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

Abstract

Aortic stenosis (AS) is characterized by highly heterogeneous progression trajectories, making personalized surveillance and timely intervention challenging. We propose AS-TIME, a time-conditioned framework that predicts patient-specific AS progression from a single baseline echocardiogram and follow-up interval ∆t. AS-TIME hierarchically integrates: (i) attention-based multiple instance learning to aggregate video and report representations, (ii) ∆t-conditioned modality gating for adaptive multimodal fusion, and (iii) a ∆t-guided mixture-of-experts head to model heterogeneous progression dynamics. AS-TIME achieves an AUC of 80.4% and balanced accuracy of 73.8%, outperforming all baselines. Temporal ablation demonstrate that explicit ∆t conditioning improves both modality fusion and expert routing. The model provides multi-level explainability through modality fusion weights and expert utilization. Our code is available at: https://github.com/DeepRCL/AS-TIME.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6082_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/DeepRCL/AS-TIME

Link to the Dataset(s)

N/A

BibTex

@InProceedings{KimDia_ASTIME_MICCAI2026,
        author = { Kim, Diane AND To, Minh Nguyen Nhat AND Wu, Victoria AND Tsang, Teresa S. M. AND Luong, Christina AND Abolmaesumi, Purang},
        title = { { AS-TIME: Time-Conditioned Multimodal Modeling of Aortic Stenosis Progression } },
        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 main contributions are the formulation of AS progression prediction as a time-conditioned multimodal problem using only a single baseline study and elapsed time, and the proposed hierarchical architecture that combines MIL-based aggregation, time-conditioned modality gating, and a mixture-of-experts head. The paper also demonstrates that incorporating time into fusion and prediction improves performance over standard multimodal baselines.

  • 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 clinically meaningful and challenging problem, and the overall approach is well motivated. The use of multimodal data (echo videos and reports) combined with explicit conditioning on elapsed time is intuitive and aligns with how clinicians think about disease progression. The architecture is clean and modular, and the results show consistent (if modest) improvements over reasonable baselines. I also appreciated the effort toward interpretability, with modality weighting and expert routing providing some insight into model behavior, which is important for clinical applications.

  • 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 primary weakness I see is the evaluation. The dataset is relatively small and limited to a single institution, with no external validation, making it difficult to assess generalization. The performance gains over baselines are also modest, and it is unclear whether they are statistically significant. While the paper emphasizes clinical relevance and robustness, these claims are not strongly supported by the experiments, which limits confidence in the practical impact of the method. In addition, despite the “longitudinal” language, the model only uses Delta t for its predictions. The model cannot actually condition on trajectories, just on the elapsed time. Finally, the paper does not sufficiently cite many of the methods that it integrates. For instance, using attention aggregation for frame embeddings should be cited (added references below).

  • 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

    Video aggregation is the same as (general) Ilse, Maximilian, Jakub Tomczak, and Max Welling. “Attention-based deep multiple instance learning. “ In International conference on machine learning, pp. 2127-2136.PMLR, 2018. (echo context) Smith, D. Hudson, John Paul Lineberger, and George H. Baker. “On the relevance of temporal features for medical ultrasound video recognition. “ In International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 744-753.Cham: Springer Nature Switzerland, 2023.

  • 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 is clinically motivated, but the limited novelty, modest gains, and lack of strong validation leave it below the bar for me.

  • 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.

    Given the planned improvements to clarity and literature support, I am happy to support acceptance.



Review #2

  • Please describe the contribution of the paper

    This paper proposes AS-TIME (Temporal Inference Model with Echocardiography), a time-conditioned multimodal framework for patient-specific aortic stenosis (AS) progression prediction using only a single baseline echocardiography study and elapsed time to follow-up (Δt). The core innovations include:

    1.First Time-Conditioned Echocardiography Model: To the authors’ knowledge, this is the first work exploring explicit time conditioning for longitudinal echocardiography analysis in aortic stenosis progression prediction. Unlike traditional static prediction models, AS-TIME treats Δt as a structural signal that reshapes how echo videos and reports are interpreted.

