalities. SYNPRED is evaluated on the binary classification of pathological complete response (pCR) of breast cancer patients from DCE MRI and RNA (gene-expression profiles). Results show improvements over baselines corresponding to different multimodal approaches, especially when modalities are missing at test time.

  • 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 problem addressed in the paper, i. e. fusing modalities that can be factorised into shared and modality-specific information, is quite relevant and affects many real-world healthcare applications.
    • The method is formulated in a generic way and does not rely on application or dataset-specific assumptions. Since it also appears relatively easy to implement, it has the potential to be easily tested on other complementary multimodal applications.
    • The proposed method is tested against multiple baselines (single-modal models + 4 multimodal baselines) that are representative of different approaches in the literature (naïve fusion, alignment-based methods, and multimodal disentanglement).
  • 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) Regarding the proposed method, I feel that some important design choices were not discussed or justified sufficiently: 1.1) p. 4, Multimodal Fusion: The modality-specific latent posteriors are combined using a mixture-of-experts (MoE) approach, i. e. a weighted sum. I find this choice counter-intuitive, as the goal is to extract shared information between the posteriors, and MoE is typically seen as a way to query/route to specific experts. I would have thought a product-of-experts (PoE) to be a more natural fit, since a product intrinsically models where all latent posteriors are dense. PoE is even used later in the pipeline to fuse latents for downstream predictions. Have the authors considered using PoE to combine latent posteriors, and if so why did they prefer to use MoE? 1.2) p. 4, Multimodal Fusion: The original MoE paper [R1] suggested gating networks to predict data-dependent mixture weights. However, more recent MoE-based multimodal VAEs [23,24] have gone the other direction, with fixed mixture weights corresponding to the arithmetic mean of the latent posteriors. This has been done to improve robustness and to avoid the model collapsing to a single modality. The learnable, but data-independent, mixture weights proposed here fall somewhere between these two approaches. Have the authors looked at the values of the learned mixture weights, to see whether they are balanced or if they suffer from modality collapse? 2) While the evaluation shows improvements from the proposed method over existing baselines, some of the claims made by the authors do not appear to me to be properly justified by the results. 2.1) The authors propose their method broadly for “medical modalities with limited semantic overlap” (p. 2, Contribution). However, the method is only tested on one dataset (BreastD-CEDL) with some apparent limitations (see comment #2.2). Since the authors reference other works combining radiomics and radiogenomics for non-breast cancer applications [12,26], why didn’t they test their method on the application/dataset from one of these papers? Either the evaluation should be extended, or the broad claims should be narrowed to the tested application to avoid generalizing too much from the currently limited evaluation. 2.2) p. 6, Table 1: It seems that RNA is a discriminative modality on its own, and that there is little benefit to adding IMG to it at test time. While no statistical significance results are reported, given the provided confidence intervals, I would guess multimodal SYNPRED lin. prob. is not significantly better than RNA-only SYNPRED ft. The same pattern appears on multimodal baselines. This observation does not invalidate improvements from the proposed method over baselines, but it does put some conclusions in perspective. As such, the proposed method apparently leverages multimodal data during training to better align representation for downstream prediction (c. f. improved single-modality results over unimodal and multimodal baselines). However, multimodal data not improving classification at test-time suggests BreastDCEDL is a poor benchmark of synergistic interactions between modalities (c. f. when properly pretrained, RNA is sufficient on its own to solve the task). The ablation study from Table 2 further supports this interpretation. Removing the fusion latent component does not significantly impact RNA-only and multimodal versions of the model (which can rely on the RNA component), but it affects the IMG-only version (by no longer being able to align with the RNA latent during training). Thus, claims about synergistic effects should be revised to specify train-time effects, as test-time synergistic effects have not been shown yet. 2.3) p. 8, Discussion: “This work addresses a central limitation of existing multimodal learning approaches for medical data. ” CoMM [5] (often cited in the paper) proposes an alignment-based contrastive strategy addressing the same limitation in prediction-oriented frameworks. The advantage of the proposed method is that it tackles this limitation in VAEs, a generative framework. Thus, I think the claim should be clarified to be in the context of generative models.

