List of Papers Browse by Subject Areas Author List
Abstract
Understanding the dynamic transitions of brain states is crucial for comprehending the working mechanisms of the brain. However, accurately decoding these states from functional networks remains challenging, as existing methods lack the interpretability to reveal how internal hubs drive brain integration. To address this issue, we propose Brain-TM, a hierarchical Transformer-Mamba framework designed to decode brain states from functional networks. Evaluated on the HCP dataset across 7 cognitive tasks, Brain-TM achieves a superior classification accuracy of 94.7%. Beyond its predictive performance, it offers a biologically interpretable framework by proceeding at three analytical levels. Firstly, we identify task-specific network representations, demonstrating that latent brain states correspond to distinct cognitive demands across varying temporal windows. Secondly, we reveal a hub-based mechanism across brain states, showing that global network integration is fundamentally orchestrated by localized hub activations. Finally, we investigate the topological characteristics of functional hubs, validating their dominance via centrality metrics and benchmark comparisons. In summary, Brain-TM successfully decodes spatiotemporal dynamics, revealing that internal hubs play a pivotal role in driving brain states.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1279_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{LiuXua_BrainTM_MICCAI2026,
author = { Liu, Xuan AND Wang, Xuhui AND Yu, Sigang AND Zhu, Di AND Yao, Yincheng AND Zhang, Shu},
title = { { Brain-TM: Decoding Brain States from Functional Networks Using a Hierarchical Spatiotemporal Transformer-Mamba } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces Brain-TM, a novel hierarchical spatiotemporal framework that combines GIN), Spatial Transformer, and State Space Model (Mamba) to decode dynamic brain states. The framework successfully extracts spatial functional topographies and long-range temporal dynamics, achieving a 94.7% classification accuracy across seven cognitive tasks on the HCP dataset. Beyond predictive performance, the study provides a robust interpretability analysis, demonstrating that global brain network integration during specific cognitive tasks is fundamentally driven by localized hub activations.
- 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 model effectively bridges local graph topology and global temporal dynamics. It uses GIN for local node embedding, a masked Cross-Attention mechanism to aggregate nodes into 17 specific functional networks, and the Mamba architecture to efficiently model long-range temporal state transitions without the quadratic computational bottleneck of standard temporal Transformers. 2.The authors map latent temporal features to task-specific spatial attention weights, utilize Gini coefficients to quantify “Distributed vs. Hub-Driven” network patterns, and validate the topological dominance of these hubs.
- 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.Brain-TM is designed as a spatiotemporal model that captures dynamic functional connectivity across sliding windows. However, the chosen baselines (e. g. , GAT, GCN, BrainNetCNN, BrainGNN) are predominantly static spatial graph models. Comparing a dynamic sequence model against static snapshot models inherently favors the former. The paper lacks a comparison against contemporary dynamic spatiotemporal models (such as ST-GCN or dynamic variants of the Brain Network Transformer) 2.The experimental methodology states that performance was assessed using a “5-fold stratified cross-validation”. However, it is not clear whether this split was strictly bounded at the subject level. If sliding windows from the same subject are randomly distributed across training and testing sets, the model may simply learn to identify subject-specific baselines rather than generalizing cognitive tasks. 3.In the ablation study, replacing the Mamba module with a standard Transformer resulted in a severe performance collapse, dropping from 94.71% to 82.72%. Given that Transformers typically excel at capturing temporal dependencies, this 12% drop is very strange and may raise critical concerns regarding whether the baseline Transformer was optimally tuned (e. g. , adequate layers, heads, and positional encodings). 4.The model uses a sliding window strategy with window sizes ranging from 10 to 35 TRs. Considering that HCP task-fMRI data utilizes a block design with relatively short task durations, larger windows (e. g. , 30-35 TRs) will span across multiple cognitive states for most of tasks. This temporal blurring may introduce highly mixed signals into a single dynamic adjacency matrix, which contradicts the paper’s claim of identifying task-specific functional representations.
- 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?
Innovative Architecture: This paper proposes a novel hierarchical spatiotemporal framework that integrates GIN for local node embedding, a masked Cross-Attention mechanism to aggregate nodes into functional networks, and the Mamba architecture to efficiently model long-range temporal state transitions. High decoding performance: The model achieves an accuracy of 94.71% for decoding seven tasks on HCP datasets and demonstrates superiority over several established graph neural network baselines.
- 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.
Baselines: The accuracy scores for ST-GCN (93.42%) and Brain Network Transformer (94.37%) should be added to the main text, alongside other recent models. State-of-the-art performance for 7-way task domain classification can reach 99%, far exceeding the results reported here. Have the authors considered decoding fine-grained brain states (over 20 task conditions) to better demonstrate the method’s capability? Data splitting: The manuscript should explicitly state “strictly subject-level data splitting” in the Methods section to preclude any data leakage concerns. Ablation analysis: The 12% performance drop when substituting Mamba with a Transformer warrants careful examination. The current analysis is unconvincing. The manuscript should also temper its claim that Mamba is “inherently” superior by such a substantial margin, given the limited hyperparameter tuning performed for the Transformer baseline in the fMRI context.
Review #2
- Please describe the contribution of the paper
1.A hierarchical spatiotemporal architecture (Brain-TM) combining GIN, Transformer, and Mamba for cognitive task classification by decoding from dynamic functional connectivity, achieving state-of-the-art performance on HCP. 2.The study shows that latent brain states learned by the model correspond to distinct cognitive tasks, with functional network activation patterns varying systematically across temporal window sizes. 3.The study also reveals that global brain network integration is associated with localized hub activations rather than distributed activity, validated through Gini coefficient analysis and network centrality benchmarking
- 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.State-of-the-art classification performance with strong ablation evidence justifying each architectural component. 2.Interpretability is built into the architecture, with attention weights serving dual roles as both computational mechanisms and biological readouts. 3.Hub-driven findings are consistent with established neuroscience literature (DMN, Salience, Control networks).
