List of Papers Browse by Subject Areas Author List
Abstract
Early and accurate prediction of pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) is critical for personalized breast cancer treatment. However, modeling complex longitudinal dynamics remains challenging in small-sample settings. Although biomedical foundation models offer strong representations, most 2D architectures ignore 3D spatial continuity and multi-scale tumor morphology, and existing approaches mainly rely on statistical correlations without considering underlying causal mechanisms. To address these issues, we propose a novel framework that integrates spatiotemporal modeling leveraging a biomedical foundation model with causal graph reasoning for pCR prediction. Our method introduces a Spatiotemporal-aware Attention (STA) mechanism, consisting of Multi-scale Adaptive Spatial-aware Attention (MASA) to capture 3D volumetric tumor morphology and Temporal-aware Attention (TA) to model treatment-induced changes across longitudinal MRI scans. Furthermore, we incorporate Causal Graph Reasoning to uncover latent causal pathways among longitudinal imaging dynamics, clinical biomarkers, and pCR, enabling unbiased estimation of treatment effects. Experiments on a multicenter cohort of 438 cases show that our method outperforms existing approaches. The causal model further surpasses its non-causal counterpart and identifies consistent causal effects of key biomarkers (ER, PR, HER-2) across centers. These results suggest that our method improves both predictive accuracy and interpretability.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0448_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LinHai_Spatiotemporal_MICCAI2026,
author = { Lin, Haiwei AND Huang, Bin AND Wang, Meng AND Ren, Ya AND Liu, Zhou AND Huang, Bingsheng},
title = { { Spatiotemporal Modeling of Longitudinal MRI with Causal Graph Reasoning for Breast Cancer pCR Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a framework integrating spatiotemporal modeling with causal graph reasoning for predicting pathological complete response to neoadjuvant chemotherapy in breast cancer. The main contributions are: 1.Spatiotemporal-aware Attention (STA): A two-part attention mechanism consisting of (a) Multi-scale Adaptive Spatial-aware Attention, which leverages a frozen BiomedCLIP encoder to extract 2D slice features and aggregates them into a 3D volumetric representation via a CLS token and self-attention, and (b) Temporal-aware Attention, which uses cross-attention between pre-NAC and early-NAC time points to model treatment-induced dynamic changes. 2.Causal Graph Reasoning: The predicted pCR probabilities from the imaging model are combined with clinical biomarkers and fed into NOTEARS to learn a DAG. A structural causal model is then constructed to estimate individual and average causal effects of biomarkers on pCR. 3.Multicenter Validation: Experiments on 438 cases across three centers demonstrate improved predictive performance (mean AUC 0.823 for the causal model) and clinically consistent causal effect estimates across centers.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clinically motivated and well-structured framework: The paper addresses a genuine clinical need — early non-invasive pCR prediction — with a clearly structured two-stage pipeline. The combination of imaging dynamics with causal inference is a thoughtful approach to improving both accuracy and interpretability. 2.Effective use of a biomedical foundation model: Leveraging the frozen BiomedCLIP encoder for feature extraction is a pragmatic choice for small-sample clinical settings, avoiding the need for training from scratch on limited data while still extracting strong visual representations. 3.Multicenter external validation: The use of three centers strengthens the generalizability claim. This is particularly valuable in medical imaging where single-center models often fail to generalize. 4.Comprehensive experimental comparison: The paper provides systematic comparisons with multiple spatial modeling methods (ABMIL, TransMIL) and temporal modeling methods (LSTM, BiLSTM, GRU, BiGRU, iMRrhpc), as well as ablation studies demonstrating the contribution of each component. 5.Clinically consistent causal effects: The identified causal effects of ER, PR, and HER-2 on pCR are consistent across all three centers and align with established clinical knowledge, supporting the plausibility of the causal reasoning. 6.Interpretability through causal analysis: The ICE/ACE framework provides patient-level and population-level interpretability, which is more actionable for clinicians than purely associative models.