    2.Hierarchical Fusion Architecture with Three-Level Time Integration:

    • Token-level MIL: Attention-based multiple instance learning aggregates variable-length tokens within echo videos and clinical reports
    • Modality-level Δt-conditioned gating: Adaptively weights video and text representations based on disagreement and elapsed time
    • Prediction-level Δt-guided Mixture-of-Experts (MoE): Routes fused representations through 12 experts with time-dependent specialization

    3.Novel Disagreement-Aware Modality Gating: Computes cosine similarity between video and report embeddings, concatenates with Δt encoding, and learns adaptive modality weights (observed: 96% video, 4% report on average), reflecting the inherent data generation structure where reports are clinician interpretations of videos.

    4.Explainability at Multiple Levels:

    • Instance-level attention weights show which echo views and report sentences are important
    • Modality contribution scores reveal video vs. report relative importance
    • Expert routing probabilities expose Δt-dependent prediction patterns (e.g., Expert 8 for Δt<8 years, Expert 3 for longer intervals)
    • Model is well-calibrated (Figure 2), essential for clinical deployment

    5.Strong Clinical Validation: Evaluated on 4,124 AS patient pairs from a tertiary medical center, achieving 80.4% AUC and 73.8% balanced accuracy, outperforming all baselines including recent temporal methods (TAMME).

  • 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.

    Strength 1: Novel and Well-Motivated Time-Conditioned Architecture

    • Explicit Δt conditioning at multiple hierarchy levels (modality fusion and expert routing) is innovative and well-motivated for trajectory-aware prediction.
    • The disagreement-aware gating mechanism is clever: using cosine similarity between modalities reveals complementarity (avg: -0.33, suggesting modalities capture different information), which informs adaptive weighting.
    • MoE with time-dependent expert specialization is elegant and interpretable: different experts activate for different Δt intervals, reflecting heterogeneous AS progression patterns.
    • The hierarchical design (token MIL → modality gating → MoE) is principled and mirrors clinical reasoning about disease progression.

    Strength 2: Strong Empirical Validation with Rigorous Ablation Studies

    • Comprehensive baseline comparisons: temporal embeddings alone, unimodal (video/report), multimodal fusion strategies (TAMME, concatenation, cross-attention), and foundation models (EchoPrime, BERT, MedGemma).
    • Strong results: 80.4% AUC, 73.8% BAcc, up to 15.29% AUC improvement over baselines.
    • Systematic temporal ablation (Table 2) demonstrates clear value of Δt conditioning: no Δt (73.91% AUC) → gated fusion (79.42%) → full AS-TIME (80.40%), with reduced variance suggesting stabilization.
    • Statistical rigor: results reported with standard deviations across 5 seeds, showing reproducibility.
    • Reasonable dataset size: 4,124 patient pairs with 35,658 total echo videos across 3 views (A5C, PLAX, PSAXAo).

    Strength 3: Multi-Level Interpretability and Clinical Alignment

    • Explainability at instance, modality, and expert levels provides transparency for clinical deployment.
    • Observed modality weights (96% video, 4% report) align with clinical practice where reports are interpretations of videos.
    • Expert routing patterns show meaningful temporal specialization (Expert 8 for early intervals, Expert 3 for longer follow-up), suggesting the model learns clinically relevant progression dynamics.
    • Well-calibrated predictions (Figure 2) are essential for clinical trust and decision-making.
    • Only requires single baseline echo without additional clinical variables, addressing practical deployment challenges.

    Strength 4: Strong Clinical Motivation and Impact Potential

    • AS is a prevalent and growing problem (global aging population), and current surveillance is suboptimal.
    • Heterogeneous progression patterns make individualized risk stratification critical for optimizing follow-up schedules.
    • The framework supports adaptive, risk-guided surveillance, particularly valuable for resource-limited settings and rural/remote communities.
    • Clear future directions outlined: subgroup analysis, multi-visit longitudinal modeling, time-to-progression estimation, external validation.
  • 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.

    Weakness 1: Missing Comparison to Time-Aware Baselines and Simple Ablations

    • Issue: The reviewer specifically notes: “No comparison to simple time aware model that only use either report or images alone are done, in this past few years multiple methods were proposed across MICCAI and ISBI.”
    • While the paper compares against TAMME [10] (which underperformed), there are no comparisons to other recent time-aware multimodal methods from MICCAI/ISBI 2024-2025.- Critically missing: simple baselines using EchoPrime+Δt or BERT+Δt with time-conditioned architectures similar to AS-TIME (e.g., MoE routing on unimodal features). This would isolate the value of multimodal fusion vs. sophisticated temporal modeling on single modality.
    • The TAMME underperformance is attributed to “requiring multiple input data across multiple time points,” but this should be investigated more rigorously. Why not adapt TAMME to single timepoint?
    • Specific references: Recent time-aware medical imaging methods should be compared: Liu et al. 2025 [3] (cited in paper but not compared), Susetzky et al. 2025 [10] (TAMME), and other temporal modeling approaches from recent MICCAI proceedings.