    References [R1] R. A. Jacobs et al. , “Adaptive Mixtures of Local Experts,” in Neural Computation, March 1991, doi: 10.1162/neco. 1991.3.1.79.

  • 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

    1) p. 2, Related Work: “complementary information [. . . ] can induce synergy in both multimodal prediction and single-modality representations after multimodal training” Synergy, as defined in [5], is the interaction between modalities that bring different and complementary information. Thus, by definition, I don’t understand what the authors mean when they say synergy can be induced in single-modality representations. I would suggest they rephrase/clarify this assertion. 2) p. 4, Section 2.1: $x$ has two definitions in the mathematical notation, leading to confusion. $x$ is used for inputs, which is fine and conventional in the field, but then $x$ is also used to describe modalities whe used in subscript, e. g. $x_x$ or $z_x$. This is needlessly confusing, and I would suggest using a different notation for one of the meanings of $x$, e. g. $m$ for modalities like in [24]. 3) p. 4, Section 2.1: In the “Latent Decomposition and Reconstruction” section, there is quite a bit of redundancy between equations and the text, especially the paragraph after Eq. (1). Both are clear on their own, so the authors could shorten this section. 4) p. 5, Eq. (6): Have the authors considered using two different \beta, one for each KL-based terms? Considering that $L_{KL-inst}$ and $L_{KL-task}$ enforce very different regularisation (KL-inst being a classic global prior, and KL-task encouraging a cluster per combination of modality and target label), my intuition is that the method could benefit from allowing each regularisation to be weighted independently. 5) p. 6, Comparison with SOTA: Why was ContIG [25] chosen as the alignment-based pretraining baseline? CoMM [5] appears equaly applicable, and was even cited multiple times in the introduction to justify the limitations of current limitations of redundancy-oriented contrastive learning approaches. Were there specific reasons the authors chose not to compare to CoMM instead? 6) p. 6, end of 1st paragraph: “Following prior work [24], multimodal performance for the SOTA methods is reported as the average of single-modality results; [. . . ]” Tables 1 and 2 felt quite self-explanatory, so the quoted excerpt confuses me more than anything. It seems to me that single-modality results are reported independently (IMG-only and RNA-only columns), so I don’t see which results correspond the “average of single-modality results”. Can the authors clarify this sentence? 7) p. 7, Analysis of Explainability: This evaluation is tangential to the main message of the paper and does not really affect the rest of the paper, so I will not comment on it too much. Still, there are not enough details about the methods used to generate explanations to properly evaluate this section of the paper. The most details are that these are “saliency-based explanations”. I would suggest authors revise this section, potentially to give a bit more details about how the saliency maps were generated. 8) Missing words and typos throughout the paper: 8.1) p. 1, 2nd paragraph: Missing article: “we propose multimodal generative framework” -> “we propose a multimodal generative framework” 8.2) p. 4, Latent Decomposition and Reconstruction: Missing word: “d_t is the dimensionality of each task-relevant and […]” -> “d_t is the dimensionality of each task-relevant latent and […]” 8.3) p. 4, Latent Decomposition and Reconstruction: In the sentence beginning with “The decoder […] is reconstructing”, $z^{uni}$ in the equation is missing the modality subscript x. It should be $z^{uni}_x$. 8.4) p. 4, Multimodal Fusion: “multmimodal” -> “multimodal” 8.5) p. 6, Analysis of Explainability: “prediciton” -> “prediction” 9) Reference [5] has been published in the proceedings of ICLR 2025 (available online at: https://proceedings. iclr. cc/paper_files/paper/2025/hash/108030643e640ac050e0ed5e6aace48f-Abstract-Conference. html) and should be referenced as such over the arXiv version.

  • 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?