- 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.Limited temporal modeling with only two time points: The framework uses only pre-NAC and early-NAC MRI scans. With only T=2 time points, the “temporal modeling” claim is significantly constrained — it is essentially a pairwise comparison rather than true longitudinal modeling. The TA module reduces to cross-attention between two feature sets, which limits the expressiveness of the temporal component. The authors should discuss this limitation and how the method could extend to more time points. 2.Loose coupling between spatiotemporal modeling and causal reasoning: The two main components appear as a two-stage pipeline rather than a truly integrated framework. The imaging model first produces a scalar probability, which is then used as a single variable in the causal graph. This lossy compression from rich imaging features to a single probability may discard informative causal signals. A tighter integration (e. g. , learning causal structure over latent imaging features) would be more principled. 3.Modest and inconsistent improvements: The improvement from the non-causal to causal model is small (mean AUC: 0.814 → 0.823), and in Center 3, AUC actually slightly decreases (0.826 → 0.824). No statistical significance tests (confidence intervals, p-values, or paired tests) are reported, making it difficult to judge whether these improvements are meaningful or within noise margins. 4.Methodological novelty is incremental: MASA is essentially a standard Transformer encoder with a CLS token and positional embeddings applied to slice-level features — this is a well-established architecture. TA uses cross-attention which, while effective, is not particularly novel. The causal reasoning component applies an existing method (NOTEARS) off-the-shelf. The novelty lies more in the combination than in any individual component. 5.Missing important experimental details and baselines: +No comparison with recent pCR prediction methods that also use longitudinal MRI and clinical data (e. g. , the cited iMRrhpc [6] is only compared on temporal modeling, not as a full pipeline). +The non-linear NOTEARS-MLP is mentioned in the implementation but the formulation in Section 2.3 describes a linear SEM — this inconsistency is confusing. +No sensitivity analysis on key hyperparameters (τ threshold for DAG sparsity, Nmax for padding, λ for regularization). 6.Imbalanced dataset without adequate handling: The pCR/non-pCR distributions are highly imbalanced (e. g. , 25/94 in Center 2). While the paper reports sensitivity and specificity separately, there is no discussion of how class imbalance is handled during training (e. g. , oversampling, weighted loss). This could significantly affect the reliability of the results. 7.Limited discussion on causal assumptions and validity: Causal structure learning from observational data requires strong assumptions (causal sufficiency, faithfulness, no hidden confounders). The paper does not discuss whether these assumptions are plausible in this clinical context, nor does it perform any sensitivity analysis on the learned DAG structure.
- 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 paper addresses a clinically important problem with a reasonable framework combining spatiotemporal attention and causal graph reasoning. The multicenter validation on 438 cases and the clinically consistent causal effect estimates are notable strengths. However, several factors temper my enthusiasm: 1.Limited temporal scope: With only two time points, the “temporal modeling” claim is overstated, and the TA component essentially performs pairwise cross-attention rather than true longitudinal modeling. 2.Incremental methodological novelty: The individual components are largely existing techniques applied in combination. The integration is somewhat loose — the causal module receives only a scalar probability rather than rich imaging features. 3.Insufficient statistical rigor: The improvements are modest and lack statistical significance testing. The slight AUC decrease in Center 3 for the causal model raises questions about the robustness of the causal reasoning component. 4.Missing comparisons and ablations: No comparison with state-of-the-art pCR prediction pipelines as complete systems, and no sensitivity analysis on critical hyperparameters or causal assumptions. On balance, the paper makes a reasonable contribution to the field through its combination of foundation model-based spatiotemporal modeling with causal reasoning, supported by multicenter validation. The clinical interpretability via causal effects is a genuine plus. If the authors can address the statistical significance concerns and better justify the causal assumptions in their rebuttal, I would be inclined to strengthen my recommendation.
- 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.
Basically, it solved my confusion.