    Weakness 2: Insufficient Dataset and Experimental Details

    • Issue: The reviewer requests: “Very detail on the split distribution and how was the ablation study perform and which set the results are reported.”
    • The paper states “70% training, 15% validation, 15% test set, stratified by class distribution while ensuring patient exclusivity” but lacks critical details:
    • Exact sample sizes for train/val/test splits not reported
    • Event rate (significant AS %) in each split not provided (only overall: 63.4% early, 36.6% significant)
    • How is patient exclusivity ensured when there are 4,124 pairs from 2,605 patients? Are pairs from the same patient split consistently?
    • Temporal distribution of Δt in each split not reported - is there temporal bias?
    • Random seed for split not mentioned
    • Ablation study details: Are all experiments (baselines, ablations) evaluated on the same test set? Or is hyperparameter tuning on validation set followed by final evaluation on test set?
    • No external validation cohort - all experiments are on single institution data, limiting generalizability claims.

    Weakness 3: Limited Analysis of MoE Expert Behavior and Architectural Justification

    • Issue: While expert routing patterns are described (Expert 8, Expert 3, etc.), the analysis is superficial:
    • Why 12 experts? Was this number ablated (e.g., 4, 8, 16 experts)?
    • How diverse are the learned expert behaviors? Are some experts redundant?
    • No visualization of what features each expert focuses on (e.g., saliency maps per expert)
    • No analysis of failure cases: When does the MoE fail? Which experts misroute?
    • The modality gating shows strong video preference (96%), raising the question: Why not just use video-only model with better architecture? The 4% report contribution seems marginal.
    • Missing architectural ablations: What if we use attention pooling instead of MIL? What if we use different temporal encoding (learnable vs. sinusoidal)?

    Weakness 4: Limited Reproducibility

    • Issue: No statement about code/model release is provided.
    • Dataset is private (single tertiary medical center), limiting reproducibility.
    • Some implementation details are provided (AdamW, lr=3.8e-6, batch=16, 50 epochs), but others are missing:
    • MIL architecture details (attention mechanism, hidden dimensions)
    • Gating network architecture (number of layers, activation functions)
    • MoE routing mechanism specifics (top-k selection? soft vs. hard routing?)
    • How are 32 frames sampled from videos (temporal distribution, selection criteria)?
    • Foundation model versions (EchoPrime, BioClinical-ModernBERT specific checkpoints)
  • 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 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 - No ethics concerns identified. The paper uses retrospective echocardiography data from a tertiary medical center, which presumably has appropriate IRB approval and patient consent.

  • 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

    Major Comments: 1.CRITICAL: Add comparisons to recent time-aware multimodal methods from MICCAI/ISBI 2024-2025.Include simple but strong baselines: EchoPrime+Δt with MoE (no report), BERT+Δt with MoE (no video) to isolate multimodal fusion value. 2.Provide complete dataset split details: exact train/val/test sizes, event rates per split, Δt distribution per split, handling of multiple pairs from same patient, random seeds. 3.Clarify ablation study methodology: Are all experiments on the same test set? How are hyperparameters tuned? Are results on validation or test set? 4.Add external validation on independent institution to support generalizability claims. 5.Expand MoE analysis: Ablate number of experts, visualize per-expert attention/saliency, analyze failure cases and expert misrouting. 6.Commit to code release for reproducibility.

    Minor Comments: 1.Justify architectural choices: Why 12 experts? Why sinusoidal encoding for Δt vs. learnable? 2.Analyze the 96% video vs. 4% report weighting: Is the report truly contributing? Compare against video-only with similar architecture. 3.Provide more MIL implementation details: attention mechanism type, hidden dimensions, number of layers. 4.Include confidence intervals or statistical significance tests for comparisons in Table 1.5.Discuss computational costs: training time, inference speed, memory requirements. 6.Add failure case analysis: When and why does AS-TIME fail? Are there specific Δt ranges or AS subtypes where performance degrades? 7.Clarify MoE routing: Is it soft (weighted combination) or hard (top-k selection)?

