Abstract

Early diagnosis of Alzheimer’s disease (AD) is paramount for effective intervention. Resting-state functional magnetic resonance imaging (rs-fMRI) offers critical insights into how brain connectivity networks transition from cognitively normal (CN) to mild cognitive impairment (MCI, the early stage of AD) status. Most existing methods predominantly rely on static functional connectivity (FC), failing to capture directional (i.e., causal) and dynamic high-order synergistic interactions among brain regions. None of these methods utilize robust features from noisy and heterogeneous multi-view rs-fMRI representations. We propose HyperBrainNet, a causality-aware hybrid hypergraph network for better capturing subtle alterations of such complex interactions. Starting from whole-brain effective connectivity (EC) with causal inferences, we construct six complementary feature views by integrating the EC with 1) time-frequency statistics, 2) raw temporal activities, 3) static FC, 4) graph-theoretic topology, and 5) nodal entropy. We design Adaptive Reliability Fusion (ARF) to mitigate large inter-subject variations in noise levels/distributions across different views, a great challenge in conventional multiview learning. The ARF synthesizes sample-wise quality assessment networks with global learnable weights to achieve dynamic, weighted multiview feature fusion. A hybrid hypergraph then integrates these multiview features through learnable hyperedges. A multi-layer hypergraph convolution mechanism is employed to model high-order, inter-regional dependencies. Finally, dual-path attention pooling on nodes and hyperedges isolates critical hubs to provide interpretability. Our model significantly outperforms other similar methods in CN versus MCI classification, offering a novel solution for comprehensively investigating early pathological mechanisms of AD.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0825_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/eyeopeners/HyperBrainNet

Link to the Dataset(s)

ADNI dataset: https://adni.loni.usc.edu/

BibTex

@InProceedings{JiaYif_AMultiview_MICCAI2026,
        author = { Jia, Yifan AND Wang, Luoyu AND Tao, Yitian AND Zhu, Zihao AND Zhang, Yawen AND Zhang, Han},
        title = { { A Multi-view, Hybrid, Hypergraph Learning Framework with Causal Perturbation for Early Alzheimer’s Disease Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16885},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    In this paper the HyperBrainNet model, which is designed for early diagnosis of Alzheimer’s disease (CN vs MCI), is proposed. It integrates causal effective connectivity (EC), multi-view features, hybrid hypergraph and adaptive fusion mechanism, and achieves better performance than mainstream methods on the ADNI dataset.

    1.This work is the first to systematically integrate NPI-derived EC with multi-view features and hybrid hypergraph learning, effectively addressing three major limitations of traditional GNNs including reliance on undirected static FC, neglect of higher-order interactions, and the heterogeneity of noise across multiple views. The ARF module, hybrid hyperedge design, and dual-path attention pooling demonstrate clear originality and neuroscientific relevance. 2.The overall framework (Fig. 1) is logically well-structured, with each step supported by clear motivation, mathematical formulation, and implementation details. 3.Subject-level stratified 5-fold cross-validation is adopted, with comprehensive baseline comparisons. Ablation studies quantitatively validate the contribution of each component, and interpretability analyses are highly consistent with known pathological characteristics of AD. 4.In early screening scenarios for CN vs. MCI, the method achieves a good balance between sensitivity and specificity, provides interpretable biomarkers, and shows strong potential for real-world clinical translation.

  • 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.This work is the first to systematically integrate NPI-derived EC with multi-view features and hybrid hypergraph learning. 2.The overall framework (Fig. 1) is logically well-structured, with each step supported by clear motivation, mathematical formulation, and implementation details. 3.Comprehensive baseline comparisons are given. Ablation studies quantitatively validate the contribution of each component, and interpretability analyses are highly consistent with known pathological characteristics of AD. 4.In early screening scenarios for CN vs. MCI, the method achieves a good balance between sensitivity and specificity, provides interpretable biomarkers, and shows strong potential for real-world clinical translation.