Review #2
- Please describe the contribution of the paper
1.A spatiotemporal-aware attention mechanism is proposed for longitudinal MRI to capture 3D spatial features and temporal changes. 2.Causal graph reasoning is incorporated to reveal relationships among imaging features, clinical biomarkers, and outcomes for tumor response 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.The causal graph reasoning component is interesting because it integrates both imaging and clinical data and provides interpretability, which is particularly important for clinical tasks. 2.The method section is relatively well organized, and the paper provides component-wise comparisons as well as 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.
1.The comparison setup is not fully convincing. The proposed method is compared mainly against methods for either spatial modeling or temporal modeling, but it lacks stronger comparisons with existing spatiotemporal methods. 2.The paper provides a clear high-level description of the proposed framework, but several implementation details remain insufficient for full reproducibility for the proposed architecture, such as N_max. 3.The dataset is relatively limited, with only 438 cases, which may not be sufficient to support robust spatiotemporal modeling.
- 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?
The causal graph reasoning component is interesting because it integrates both imaging and clinical data and provides interpretability, which is particularly important for clinical tasks. However, the comparison setup is not fully convincing, and the reproducibility of the method is limited due to the use of a private dataset and the lack of sufficiently detailed methodological and parameter descriptions.
- 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.
I find the clarification regarding the baseline comparison against the spatio-temporal model reasonable and accept it.
Review #3
- Please describe the contribution of the paper
This paper proposes a longitudinal breast DCE-MRI-based framework for predicting pCR after NAC. The proposed method combines a frozen BiomedCLIP image encoder, a spatiotemporal attention module, and a NOTEARS based causal graph reasoning component. Specifically, MASA is used for inter slice spatial aggregation of tumor containing slices, while TA models temporal interactions between pre-NAC and early NAC MRI scans. The causal graph reasoning module is further used to identify potential relationships among image-based probabilities, clinical information, and pCR. The study is based on three retrospective multicenter cohorts, where Center 1 is used for model development and internal validation, and Centers 2 and 3 are used as external validation cohorts. The paper provides a relatively clear set of ablation experiments to evaluate the proposed spatial modeling, temporal modeling, multimodal integration, and causal graph reasoning components.
- 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 proposed method attempts to model both spatial and temporal structures in longitudinal DCE MRI. MASA, based on self attention, aggregates information from tumor containing slices, while TA, based on cross attention, explicitly models the interaction between pre NAC and early NAC MRI. This design is more appropriate for longitudinal treatment response prediction than simple feature concatenation, and the experimental results suggest that this strategy can improve pCR prediction.
2.The authors train and internally validate the model on Center 1, and then evaluate it on two external centers. This design is stronger than a purely single center evaluation and provides some evidence for the generalizability of the proposed method, although the two external validation cohorts are still relatively small.
3.The ablation experiments in Tables 1 to 4 are useful and relatively clear. The authors compare different spatial aggregation strategies, temporal modeling methods, clinical data integration, and the causal graph reasoning module. These experiments help readers understand the contribution of each component.
4.The proposed framework not only improves MRI based modeling but also incorporates clinical information through the causal graph reasoning module. Although the current interpretability analysis remains limited, Table 4 shows that the causal graph reasoning component may improve overall predictive performance.
- 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.According to Section 3, the study relies on manually segmented tumor ROIs. This limits the clinical applicability of the proposed method because manual segmentation is labor intensive and may introduce reader dependent variability. Importantly, this segmentation step is not clearly shown in Fig. 1 as an important prerequisite of the proposed pipeline.
2.Section 2.3 provides the mathematical formulation of causal graph reasoning, but the motivation, assumptions, and implementation details are insufficiently explained. The authors should not only describe how NOTEARS and SCM based inference are performed, but also explain why this causal formulation is appropriate for this clinical problem. Without these details, the causal reasoning component remains difficult to understand.