    For extension: 1.How does performance vary across different Δt ranges? Is there a minimum or maximum Δt where the model becomes unreliable? 2.How does the model handle edge cases: Δt = 0 (same-day prediction), very long Δt (>10 years)? 3.Can the framework be adapted to predict time-to-progression (Δt) given current and target AS severity?

  • 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?

    Recommendation: Accept. The paper makes significant methodological and clinical contributions. Weaknesses are addressable and don’t diminish the core novelty and impact.

    Positive Factors:

    • Novel and well-motivated contribution: First time-conditioned echocardiography model for AS progression prediction with hierarchical Δt integration at fusion and routing levels
    • Strong technical design: Disagreement-aware modality gating, Δt-guided MoE routing, and multi-level interpretability are innovative and principled
    • Impressive empirical results: 80.4% AUC, 73.8% BAcc, outperforming all baselines with up to 15.29% AUC improvement
    • Rigorous ablation studies: Systematic temporal ablation (Table 2) demonstrates clear value of Δt conditioning at both fusion (5.51% gain) and routing levels
    • Multi-level explainability: Instance attention, modality weights, and expert routing provide clinical interpretability
    • Practical clinical utility: Single baseline echo study without additional variables, suitable for resource-limited settings
    • Clinical relevance: AS surveillance is an important problem with heterogeneous progression patterns requiring personalized risk stratification
    • Well-calibrated predictions: Figure 2 shows good calibration, essential for clinical deployment
    • Clear presentation: Well-written with good figures and comprehensive tables

    Concerns:

    • Missing comparison to recent time-aware multimodal baselines and simple unimodal+MoE ablations
    • Insufficient dataset/experimental details (split sizes, Δt distributions, ablation methodology)
    • No external validation (single institution only)
    • Limited MoE expert analysis (no ablation on number of experts, no per-expert visualizations)
    • No code release commitment
    • 96% video weighting raises question about marginal report contribution
  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    I have changed my mind concerning paper, while the author provide clarification on certain elements .There are crutial experiments that need to be run and cannot be run during the rebuttal phase.



Review #3

  • Please describe the contribution of the paper

    This paper proposes AS-TIME, a time-conditioned multimodal model for predicting aortic stenosis progression from a single baseline echocardiography study and its report. Its main contribution is treating the time interval as a conditioning signal that dynamically controls both modality fusion and prediction (via mixture-of-experts), enabling trajectory-aware prediction with minimal input data.

  • 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.
    • Addresses patient-specific progression prediction for aortic stenosis, which is important for follow-up planning.
    • Good completeness as it outperforms baselines with supporting ablation studies.
  • 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 model assumes delta-t is given at inference time, but in practice, clinicians want to decide when to follow up, not predict risk at a predefined delta-t. This limits direct clinical usability.
    • There is no detailed breakdown of performance across different delta-t ranges (e.g., short vs long intervals), which is critical since time is central to the method.
  • 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
    • Consider including simpler time-aware baselines (e.g., survival models or regression on time-to-progression) to better justify the complexity of the proposed architecture.
    • It would strengthen the paper to clarify better the clinical use case, especially how the model would be used in practice, given that delta-t is required as an input. In reality, clinicians often need to decide the follow-up interval.
    • The color scheme in the Modality-Level MIL block is confusing. Red is not clearly introduced, green may be mistaken for time, and blue/purple appear inconsistently used. I think these likely represent attention weights. Clearer and consistent color coding would improve readability.
    • It would be helpful to include performance stratified by delta-t ranges to better understand model behavior across short vs. long-term prediction, along with a brief analysis of failure cases, particularly where the model is confident but incorrect.
  • 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?

    This is a solid and clean piece of work with good overall completeness. While the technical novelty is somewhat incremental and the clinical use case could be better clarified, the paper is well-executed and provides a practical and meaningful contribution.

  • 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.

    This paper is well-executed, clinically well-motivated, and shows strong practical potential. While the methodological novelty is somewhat incremental, the overall completeness and empirical support are sufficient. The authors also addressed several concerns in the rebuttal and clarified important experimental details. Overall, I lean toward acceptance.