  • 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.Insufficient details on NPI/EC extraction: in Section 2.2 only Equation (1) is provided, but key implementation details such as the architecture of the surrogate ANN, training hyperparameters, perturbation intensity settings, and post-processing of the EC matrix (e. g. , thresholding/normalization) are lacked. 2.Superedge K-value definition: The authors set the number of learnable superedges to K=10 to fill the gap between Schaefer 7 and 17 networks. It is suggested that the authors supplement this with a sensitivity analysis of the K-value (e. g. , K= {5, 10, 15, 20}) to demonstrate the stability of this parameter selection. 3.Overly brief description of the multi-view encoders: The “dedicated encoder” for the five additional views is only mentioned in one sentence, without details on network architecture, parameters, feature dimensions, or preprocessing steps. 4.Lack of statistical significance analysis: Although in Table 1 clear performance gaps are showed compared to baselines, no p-values or other statistical test results are reported. 5.Table 2 formatting is slightly disorganized; reformatting is recommended. 6.Some long sentences (e. g. , at the end of the Introduction and in Section 2.3) could be split to improve readability. 7.The hyperedge degree (De) should be explicitly defined before Equation (4). 8.The future work plan is too general and does not specify any concrete directions for expansion. 9.Computational efficiency: Considering that the model integrates CNN, Transformer, and hypergraph convolution, how are its parameter count and inference speed? It is recommended to supplement the FLOPs or runtime comparison of the main comparison models in the reply or body text.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    In this paper HyperBrainNet for early AD diagnosis using causal multi-view hypergraph learning is presented . The topic is important, and the experiments are comprehensive. Despite some flaws in detail and presentation, the work is solid and valuable.

  • 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 recommend accepting the paper, although I still view it as a borderline/weak-accept case rather than a strong accept.

    My initial concerns were mainly related to insufficient methodological details and presentation clarity, including the implementation of the NPI-based EC extraction, the construction of the multi-view features, the definition of the hypergraph structure, the choice of the learnable hyperedge number K, the lack of statistical significance analysis, and the computational cost of the proposed model. The rebuttal provides useful clarifications on these issues. In particular, the authors explain the surrogate model used for EC estimation, the computation of the additional feature views, the total hyperedge construction, the definition of the hyperedge degree, the rationale and sensitivity analysis for K, and the parameter/FLOP cost of the model. These responses partially alleviate my concerns regarding reproducibility and implementation transparency.

    I also acknowledge the concern that the causal claim may be overstated. However, the authors reasonably clarify that the method should be interpreted as using directed or model-inferred effective connectivity rather than interventionist causality. Their explanation that the EC matrix is encoded row-wise, so that each ROI embedding carries an outgoing directed profile, partially addresses the concern that directional information is completely lost in the downstream architecture. While this explanation is not perfect and should be made much clearer in the final paper, I do not think this issue alone warrants rejection after the authors have agreed to moderate the causal terminology.

    Overall, the paper addresses an important clinical problem, proposes a coherent integration of directed effective connectivity, adaptive multi-view fusion, and hybrid hypergraph learning, and reports competitive results with ablation and interpretability analyses. Some claims should be toned down, and several details from the rebuttal must be explicitly incorporated into the final manuscript. Nevertheless, after considering the rebuttal, I believe the paper is marginally above the MICCAI acceptance threshold.



Review #2

  • Please describe the contribution of the paper

    This paper proposes HyperBrainNet, a multi-view hybrid hypergraph neural network for early Alzheimer’s disease diagnosis via CN vs. MCI classification from resting-state fMRI. The framework addresses three limitations of existing graph-based approaches: lack of directional (causal) connectivity modeling, insufficient capture of high-order multi-region interactions, and suboptimal multi-view feature fusion under heterogeneous noise. The method extracts effective connectivity (EC) via Neural Perturbational Inference and constructs six complementary feature views (EC, FC, raw temporal, time-frequency statistics, graph topology, nodal entropy). An Adaptive Reliability Fusion (ARF) module dynamically re-weights views per subject based on learned quality scores. Fused features are then processed through a hybrid hypergraph that combines fixed neuroscientific priors (Schaefer’s 7-network parcellation) with orthogonally regularized learnable hyperedges, followed by multi-layer hypergraph convolution and dual-path attention pooling over nodes and hyperedges. Evaluated on ADNI (462 CN, 242 MCI) with five-fold cross-validation, HyperBrainNet achieves 74.4% accuracy and 74.9% AUC, outperforming GCN, GAT, BrainGNN, BrainNetTransformer, and other 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.