3.Tables 1 to 3 report AUC, accuracy, sensitivity, and specificity, but the discussion focuses mainly on AUC. The other metrics are not sufficiently interpreted, and the trends across centers are not always consistent. ROC curves might make the comparison clearer. Moreover, the paper does not report variance across the five-fold cross-validation, although this method is mentioned in Section 3.1. 4.The paper does not clearly describe which clinical variables were used, how missing values were handled, how biomarkers were encoded, or whether treatment regimen information was included. From Fig. 1, it seems that the clinical variables may include Ki67, ER, PR, HER2, maximal diameter, molecular subtype, and age, but this should be explicitly stated in the paper.
5.Only ER, PR, and HER2 were analyzed in the causal graph reasoning component. My concern is that this module may mainly capture known associations between receptor status and pCR, because ER, PR, and HER2 are strong predictors of pCR (Ref. 1). The clinical value of causal graph reasoning at the individual patient level remains unclear. In practice, clinicians would expect a model to provide stable performance across different patient subgroups. If the model mainly relies on receptor status or molecular subtype, then its added value beyond existing clinical decision factors may be limited.
[Ref. 1] https://doi.org/10.1186/s13058-018-0950-x
- 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.
(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?
Overall, the paper is clearly written, technically reasonable, and clinically relevant. The inclusion of external validation is valuable, and the proposed framework addresses an important problem in breast cancer treatment response prediction. However, the explanatory analysis of causal reasoning is still weak.
- 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 reading the authors’ rebuttal, I maintain my original Weak Accept recommendation. The paper discusses a clinically relevant problem and the inclusion of multicenter external validation, which remains a potential strength. However, several of my original concerns were not substantially addressed or responded to in the rebuttal. These include the reliance on manually segmented tumor ROIs, the limited explanation of the clinical variables, and the unclear added clinical value of the causal graph reasoning module beyond known biomarker associations such as ER, PR, and HER2.It also remains uncertain whether these issues will be adequately clarified in the revised manuscript.
Author Feedback
Thanks to the meta-reviewer and reviewers for your valuable comments on our work. We appreciate your recognition of:
- Reproducibility: “a clear and detailed description of the algorithm to ensure reproducibility” (R1, R4)
- Well-structured Framework: “a thoughtful approach to improving both accuracy and interpretability” (R1)
- Multicenter Validation: “The use of three centers is particularly valuable” (R1) and “provides some evidence for the generalizability” (R4)
- Experimental Rigor: “systematic comparisons and ablation studies” (R1, R2) “are useful and relatively clear” (R4)
- Interpretability: “The causal graph reasoning component is interesting” which “provides interpretability” (R2) and “is more actionable for clinicians than purely associative models” (R1)
Q1: Temporal modeling with only two time points (MR, R1) A1: Our method can be extended to multiple time points using either sequential cross-attention between adjacent stages or global self-attention across all time points. However, our design is clinically motivated. In NAC, the key objective is early prediction of pCR as soon as possible after treatment initiation, when therapy can still be adjusted to enable timely decisions such as switching ineffective regimens or avoiding unnecessary toxicity. In contrast, mid- or post-NAC MRI is mainly confirmatory or used for pre-surgical planning (e.g., assessing residual tumor burden and guiding surgical extent).
Q2: Justification of causal graph reasoning (MR, R1, R4) A2: The causal graph is used as a causally informed and interpretable modeling tool under standard assumptions to reduce spurious correlations and integrate imaging and clinical variables. The learned structure is clinically plausible, where known relationships such as Ki-67, ER, PR, and HER2 → molecular subtype follow established clinical definitions, and associations such as HER2/ER/PR → pCR are supported by prior breast cancer studies.
Q3: Modest performance gains and inconsistent trends across centers (MR, R1, R4) A3: This work focused on enhancing model interpretability through causal reasoning rather than performance improvement. Importantly, the causal module improves interpretability while also bringing performance gains, particularly showing noticeable improvement in generalization on the external test set (Center 2), while the slight variation in Center 3 is expected due to inter-center distribution differences.