Author Feedback

We extend our sincere gratitude to the Area Chair for the opportunity to clarify our work. We appreciate the reviewers’ thoughtful comments and are encouraged that they found AS-TIME clinically meaningful and well motivated. We address the key concerns below: 1.Model Framing/Clinical Relevance (R1,R3): AS-TIME is intended as a horizon-conditioned risk estimator: given a baseline echo study, clinicians can query personalized progression risk across plausible follow-up intervals (0-10 years) to support individualized surveillance planning. We will revise the “longitudinal” wording to clarify this framing and discuss future extensions for multi-timepoint trajectory modeling and direct time-to-progression prediction. 2.Novelty (R1): We agree that attention-based MIL and MoE build on established methods, and sincerely thank R1 for suggesting additional references. We will include them in our paper. We would like to respectfully emphasize that our contribution is not MIL or MoE in isolation, but their integration into a time-conditioned framework where ∆t dynamically modulates multimodal fusion and expert specialization. We will make our novelty clearer, in line with R2 and R3’s recognition of time-conditioned formulation as central to our work. 3.Dataset (R2): The data were split into train/val/test sets of 2,933/594/597 study pairs, stratified by class, with about 64% no progression and 36% progression in each split. Patient exclusivity was enforced by splitting on patient IDs, thus all pairs from the same patient remained in a single split. Mean ∆t was comparable across train/val/test: 2.35/2.46/2.46 years. All results in Tables 1 and 2 were evaluated on the same held-out test set and averaged over five seeds. Hyperparameters were tuned on the val set. We will revise the Experiments section of our paper for clarity. 4.Statistical Significance (R1,2): In comparison to Cross-Attention which achieved the highest baseline, AS-TIME consistently outperformed across all seeds. Paired t-tests confirmed statistically significant differences in both AUC and BAcc. We will make this clearer in our final manuscript. 5.External Validation/Baseline (R1-3): We agree that lack of external validation limits generalizability. This was challenging because AS-TIME requires longitudinal echo studies with reports and clearly defined AS severity labels. Despite our best efforts, we were unable to identify a publicly available or accessible external dataset containing this combination of data. We will include this limitation in the final paper and acknowledge that further validation is important for assessing generalization. Additionally, we appreciate R2 and R3’s suggested extra baselines and will incorporate them into our future work. 6.Modality Weighting/MoE/Performance Across ∆t (R1-3): We agree that the strong video weighting warrants further discussion. We do not view it as evidence that reports are uninformative, but rather, it likely reflects the clinical relationship between raw echo videos and report summaries. We will acknowledge that further analysis is required to better characterize the contribution of the report modality. AS-TIME uses a soft weighted expert combination, with the number of experts selected by validation. AUROC remained stable across ∆t bins, suggesting preserved discriminative performance across time horizons. BAcc showed slight decrease in longer horizons, likely reflecting increased prediction difficulty and significantly reduced sample sizes in these subgroups. 7.Reproducibility/Citations/Figure (R1-3): We will add the suggested citations, implementation details, and release code upon acceptance. We will also revise the method figure with a clearer color scheme. We will revise the final manuscript to address feedback outlined above. If these clarifications address your concerns, we kindly ask you to consider revising your score.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper addresses an important clinical problem and presents a well-motivated time-conditioned framework for aortic stenosis progression prediction. The paper is generally well executed, with supportive ablations and good overall clarity. The main concerns are around validation and clinical framing. The reported gains, while encouraging, should be better supported with statistical testing and more complete split details.

  • 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 addresses an important clinical problem and presents a well-motivated framework for modelling aortic stenosis progression using multimodal echocardiography data and explicit time conditioning. The proposed integration of elapsed time into multimodal fusion and prediction is intuitive, clinically relevant, and supported by consistent improvements over competitive baselines. The work is strengthened by comprehensive ablation studies, interpretability analyses, and a reasonably large longitudinal cohort.

    The rebuttal satisfactorily clarified several concerns regarding data splitting, statistical significance, model design, and clinical framing. While the lack of external validation remains an important limitation, and additional time-aware baselines would further strengthen the study, these issues do not outweigh the overall methodological contribution and clinical relevance of the work. Overall, the paper presents a well-executed and practically meaningful contribution that is likely to be of interest to the MICCAI community.



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.

    All reviewers recommend acceptance and judge the paper positively.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    One reviewers highlighted several key concerns including: 1.Missing Comparison to Time-Aware Baselines and Simple Ablations 2.Insufficient Dataset and Experimental Details 3.Limited Analysis of MoE Expert Behavior and Architectural Justification 4.Limited Reproducibility

    And also highlighted that this work needs a substantial work and new experiments, which cannot be done during rebuttal phase.



back to top