    1.Incorporation of effective connectivity beyond standard FC. Most rs-fMRI-based methods rely exclusively on undirected functional connectivity, which captures correlation but not directionality. This paper introduces effective connectivity via Neural Perturbational Inference alongside FC, providing a richer characterization of inter-regional relationships. The ablation in Table 2 confirms that the connectivity view contributes substantially to performance. 2.Multi-view complementary information utilization. The design of extracting multiple complementary views from the same rs-fMRI data — spanning temporal, spectral, connectivity, and topological domains — is well-motivated. The Adaptive Reliability Fusion module provides a principled mechanism for combining these views while accounting for heterogeneous noise across subjects, and the ablation demonstrates that each view contributes non-redundantly. 3.Hypergraph modeling is well-suited for brain network complexity. The human brain exhibits complex, high-order interactions that extend beyond pairwise connections. Hypergraphs, where a single hyperedge can connect multiple nodes simultaneously, are a natural formalism for capturing such multi-region co-activation patterns. The hybrid design combining neuroscientific priors with learnable hyperedges is a thoughtful approach that balances biological grounding with data-driven flexibility.

  • 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.Overcomplicated pipeline with unclear marginal contributions. The framework integrates six feature views, six dedicated encoders, sample-wise quality networks, global learnable weights, prior and learnable hyperedges, orthogonal regularization, multi-layer hypergraph convolution, dual-path attention pooling, focal loss with label smoothing, diversity regularization, and supervised contrastive loss. While each component is individually motivated, the sheer number of interacting modules makes it extremely difficult to identify where the performance gain truly originates. The ablation in Table 2 removes components in coarse groups but never isolates closely coupled elements — for instance, EC versus FC within the connectivity view, or prior versus learnable hyperedges within the hypergraph module. Without finer-grained ablations, the reader cannot determine whether the improvement comes from the causality modeling, the multi-view fusion, the hypergraph structure, or simply from having more parameters and loss terms. 2.Insufficient description of the six feature views and brain parcellation. The paper does not adequately explain what each of the six feature views concretely represents, how each is computed, or why these particular six were chosen over other possible representations. For instance, how are the time-frequency statistics extracted? What specific graph-theoretic topology metrics are used? How is nodal Shannon entropy defined and computed on the fMRI signal? The paper provides references but offers almost no inline description, significantly increasing the barrier to understanding and reproducibility. 3.Key structural details of the hypergraph are underspecified. Several critical quantities governing the hypergraph structure are not clearly defined. How many hyperedges are there in total (the paper mentions 7 prior + 10 learnable = 17, but this is not stated explicitly)? How many nodes does each hyperedge connect — is this uniform or variable? For the learnable hyperedges, the temperature-scaled similarity in Eq. 3 produces soft membership values, but it is unclear whether a threshold or top-k selection is applied to sparsify the incidence matrix, or whether all nodes belong to all learnable hyperedges with varying weights. For the prior hyperedges, are the Schaefer 7-network assignments binary and non-overlapping? These details are essential for understanding the model’s actual computational graph and for reproducibility. 4.The “causal” claim is overstated. The paper repeatedly positions causality as a core contribution, but the effective connectivity derived from NPI is based on virtual perturbations of a surrogate neural network trained on observational fMRI data. This captures directed statistical dependencies under the model’s learned dynamics, not true causal effects in the interventionist sense. Confounding pathways, indirect effects, and hemodynamic delays inherent to BOLD fMRI are not explicitly addressed. The paper should more carefully qualify this as “directed” or “model-inferred effective connectivity” rather than broadly claiming causal inference. 5.The claimed directionality is not preserved through the architecture. The paper emphasizes directional (causal) interactions as a key motivation, yet the downstream architecture does not maintain this directionality. Specifically, the Transformer blocks in the encoder allow every node to attend to every other node symmetrically via self-attention, effectively erasing the asymmetry present in the EC matrix. Furthermore, the hypergraph convolution in Eqs. 4–5 updates node features by symmetric aggregation – information flows from nodes to hyperedges and back without any directional distinction. This means that even though the EC input encodes that ROI i causally drives ROI j but not vice versa, this asymmetry is lost once the features pass through the Transformer and hypergraph layers. The paper does not discuss this architectural tension, nor does it demonstrate that the directional information in EC survives to influence the final classification. A directed message-passing scheme or an asymmetry-preserving attention mechanism would be needed to substantiate the claim that directionality is meaningfully exploited.

