List of Papers Browse by Subject Areas Author List
Abstract
In this paper, we address the problem of multimodal federated learning with missing modality. Existing methods utilize an additional public dataset or perform naive feature synthesis that is based solely on the available modality. To address these limitations, we propose ProMoE-FL, a Prototype-conditioned Mixture-of-Experts framework for robust missing-modality feature synthesis in multimodal federated learning. ProMoE-FL builds a global client-aware prototype bank that captures clinically meaningful modality priors across institutions. Our Mixture of Experts is conditioned on these prototypes and modality indices to enable direction-aware expert routing for dynamically synthesizing missing features. We perform extensive quantitative and qualitative evaluations on four public chest X-ray datasets (MIMIC-CXR, NIH Open-I, PadChest, and CheXpert) and demonstrate that ProMoE-FL consistently outperforms state-of-the-art methods in both homogeneous as well as the more challenging heterogeneous settings. Code available at: https://github.com/bhattarailab/ProMoE-FL.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1724_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/bhattarailab/ProMoE-FL
Link to the Dataset(s)
MIMIC-CXR-JPG: https://physionet.org/content/mimic-cxr-jpg/2.1.0/
NIH Open-I: https://www.kaggle.com/datasets/raddar/chest-xrays-indiana-university
CheXpert: https://stanfordmlgroup.github.io/competitions/chexpert/?utm_source=catalyzex.com
PadChest: https://bimcv.cipf.es/bimcv-projects/padchest/
BibTex
@InProceedings{ChhAav_ProMoEFL_MICCAI2026,
author = { Chhetri, Aavash AND Niroula, Bibek AND Vazquez, Eduard AND Shrestha, Yash Raj AND Gyawali, Prashnna AND Bazzani, Loris AND Bhattarai, Binod},
title = { { ProMoE-FL: Prototype-conditioned Mixture of Experts for Multimodal Federated Learning with Missing Modalities } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes ProMoE-FL, a prototype conditioned Mixture of Experts framework for multimodal federated learning with missing modalities. It constructs a client aware global prototype bank to capture cross institutional modality priors, and leverages a direction aware MoE routing mechanism to dynamically synthesize missing features without public data. Extensive experiments on four chest X-ray datasets show that ProMoE-FL consistently outperforms SOTA methods in both homogeneous and heterogeneous settings, with high quality synthesized features that preserve critical clinical information.
- 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.It has strong robustness and can effectively solve the data heterogeneity problem in federated scenarios. ProMoE-FL introduces a client-aware global prototype library, which can capture cross-institutional modal priors and adapt to the data differences of different medical centers. In the heterogeneous scenarios of real clinical settings, its performance is significantly better than that of traditional interpolation methods, and it solves the problem of poor generalization of traditional methods. 2.It does not rely on public data, and is more practical. Unlike existing methods such as CAR-MFL, ProMoE-FL only relies on private data from each client for training, without the need for additional public data. This avoids the problems of scarce public data and restricted privacy in medical scenarios, and better meets the actual needs of clinical federated learning. 3.The feature synthesis quality is high, and key clinical diagnostic information is retained. Through prototype condition constraints and a hybrid expert routing mechanism, the features synthesized by ProMoE-FL are highly consistent with the real features, can retain key clinical structures such as lung lesions, and avoid feature distortion caused by simple interpolation, providing support for diagnostic tasks.
- 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 prototype bank relies heavily on high-quality multimodal client data. If multimodal clients are scarce, poorly annotated, or suffer from modality bias, the quality of the global prototype bank will degrade, which in turn hurts model performance. 2.The MoE structure increases model complexity and communication overhead. Although the expert design avoids parameter redundancy, it is more complex than simple imputation methods and requires extra coordination between experts and the prototype bank during federated training, raising communication costs. 3.The model’s adaptability to extreme missing-modality scenarios is not fully validated. Experiments only cover conventional missing modality settings. In extremely unbalanced cases, the prototype bank may fail to learn reliable cross-modal relations, and the generalization ability remains untested.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(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?
Overall Recommendation: Weak Accept This paper proposes ProMoE-FL, a federated learning framework to address missing modalities in multi-modal medical data across distributed healthcare institutions. The problem is highly relevant to clinical AI, and the core idea of combining client-aware prototypes with a prototype-conditioned MoE architecture (without public data) is novel and technically sound. Experiments on four chest X-ray datasets show consistent improvements over state-of-the-art baselines. Strengths Addresses a critical real-world challenge in federated multi-modal medical AI. Novel, privacy-compliant architecture suitable for clinical deployment. Rigorous experiments and ablation studies validate effectiveness. Strong practical value for multi-center clinical AI deployment. Weaknesses & Required Revisions Lack of convergence analysis for prototype updates and MoE routing. Missing hyperparameter sensitivity analysis in ablations. Clarity issues in introduction, methodology, and related work. Limited generalization beyond chest X-ray data. No computational/communication overhead analysis. Conclusion: The paper presents a novel and impactful solution with convincing results, but requires significant revisions to address the above limitations. I recommend a Weak Accept , contingent on the authors addressing these concerns.