    Even if the method is evaluated only on one specific application, its formulation is generic and simple to implement. Thus, it could be beneficial to the community to be able to test the framework on a broad variety of real-world applications, to potentially better understand its limitations and build on this foundation. However, in its current form, the paper does not properly justify some design choices and overreaches in its claims given the evaluations performed. These claims could be true, but at this stage they are not the only plausible conclusions given the provided results. Thus, the authors should better motivate some design choices, and revise some claims to acknowledge limitations in the application evaluated, before the paper can be accepted.

  • 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 most critical concerns pointed out in the reviews were related to 1) aggregation strategy, 2) synergy definition, and 3) the limited dataset. In response, the authors clarified in their rebuttal that: 1.MoE is more robust than alternative aggregations in settings where modalities are missing; 2.There is clinical interest in a model that leverages synergy at training time to improve performance in unimodal settings at test time; 3.They will “moderate statements regarding general applicability.” Still, BreastDCEDL combines three cohors, making it multi-center, multi-manufacturer, and substantially bigger than other radiology-genomics datasets.

    These clarifications answered my concerns satisfactorily enough to recommend accepting the paper, on the ground that the paper’s methodology is generic enough to be of interest to the community.

    Still, I will insist that “statements regarding general applicability” must be carefully revised. The authors justified BreastDCEDL as the best option for radiology-genomics applications, but did not consider other paired modalities (e.g. images and tabular data) for which large datasets are available [R1] and applicable to the proposed model. Thus, I suggest authors focus on the elements put forward in their answer to “Synergy definition and RNA dominance”. Namely, the proposed model can leverage multi-modal training data to improve unimodal test-time performance, a clinically relevant problem illustrated on a radiology-genomics application. [R1] provides an example of a paper framed similarly to what I suggest, and tested on a different medical dataset.

    References [R1] Hager et al., “Best of both worlds: Multimodal contrastive learning with tabular and imaging data”, CVPR, 2023.



Review #2

  • Please describe the contribution of the paper

    This paper proposes SYNPRED, a multimodal VAE framework that explicitly disentangles the latent space into three components: shared task-relevant (fusion), modality-specific task-relevant (uni), and instance-specific (inst) factors. Task-conditioned priors are used to shape the task-relevant latents, while mixture-of-experts and product-of-experts are employed for multimodal fusion and prediction, respectively, enabling inference under missing modalities. The method is evaluated on the BreastDCEDL dataset with paired DCE-MRI and gene expression data for breast cancer pCR prediction. The authors report improved multimodal AUROC and notable gains in unimodal (especially imaging) prediction after multimodal training, which they interpret as evidence of cross-modal synergy.

  • 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 addresses an important and underexplored setting where modalities share limited semantic overlap. The argument that alignment-based methods suppress complementary signals is clearly articulated. 2.The three-way latent factorization (fusion/uni/inst) with task-conditioned priors is simple, intuitive, and well-justified. 3.MoE for training and PoE for prediction allow test-time inference with incomplete modalities without retraining. 4.Image-only AUROC improves from 0.549 to 0.693 after multimodal training, which is clinically relevant since imaging is far more available than gene expression in practice. 5.Public multi-center dataset, clinically meaningful endpoint (pCR), five-seed runs with CIs, component-wise ablation, and saliency-based explainability analysis.

  • 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 MoE fusion of Gaussian posteriors uses simple weighted averages of means and standard deviations, which is not a standard approach for combining Gaussians and could misrepresent uncertainty. It would be helpful to discuss why this was chosen over more principled alternatives such as moment matching or MoPoE. 2.The KL regularization weight beta is very small (10^-4), which suggests reconstruction loss dominates training. This raises the question of how much the task-conditioned KL terms actually shape the latent space. An ablation over beta would help clarify this. 3.The synergy claim is central to the paper but is supported only by downstream performance gains. Some form of information-theoretic quantification (e.g., PID or conditional MI) would make this claim more concrete. 4.The unimodal image baseline (convolutional VAE, AUROC 0.549) appears weak relative to discriminative models available on BreastDCEDL (e.g., ViT-based). Including stronger baselines would help better calibrate the observed gains from multimodal training. 5.Multimodal performance for SOTA baselines is reported as the average of single-modality results, which is unconventional. Genuine multimodal inference for all baselines would improve the fairness of comparison. 6.Linear probing outperforms fine-tuning in the multimodal setting, which is surprising and not discussed. This warrants some analysis.