  • 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 core ideas (EC integration, hybrid hypergraph, multi-view fusion) are reasonable and the clinical task choice is appropriate. However, two major factors limit enthusiasm. First, the paper’s central claim – that directionality matters – is architecturally undermined: symmetric Transformer attention and symmetric hypergraph message passing erase the asymmetry encoded in the EC input, and the paper provides no evidence that directional information survives to influence classification. Second, the pipeline’s complexity far exceeds the granularity of the validation: six views, multiple loss terms, and numerous hyperparameters are evaluated only on a single dataset with coarse-grained ablations that do not isolate key components (EC vs. FC, prior vs. learnable hyperedges). Insufficient exposition of critical details (feature computation, hypergraph structure) further limits reproducibility. Sound ideas, but the evidence does not yet match the claims.

  • 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.

    The authors provided additional clarifications regarding some technical details in the rebuttal. However, my concerns regarding causality, directionality, and model complexity still remain.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of HyperBrainNet, a unified framework for early Alzheimer’s disease diagnosis from rs-fMRI data. The method goes beyond conventional approaches based only on static functional connectivity by incorporating causal effective connectivity, multi-view feature representations, and a hybrid hypergraph structure to model higher-order interactions among brain regions. In addition, the paper introduces an adaptive reliability fusion module to handle noise and heterogeneity across feature views, and an attention-based pooling strategy to improve interpretability. Experimental results on CN vs. MCI classification suggest that these components together lead to better performance than several representative baseline methods.

  • 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 overall idea is interesting. The paper combines effective connectivity, multi-view fusion, and hypergraph learning in a single framework for early AD diagnosis. 2.The multi-view setting is a strength. Using several complementary rs-fMRI feature types, together with adaptive fusion, is well motivated for handling noisy and heterogeneous data.

  • 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 novelty is somewhat incremental. The paper combines several existing directions—effective/causal connectivity analysis, hypergraph learning, and multi-view fusion/transformer-based modeling. In that sense, the main contribution is more a new integration of prior ideas than a clearly new methodological concept. 2.The validation is still limited. The experiments are conducted only on ADNI and only on the CN vs. MCI task. There is no external dataset, cross-site evaluation, or more realistic clinical setting, so the generalizability of the method is still unclear.3.The performance gain is not very large. In Table 1, the proposed model improves over the strongest baselines, but the margin is still relatively modest considering the complexity of the framework. This makes it harder to judge whether all proposed components are necessary in practice.

  • 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 score is mainly based on a balance between the interesting idea and the current limitations. On the positive side, the paper proposes a reasonably well-designed framework that integrates causal connectivity, multi-view features, and hypergraph learning, and the experimental section is fairly complete with ablation studies and interpretability analysis. These aspects make the work technically solid and relevant to early AD diagnosis.However, the contribution is somewhat incremental, as it largely combines existing techniques rather than introducing a clearly new methodological component. In addition, the evaluation is limited to a single dataset and task, and the performance gain over strong baselines is moderate. Taken together, these factors limit the overall impact and generalizability of the work, which leads to a moderate overall score.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Author Feedback