Q4: Limited baselines (MR, R1, R2) A4: We already compare representative spatial and temporal baselines. Existing spatiotemporal methods (e.g., MICCAI 2022 and 2024 works, DOI: 10.1007/978-3-031-16437-8_51 and 10.1007/978-3-031-72120-5_61) and pCR baselines are not directly applicable to our dataset due to different preprocessing pipelines or ROI definitions, making fair comparison difficult. For example, iMRrhpc [6] operates on whole-breast 3D volumes with unified resizing and a 3D encoder, which differs from our multi-scale lesion-level representation and is not directly compatible with our setting.
Q5: Incomplete implementation details (MR, R1, R2, R4) A5: Clinical variables and missing-value handling follow standard practice in prior studies. For hyperparameters, we use default settings from the original NOTEARS implementation and its open-source code, which are widely adopted and validated in prior work.
Q6: Limited dataset size (R2) A6: Our study includes 438 cases from three centers, comparable to prior work. For example, a MICCAI study (DOI: 10.1007/978-3-032-05182-0_32) used 707 single-center cases without multi-center validation. Our dataset is multi-center, increasing heterogeneity.
Q7: Limited reproducibility (R2) A7: If the paper is accepted, we will release the code to ensure reproducibility.
Q8: NOTEARS formulation inconsistency (R1) A8: The linear SEM is used for clarity, while implementation uses nonlinear NOTEARS-MLP, which performed better in preliminary experiments.
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 reviews are mixed, with two weak-accept and one weak-reject recommendation. The paper addresses a clinically important problem in early pCR prediction for breast cancer using longitudinal DCE-MRI and clinical biomarkers. The reviewers recognize several strengths, including the use of a biomedical foundation model, spatiotemporal attention for pre-/early-NAC imaging, multicenter external validation, ablation studies, and an attempt to improve interpretability through causal graph reasoning.
However, the reviews raise several substantive concerns. The main issues are: (1) the temporal modeling claim is somewhat limited because only two time points are used, making the method closer to pairwise longitudinal comparison than general longitudinal modeling; (2) the causal reasoning component is not sufficiently justified, with limited discussion of causal assumptions, possible hidden confounding, DAG stability, and whether the module adds clinical value beyond known ER/PR/HER2 associations; (3) the experimental gains are modest and not supported by statistical significance testing or confidence intervals, with inconsistent trends across centers; and (4) the comparison and reproducibility details need strengthening, including stronger spatiotemporal/pCR baselines, manual ROI segmentation, clinical variable definitions, missing-value handling, class imbalance handling, and key hyperparameters.
- 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 considering the reviews and rebuttal, I recommend acceptance. The paper addresses an important clinical problem in early breast cancer pCR prediction and combines longitudinal DCE-MRI, clinical biomarkers, spatiotemporal attention, and causal graph reasoning. The multicenter validation and ablation studies are useful strengths.
The rebuttal clarified the rationale for using two early time points, the causal graph component, baseline comparisons, and implementation details. Reviewer 1 changed from Weak Accept to Accept, Reviewer 2 from Weak Reject to Accept, and Reviewer 4 maintained a positive recommendation. Some concerns remain, including manual ROI segmentation and the limited added clinical value of causal reasoning beyond known biomarkers, but these limitations do not outweigh the paper’s clinical relevance and overall contribution.
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 clinically important problem and proposes a reasonable framework combining longitudinal MRI modeling with causal graph reasoning for early pCR prediction. Reviewers generally agreed that the work is well organized, clinically motivated, and strengthened by multicenter validation and comprehensive ablation studies. The main concerns were related to the limited temporal scope with only two time points, the relatively loose integration between the imaging and causal reasoning components, modest performance gains, and limited discussion of causal assumptions and reproducibility details. The rebuttal clarified several of these points, particularly the clinical motivation for early prediction, the intended role of causal reasoning as an interpretability-focused module, and the rationale behind the comparison setting.
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 reviewers are unanimous in their decision. However, R3 has a few concerns, which the authors should address in the camera-ready