- 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.
I have reviewed the authors’ rebuttal and the cross-review discussions. Recognizing MICCAI’s policy that prohibits uploading new figures or extensive experimental data during the rebuttal phase, I accept the authors’ textual clarifications regarding hyperparameter robustness and the lightweight nature (~0.3% additional parameters) of the routing mechanism. The core motivation of using a prototype-conditioned MoE to address missing modalities in non-IID clinical settings remains clinically relevant. While some deeper technical evaluations (such as rigorous empirical convergence curves and comprehensive evaluation beyond chest X-rays) are left as text promises for the final version, the current manuscript is sufficiently validated to provide a marginal contribution. Therefore, I decide to change my rating to a Weak Accept.
Review #2
- Please describe the contribution of the paper
This paper addresses multimodal federated learning under missing modalities in medical imaging-report settings. Its main contribution is a prototype-conditioned mixture-of-experts framework, ProMoE-FL, which synthesizes missing modality features by combining observed modality features with a global client-aware prototype bank. The method is designed to better handle cross-client heterogeneity and missing modalities without requiring additional public data. The paper also presents experiments on multiple chest X-ray/report datasets under both homogeneous and heterogeneous client partitions.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a relevant and practically important problem in medical federated learning: missing modalities across institutions. This is a realistic setting in clinical practice, where different sites often have incomplete access to images, reports, or other modalities.
Methodologically, the paper proposes a coherent combination of prototype-based conditioning and mixture-of-experts routing. The idea of using a client-aware prototype bank to guide feature synthesis is reasonable and well-motivated for heterogeneous federated settings. The use of a shared MoE design is also sensible from a modeling perspective, as it aims to improve flexibility while avoiding separate generators for each modality pair.
The experimental section is relatively broad. The paper evaluates the method on four public chest X-ray datasets and includes both homogeneous and heterogeneous federated partitions, which is stronger than a single-dataset or single-split evaluation. The results are generally consistent and show meaningful gains, especially in the heterogeneous setting, which is arguably the most relevant scenario.
I also appreciate that the method does not rely on an external public multimodal dataset by default, unlike some prior approaches. This makes the setting more realistic for healthcare. In addition, qualitative analyses of synthesized features and rare pathologies are useful and support the empirical findings.
- 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 main weakness is that the empirical validation is still not up to the standard expected for a top-tier conference. While the reported downstream AUC improvements are promising, the paper provides only limited direct evidence that the synthesized missing-modality features are truly faithful and clinically meaningful. The t-SNE/UMAP visualizations are useful, but they are still indirect. A more rigorous quantitative evaluation of synthesis quality would substantially improve the paper.
A second concern is the reporting of results. The paper states that results are averaged over three random seeds, but standard deviations are reported only for the proposed method, not for competing baselines. This makes it difficult to judge the statistical significance and fairness of the comparison. In addition, the evaluation relies almost entirely on macro-AUC. In a medical multi-label setting, it would also be preferable to report other metrics, such as PR-AUC, F1, sensitivity/specificity, or calibration-related measures.
A third weakness is that the computational and communication costs are not adequately analyzed. Since the method introduces multiple experts, routing, prototype banks, and extra modules beyond the base federated model, the paper should report parameter counts, training/inference overhead, and server-client communication cost. This is particularly important in federated learning, where communication efficiency is a central consideration.
Relatedly, although the paper presents gating visualizations suggesting balanced expert usage, it does not provide an explicit mechanism, regularizer, or stronger quantitative analysis to rule out expert collapse or degenerate routing behavior. The current evidence is suggestive, but not yet fully convincing.
The ablation study is also limited. The paper compares MoE against a single PCD in a single heterogeneous setting, which is helpful, but additional ablations would strengthen the claims. For example, it would be useful to isolate the effect of the prototype bank, the accumulation strategy, the number of experts, and the routing design.
A major weakness is that the proposed method appears substantially closer to prior feature-imputation approaches than the paper currently acknowledges, particularly “Multimodal Federated Learning With Missing Modalities through Feature Imputation Network”. At a high level, both works follow the same core paradigm: learning cross-modal feature completion from multimodal clients and using the synthesized bottleneck features to support unimodal clients during federated training.