We thank the AC and R1/R2/R3.Our work targets CN-vs-MCI early AD screening by integrating NPI-derived directed EC, multi-view rs-fMRI, and hybrid hypergraph learning to capture directional and higher-order changes beyond standard FC/GNN pipelines. [M1: NPI details; AC,R1,R2] Following Luo et al. (Nat. Methods), our implementation trains an MLP surrogate to predict BOLD x(t+1) from x(t-2:t) by one-step MSE. We train it only on CN subjects as a healthy reference, then perturb all CN/MCI subjects, so EC reflects subject-specific deviation from healthy dynamics. For each source ROI, an impulse increase is injected; EC_ij is the L2 deviation between perturbed and baseline predictions of ROI j, averaged over time, row-wise min-max normalized, and 99th-percentile clipped, approximating the surrogate Jacobian. [M2: multiview features; AC,R1,R2] EC is in M1, and FC is standard. R2 asked about other views: raw time series uses Conv1D+2-layer Transformer over time; time-frequency statistics are 16-D/ROI, including time-domain stats (mean/std/skewness) and Welch-PSD features (slow-band power/spectral centroid); graph topology on top-20% |FC| includes degree, betweenness, clustering, local efficiency, participation coefficient, and within-module z-score (Schaefer 7-net); nodal Shannon entropy is H_i=-sum_j p_ij log p_ij, where p_ij comes from sliding-window dFC variability. Each view is mapped to a shared latent dimension before ARF. [M3: causality; R2] We agree to qualify “causal” as “directed/model-inferred EC” rather than interventionist causality. NPI was validated against ground-truth generative EC and CCEP propagation, and moves beyond undirected FC. Directionality is not erased: the Transformer attends over time, not ROI pairs, so it does not symmetrize EC. EC is encoded row-wise; thus ROI i’s embedding carries its outgoing profile (EC_ij != EC_ji). Hypergraph convolution aggregates these direction-aware embeddings, with directionality residing in node features rather than edge operators. In finer ablations omitted for space, removing EC while keeping FC caused larger degradation than removing FC while keeping EC, supporting that directed EC reaches the classifier. [M4: hypergraph details; AC,R1,R2] Total hyperedges = 17: seven fixed Schaefer 7-network prior hyperedges plus ten learnable hyperedges. Prior incidence is binary/non-overlapping, with each ROI assigned to one Schaefer network. Learnable incidence in Eq. 3 is soft during training; at inference, top-k node selection per hyperedge yields sparse binary incidence. In Eq. 4, D_e=sum_i H(i,e), which we will define before the equation. K=10 bridges Schaefer 7- and 17-network granularities while allowing AD-specific reorganization. K in {5,10,15,20}: performance rose then fell, peaking at K=10.[M5: validation; R1,R2,R3] Table 2 groups six views into three semantic categories as a presentation choice within the page limit. The key fine-grained clarification is in M3: EC contributes beyond FC, addressing R2’s concern on whether directionality matters. The gain is not due to model size: HyperBrainNet has ~0.51M parameters and ~0.10 GFLOPs, lighter than BrainNetTransformer. The small 5-fold std in Table 1 indicates stable gains; paired tests on existing folds show significant improvements. ADNI is a multi-site/multi-phase consortium; external-cohort validation was conducted but omitted for space. Minor revisions: reformat Table 2; split long sentences; define D_e; clarify NPI/multi-view/hypergraph details; specify future work. Code will be released after acceptance. Given R1/R3’s support, we invite R2/AC to reconsider: 1) we qualify causality and use a CN-only surrogate to define disease-deviation EC; 2) the Transformer attends over time (not ROI pairs) and row-wise EC preserves EC_ij != EC_ji asymmetry; 3) fine-grained ablations support EC’s value beyond FC; 4) the lighter-than-BrainNetTransformer budget rules out parameter inflation as the source of observed gains.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

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

    This paper presents HyperBrainNet, consisting of Adaptive Reliability Fusion for multiview representations. including neural perturbational inference, Hybrid Hypergraph Learning, and Hierarchical Attention Pooling. While the method presents some novelty, several critiques have been raised that need to be addressed in the rebuttal. For example, details of NPI, multiview features, causality, and hypergraph need to be provided. The significance of validation needs to be discussed as well.

  • 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.

    Reviewers’ critiques focused on the n details of NPI, multiview features, causality, hypergraph setting, and validation, which were addressed by the authors’ rebuttal. The remaining confusion of causality and directionality need to be further clarified in the final version of the paper.



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 work is clinically relevant and technically coherent, and two reviewers support acceptance. Although Reviewer #2 raises a valid concern that the causal/directional claim is overstated and not fully preserved by the downstream architecture, the rebuttal reasonably clarifies that the method should be interpreted as using model-inferred directed EC rather than interventionist causality. The authors also address several missing details regarding NPI, multi-view features, hypergraph construction, K sensitivity, statistical testing, and computational cost. I believe the paper is marginally above the acceptance threshold, provided that the final version incorporates the rebuttal clarifications and substantially tones down the causal claims.



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 authors provide additional clarifications regarding the reviewers concerns but not fully. Issues remain regarding some technical details and overclaims, but majority of reviewers lean towards accepting the paper. Contents from the rebuttal must be explicitly incorporated into the final manuscript.



back to top