  • 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?

    The paper tackles a well-motivated problem and proposes a clean disentanglement design that yields promising results, especially the single-modality gains for imaging. However, some technical choices need better justification (MoE fusion, very small beta), the synergy claim lacks formal quantification, and the baselines could be stronger. These concerns are addressable in a rebuttal, so I lean toward 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



Review #3

  • Please describe the contribution of the paper

    The paper proposed a multimodal VAE-based framework for clinical prediction that disentangles three latent factors: shared task-relevant information, modality-specific task-relevant information, and task-irrelevant instance-specific variation. The standard alignment-based multimodal learning underuses cross-modal information, there explicit factorization enables “synergistic” prediction gains. The methods are evaluated using paired breast DCE-MRI and gene expression data for pCR prediction.

  • 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.address a clinical limitation in multi-medical learning: not all modalities are aligned and forcing cross-modal similarity may suppress complementary signal. This is an important question, especially for an imaging-genomics setting. 2.The decomposition into fusion, modality-unique, and instance-specific latent variables is intuitive and technically well-structured. The use of task-conditioned priors to shape the task-relevant latent space is a meaningful design choice over other superficial architecture tweaks. 3.Good performance gain, such as multi-modal AUROC is 0.725, exceeds the baseline methods, and the ablation studies also support this claim.

  • 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 paper is evaluated on essentially one application setting: breast MRI plus RNA for pCR prediction. This is too narrow to fully support the paper’s broader claims about multimodal medical learning with limited semantic overlap. 2.The paper repeatedly uses the term synergy, but the evidence is mostly indirect: improved multimodal performance and improved unimodal performance after multimodal training. That is suggestive, but not a rigorous demonstration. The authors should define what operationally constitutes synergy and distinguish it from ordinary regularization, auxiliary-task benefit, or representation transfer. 3.The comparison includes ContIG, MMVM VAE, DMVAE, and concatenation. This is reasonable, but still not comprehensive enough. It would be helpful to compare against stronger contemporary multimodal fusion baselines, including attention-based fusion or transformer-style fusion models, and perhaps a supervised multimodal encoder without a generative objective. Otherwise, it is not fully clear whether the gains come from disentanglement specifically or simply from a better training setup.

  • 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.

    (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’s novelty and clinical impact