In the current submission, the main differences lie in the introduction of a prototype-conditioned mechanism and a mixture-of-experts routing strategy. However, these additions appear to extend rather than fundamentally redefine the original framework. As a result, the contribution may be better characterized as an architectural enhancement of the Feature Imputation Network paradigm rather than a conceptually distinct approach.
This impression is further reinforced by the highly similar visual presentation. The overall pipeline figure adopts a closely related graphical style, module arrangement, and narrative decomposition (e.g., a federated system overview, multimodal vs. unimodal client training, and a dedicated feature synthesis block between the encoders and the classifier). Such similarity in both methodology and presentation makes the novelty claim less convincing. The authors should more clearly articulate how their method fundamentally differs from and advances beyond the Feature Imputation Network framework.
Finally, the method’s stress testing is not sufficiently comprehensive. The client splits still retain some multimodal clients. An extreme setting in which all or nearly all clients are single-modality would better test the approach’s robustness and limits. This would be especially relevant for the paper’s practical claims.
On clarity, there are also a few minor presentation issues. For example, “non-IID” is introduced without an explicit definition, and several implementation details remain underspecified. The manuscript also contains minor presentation inconsistencies, such as the inconsistent capitalization of PmcmFL in the main text/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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
This is a promising paper on an important problem, and I think the proposed direction is meaningful; however, my main recommendation is to strengthen the empirical section rather than substantially changing the core idea. To achieve this, the paper would benefit from more complete statistical reporting for all baselines, additional evaluation metrics beyond AUC, direct quantitative validation of synthesized feature quality, efficiency, and communication analysis, as well as stronger ablations and stress tests under more extreme missing-modality settings. In addition, the paper would benefit from a clearer positioning with respect to prior work, especially “Multimodal Federated Learning With Missing Modalities through Feature Imputation Network.” The current submission appears to build on a similar overall feature-imputation pipeline while introducing prototype conditioning and MoE-based routing; therefore, clarifying more explicitly what is fundamentally new, as opposed to what is an extension of the earlier framework, would help make the contribution more convincing. Finally, I encourage the authors to improve the clarity of the presentation by addressing minor inconsistencies, such as the capitalization of PmcmFL in the main text and table, defining terms such as non-IID at first mention, and clarifying several implementation details of the router and expert configuration.
- 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?
I find the problem addressed in this paper important and practically relevant, particularly for multimodal federated learning under missing-modality settings in healthcare. The proposed framework is technically sound, and the idea of combining prototype conditioning with mixture-of-experts routing is interesting. The experimental results are generally consistent and demonstrate improvements over several baselines, especially under heterogeneous settings, which is a notable strength.
At the same time, my score reflects some reservations regarding novelty and experimental completeness. From a methodological perspective, the proposed approach appears closely related to prior feature-imputation-based methods, particularly “Multimodal Federated Learning With Missing Modalities through Feature Imputation Network.” While the introduction of prototype conditioning and MoE routing extends this paradigm, the overall pipeline and core formulation remain similar, and the paper would benefit from a clearer articulation of what constitutes the fundamental novelty beyond these extensions.
This impression is further reinforced by the similarity in the overall framework design and visual presentation (e.g., pipeline structure and module organization), which makes it somewhat challenging to clearly distinguish the work conceptually from prior approaches.
In addition, the experimental section could be strengthened. The evaluation relies primarily on AUC, with incomplete statistical reporting across baselines. There is also limited direct quantitative validation of the synthesized feature quality, and no analysis of computational or communication overhead, which is important in federated settings. Furthermore, ablation studies and stress tests under more extreme missing-modality scenarios are relatively limited.
Overall, despite these limitations, I believe the paper presents a meaningful extension of existing approaches with promising empirical results. With clearer positioning of novelty and stronger experimental validation, it could become a solid contribution.
- 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.
After carefully reading the rebuttal, I appreciate that the authors responded seriously to most of the concerns and clarified several points that were previously underdeveloped. In particular, the rebuttal improves the positioning relative to prior feature-imputation methods by emphasizing the role of the global prototype bank, client-aware prototypes under non-IID settings, and the dynamic MoE routing mechanism. While I still view the work as methodologically close to the Feature Imputation Network paradigm rather than a fundamentally new formulation, I now think the paper provides a clearer case that the proposed extensions are meaningful and practically relevant rather than purely incremental.