  • 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



Author Feedback

We thank the reviewers for constructive feedback. SYNPRED enables synergistic gains under partially observed modalities through disentangled shared, modality-specific, and instance-specific latent components. Multimodal pretraining improves the weaker but clinically deployable imaging modality from AUROC 0.549 to 0.638 (0.693 after fine-tuning). The contribution of synergy is isolated by a drop from 0.638 to 0.603 AUROC when ablating the fusion loss, while leaving RNA-only essentially unchanged. This demonstrates the impact of task-conditioned shared latents rather than generic multi-modal regularization. Aggregation Strategy & KL (R1.1, R2.1). We use MoE rather than PoE-based fusion during representation learning because of its greater robustness under missing modalities [Daunhawer et al., 2021]. While MoPoE improves robustness compared to standard PoE, it still relies on product-based aggregation. Learned MoE weights remained balanced (~0.5), indicating no modality collapse. The fused (μ, σ) serves as a moment-matched summary used only to drive L_KL-task, not a generative density. We will clarify these points in the revised manuscript. In pretraining, a small shared β avoided posterior collapse while preserving active KL regularization, as reflected by strong frozen linear-probing performance. Synergy Definition and RNA Dominance (R1.2.2, R2.3, R3.2). We evaluate empirical rather than formal information-theoretic synergy in real-world multimodal data. Synergy in [5] is defined through complementary cross-modal interactions in a synthetic-data setting. Since the underlying generative factors are unknown in real-world datasets such as BreastDCEDL, we follow [23] and assess synergy empirically through improved unimodal representations learned across multimodal observations. R1 notes that RNA-only is close to multi-modal prediction, an accurate diagnosis of where synergy matters clinically: imaging is routinely acquired, RNA rarely is. Multimodal pretraining improves image-only AUROC from 0.549 to 0.693, showing molecular information improves the weaker imaging modality via synergy even when RNA is unavailable at inference time. The ablation study further isolates this effect: removing the shared fusion latent selectively degrades image-only performance, suggesting the gains arise specifically through the shared fusion latent rather than from disentanglement or regularization alone. Dataset Scope (R1.2.1, R3.1). We agree a broader evaluation would strengthen claims and will moderate statements regarding general applicability accordingly. BreastDCEDL nonetheless combines three independent breast cancer cohorts, yielding a multi-center, multi-manufacturer dataset with meaningful clinical heterogeneity. Publicly available paired radiology-genomics datasets remain very limited. The datasets [12,26] are substantially smaller. Baselines and Evaluation (R1.5, R1.6, R2.4, R2.5, R3.3). We agree that this needs clarification in the manuscript: (1) baselines were chosen to compare contrastive alignment, disentanglement, multimodal regularization, and naive fusion effects. (2) R1.5: we agree CoMM is the more appropriate alignment baseline given our framing; ContIG was chosen for its radiogenomics specificity, and both belong to the same alignment family. We will clarify this rationale. (3) multimodal evaluation (R1.6, R2.5) needs clarification: the baselines report modality-specific evaluation and lack native multimodal evaluation under missing modalities. Following [24], we therefore report multimodal performance as an average across modality-specific evaluations. (4) a stronger image backbone (e.g., ViT; R2.4) could improve performance, but our focus was to investigate multi-modal learning effects under matched encoders Replicability (R3) Code, configurations, preprocessing scripts, and a reproducibility appendix will be released upon acceptance. Notation, missing subscript, typos, and CoMM ICLR’25 citation will be corrected. (R1)



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.

    Reviewers found the evaluation on one specific application to be problematic and suggested that the paper may be overreaching in its claims. Please prioritize addressing these concerns, including strengthening the baseline comparisons to clarify whether the gains come from disentanglement specifically or from a stronger training setup, and more clearly defining or formalizing the claimed synergy.

  • 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 paper proposes SYNPRED, a multimodal variational autoencoder framework that factorizes information into task-relevant multimodal, task-relevant unimodal, and task-irrelevant instance-specific components. The work is well motivated and addresses an important problem in multimodal medical prediction, and several reviewers have acknowledged the potential value of the proposed formulation. The most critical concerns raised in the reviews concerned the aggregation strategy, the definition of synergy, and the limited dataset setting. The authors’ rebuttal provides reasonable clarifications on these points: they explain that the MoE aggregation is more robust than alternatives when modalities are missing, motivate the clinical relevance of using multimodal synergy during training to improve unimodal performance at test time, and commit to moderating claims regarding general applicability. While the concern about generalizability is not fully resolved, the BreastDECDL dataset combines 3 cohorts, making it multi-center, multi-manufacturer, and substantially larger than many radiology-genomics datasets. The authors’ planned clarification in the final revision appears sufficient. Overall, the rebuttal addresses the main reviewer concerns to an acceptable degree, and the remaining limitations do not outweigh the paper’s novelty and potential impact. The paper is therefore recommended for acceptance.



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.

    After carefully reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors directly addressed key concerns including justification of MoE over PoE for robustness under missing modalities, clarification of empirical synergy evaluation with ablation isolating the fusion latent’s contribution, moderation of generalizability claims, and improved unimodal imaging performance (AUROC 0.549→0.693) demonstrating clinical relevance.



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 some flaws in the experiments, the novelty and motivation of the paper are recognized by all reviewers.



back to top