I also appreciate the clarification regarding communication overhead and the explanation that the additional modules account for only a very small fraction of total parameters. The discussion about scalability with modality count is helpful and addresses part of my earlier concern regarding federated efficiency. Likewise, the commitment to include standard deviations for all baselines in the revision significantly improves the fairness and interpretability of the empirical comparisons.
The rebuttal also partially addresses my concerns about robustness under heterogeneous settings. The fact that the paper evaluates nine client configurations, including highly constrained multimodal-client ratios such as 8:0:2 and 0:8:2, strengthens the empirical validation relative to prior work. I still believe that even more extreme scenarios with fully unimodal federated populations would be valuable, but I no longer consider this omission fatal for the current submission.
That said, some weaknesses remain. The paper still lacks sufficiently direct quantitative evidence that the synthesized features are semantically faithful beyond downstream classification performance. The current justification that good synthesis should implicitly improve AUC is reasonable, but stronger dedicated synthesis-quality evaluations would make the claims more convincing. Similarly, the analysis of expert routing behavior and collapse prevention remains somewhat qualitative. The rebuttal acknowledges that deeper routing analysis is future work, which is fair, but this still leaves part of the MoE contribution insufficiently characterized.
Overall, although I continue to think the work is closer to an architectural enhancement of existing feature-imputation federated learning frameworks than a fundamentally new paradigm, the method appears technically sound, practically motivated, and empirically stronger than I initially assessed. The rebuttal successfully resolves several of my concerns regarding fairness, robustness, and implementation clarity. I therefore revise my choice to Weak Accept.
Review #3
- Please describe the contribution of the paper
1.Propose a novel federated medical image classification method specifically under missing modality situation. 2.Develop Prototype Construction and Alignment, Prototype Conditioned Decoder to improve the performance on local clients. 3.Achieves state-of-the-art performance on four chest-xray classification datasets.
- 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 is a timely research in federated medical image classification with missing modality setting. 2.The paper is easy to follow, with clear flowchart to understand.
- 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.Why in Homogenous setting, the use of 6:0:4 leads to a lower performance compared to 4:0:6, while in Heterogenous setting, the use of 6:0:4 provides a higher performance compared to 4:0:6? This discrepancy should be explained and discussed. Does this mean multi-modal contributes more to the performance in data heterogeneity? 2.In Quantitative Results, the reason for considering 8:0:2 heterogeneous setting to ablate MoE should be clarified. For example, why not consider other settings? 3.Since the study uses image with text encoders, it is possible to compare with CLIP-based FL approaches, such as FACMIC [1] and pFedST [2]? 4.The proposed approach consists of several modules, e.g., PCD, MoE. I suggest the authors to provide some information about computation overhead to see if its efficient or no. 5.In the manuscript, it mentioned that the results are averaged over three random seeds. However, why in Table 1, only the proposed method has STD? Since all methods are averaged over three random seeds, the comparison should be consistent. 6.In Qualitative Results, To the best of … I dont understand why put this statement in t-SNE and UMAP visualization. From the image I can not get the same conclusion. Please explain it.
[1].Facmic: Federated adaptative clip model for medical image classification. [2].Personalized Federated Side-Tuning for Medical Image Classification.
- 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.
(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 overall contribution is satisfactory for MICCAI society, with some flaws.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors rebuttal addressed the comments, and i prefer to hold “Weak Accept”.
Author Feedback
We thank all reviewers for recognizing the clinical relevance(R1-R3), technical soundness(R2), rigorous validation(R1, R2), practical value(R1, R2), novelty(R1) and clarity(R3) of our work. Clarification of Novelty [R2]: Like FeatImp([21]), ProMoE-FL operates in feature space and synthesizes missing-modality features, but is fundamentally distinct in robustness and design. Rather than relying solely on present features, we guide synthesis via a global prototype bank as a federation-wide prior capturing complementary modality information. Client-aware prototypes handle non-IID shifts, mitigating client drift (Tab.1, hetero.). Our MoE architecture dynamically routes transformations per instance, avoiding rigid static mappings. Deviations and More Metrics [R2, R3]: We follow the evaluation setup and AUC metrics of prior works ([21, 22]). We report std. for ProMoE-FL to demonstrate stability across random seeds; notably, our lower bound consistently surpasses the mean of the strongest baseline e.g., in 8:0:2 hetero., our lower bound of 78.16 exceeds FeatImp’s mean of 77.15.Further, our avg std. of 0.78 is lower than FeatImp’s 0.88, showing greater stability. Per reviewer feedback, we will include standard deviations for all baselines upon revision. Communication Overhead [R1-R3]: The only additional parameters are the lightweight prototype bank and routing mechanism, comprising ~0.3% of total parameters. MoE also enables favorable scaling with modality count rather than combinatorial growth. We will clarify this in the final version. Multimodal Reliance & Extreme Missing Modalities [R1, R2]: Our client-aware prototype bank incorporates unimodal client prototypes, mitigating reliance on multimodal clients. Following prior works ([21, 22]), our 8:0:2, 4:4:2, and 0:8:2 configurations leave only 20% of clients multimodal, reflecting realistic extreme settings. Unlike prior works evaluating a single heterogeneous setting, we also stress-test across 9 configurations. Clarity, Hyperparameters & Convergence [R1-R3]: All presentation issues (reference formatting, non-IID definition, capitalization, placement of the “to the best of” statement) will be corrected in the revision. The hyperparameter λ = 0.05 is tuned over {0.01, 0.05, 0.10} and remains robust across all configurations. Convergence was stable across runs, evidenced by consistent validation metrics. Implementation details will be released with source code and replication guidelines upon acceptance. Generalization beyond Chest X-ray [R1]: ProMoE-FL is modality-agnostic by design, formulated for an arbitrary set of modalities. Chest X-ray and radiology reports were chosen for their large-scale heterogeneous datasets suitable to stress-test a federated architecture. MoE Ablations [R2, R3]: The 8:0:2 setting was chosen as the most challenging configuration (Tab.1), making component contributions most distinguishable under stress. Our ablation explicitly isolates contributions of the prototype bank and MoE module. Fig. 2(c) shows balanced expert utilization; experts are not specialized per synthesis direction. Routing behavior analysis remains a valuable future direction. Feature Visualizations [R2]: Downstream AUC is a rigorous measure of synthesis quality as poor synthesis would directly degrade classification performance, strongly supported by our consistent SOTA results. t-SNE/UMAP visualizations provide complementary qualitative evidence. CLIP-based Approaches [R3]: FACMIC and pFedST do not handle missing modalities, making direct comparison methodologically unfair. Our framework is encoder-agnostic and makes no design assumptions specific to image-text pairs. Specific Performance Discrepancy [R3]: The 4:0:6 vs. 6:0:4 trend is consistent across all methods. In homogeneous settings, fewer multimodal clients reduce available cross-modal prior information. Under heterogeneity, multimodal priors may introduce distributional bias during unimodal synthesis, degrading performance.
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 main concerns to address are: (1) clarify novelty versus prior feature-imputation work (e.g., Feature Imputation Network), (2) add missing standard deviations for baselines and more metrics (e.g., PR-AUC), (3) analyze computational/communication overhead, (4) test extreme missing-modality scenarios, and (5) fix reference formatting (e.g., malformed author string).
- 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 reviewing the authors’ rebuttal and the post-rebuttal recommendations from reviewers, I recommend Accept. The rebuttal effectively addresses the key concerns.
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 paper addresses a timely and practically important problem in multimodal federated learning, namely missing modalities under heterogeneous clinical data distributions. The proposed ProMoE-FL framework is technically coherent, combining a client-aware prototype bank with prototype-conditioned MoE routing to synthesize missing modality features without relying on public multimodal data. The experimental evaluation across four chest X-ray datasets and multiple homogeneous and heterogeneous client configurations provides reasonably strong evidence of effectiveness, especially in more challenging non-IID settings. The reviewers raised valid concerns about the relationship to prior feature-imputation methods, incomplete statistical reporting, limited direct evaluation of synthesized feature quality, and computational or communication overhead. The rebuttal addresses these issues to a meaningful extent. In particular, the authors clarify how ProMoE-FL differs from earlier feature imputation frameworks through federation-wide prototype conditioning and dynamic expert routing, explain that the additional modules add only a small parameter overhead, and commit to reporting standard deviations for all baselines. The stress testing across several client configurations also strengthens the practical relevance of the evaluation. Some limitations remain, including the lack of direct quantitative assessment of feature synthesis fidelity and limited validation beyond chest X-ray/report data. Nevertheless, the problem setting is important, the method is sound, and the empirical improvements are consistent. I therefore support acceptance, with the expectation that the final version will include the clarified novelty discussion, full statistical reporting, overhead analysis, and corrected reference formatting.
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 paper proposes ProMoE-FL, a prototype-conditioned Mixture of Experts framework for multimodal federated learning with missing modalities.
Based on the reviewers’ feedback, all three reviewers converged to accept this paper. The authors have addressed key concerns including computational overhead (~0.3% additional parameters), standard deviations for baselines, robustness under heterogeneous settings. Some remaining concerns, such as deeper technical evaluations should be addressed in the final revision to strengthen the overall contribution.