- 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.
1.5-fold cross-validation is used for classification performance, but the interpretability analysis in 2.3 appears to be done on the full dataset without cross-validation. The paper doesn’t describe how clustering or hub identification varies across folds. If hub identity shifts across folds, the discovery is fragile. 2.The number of brain states is constrained to equal the number of task classes, which almost guarantees task-state alignment by construction. What if the cluster number k is not 7? 3.No test across different parcellations (only Schaefer 200 + Yeo 17 used). Hub findings could be parcellation-dependent. 4.The paper doesn’t discuss if the discoveries truely reflect biology or they are caused by architecture bias. For example, cross-attention mask design directly induces hubs. High Gini in attention weights could simply reflect that some ROIs happen to have stronger initial BOLD signal variance, which the attention mechanism amplifies, rather than true biological hub organization. The authors could test whether the same hub pattern emerges with a different aggregation scheme (e.g. mean pooling instead of cross-attention). The authors could also compare hub identity against known neuroanatomical hubs (e.g. precuneus, posterior cingulate, angular gyrus). Another potential method is to run the same Gini analysis on a shuffled label control (i.e. retrain the network with shuffled labels). If hub patterns persist even with randomized task labels, they reflect architecture not biology. 5.The claim that global functional integration is ‘driven by’ localized hub activations overstates what the evidence supports. The analysis relies on attention weight distributions and centrality metrics derived from a model trained on task classification, which are correlational observations, not causal evidence. 6.It would be helpful to clarify whether the node features projected into the embedding space in GIN are the thresholded functional connectivity values between each node and all other nodes.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While the classification performance is compelling, a major contribution of this paper is its biological interpretability claims, which are not sufficiently substantiated. First, the interpretability analysis lacks reproducibility: the hub identification is not cross-validated across data splits to assess stability. Second, the paper does not rule out architecture bias as an alternative explanation for the observed hub patterns.
- 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.
My first concern (R2.1) is addressed, whereas my opinions on other concerns stay the same. I recommend acceptance. However, I would like to point out that the hub identification can be biased by model architecture and hyperparameter selection, which is not addressed by comparing discoveries across different settings. With that being said, the work is innovative and brings value to the community.
Review #3
- Please describe the contribution of the paper
This paper proposes a hierarchical spatiotemporal modeling framework termed Brain-TM, which, for the first time, systematically integrates a Graph Isomorphism Network, a Hierarchical Transformer, and the Mamba selective state space model into a unified architecture. The framework adopts a three-stage progressive design: first, a GIN is employed to encode node-level topological features from dynamic functional connectivity matrices constructed via a sliding window; second, a Cross-Attention mechanism augmented with an anatomical structure mask strictly aggregates the 200 brain regions into 17 functional networks, followed by self-attention to model inter-network interactions; third, a Mamba module performs efficient long-range temporal modeling on the sequence of network-level features, capturing the dynamic dependencies underlying transitions between brain states while maintaining linear complexity. In extensive validation on the HCP dataset encompassing seven cognitive tasks, Brain-TM achieves a classification accuracy of 94.7%, outperforming static baselines such as GAT, GCN, and BrainNetCNN as well as the brain-specific model BrainGNN. Ablation experiments confirm that the Mamba module contributes most significantly to performance, with accuracy dropping sharply to 82.7% upon its removal. Regarding interpretability, the paper back-projects temporal features onto spatial attention weights to construct a three-tiered analytical pipeline: the first tier identifies seven latent brain states via feature clustering, demonstrating significant correspondence with the seven cognitive tasks and revealing a temporal evolution pattern wherein short windows rely on sensorimotor networks while longer windows engage default mode and control networks; the second tier introduces Gini coefficient analysis of attention distributions, uncovering that global network integration is predominantly driven by localized hub activations within the default mode, salience, and control networks, thereby establishing a “hub-driven” distributed expression pattern; the third tier validates through graph-theoretic centrality tests and subnetwork reconstruction that the top 51 high-contribution hubs exhibit significantly higher degree, eigenvector, and betweenness centrality than random controls, confirming their role as the stable core skeleton of the whole-brain functional connectome.
- 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 core strength of Brain-TM lies in the anatomical structure mask constraint mechanism introduced within its spatial aggregation module: by embedding a binary mask derived from the Schaefer-17 functional atlas into the Cross-Attention operation, the framework enforces that the representation of each functional network is generated exclusively from the weighted embeddings of the brain regions that anatomically belong to it, thereby eliminating cross-network information contamination. This hard-constraint design not only ensures the biological fidelity of the subsequent Intra-Net Attention weights—allowing the contribution of each region to its network to be rigorously quantified—but also provides a generalizable embedding paradigm for incorporating prior neuroanatomical knowledge into deep graph models, effectively bridging the semantic gap between data-driven feature learning and established functional system boundaries.
- 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.
Major Weakness 1: Outdated and Unfair Baseline Comparisons, Failure to Align with True State-of-the-Art Methods. In Table 1, the authors compare Brain-TM against GAT (Veličković et al., 2018), GCN (Kipf & Welling, 2017), GraphSAGE (Hamilton et al., 2017), BrainNetCNN (Kawahara et al., 2017), and BrainGNN (Li et al., 2021). All of these methods are static functional connectivity classification models, whereas Brain-TM takes dynamic functional connectivity sequences as input. This comparison is severely imbalanced in terms of information content and cannot substantiate claims of architectural superiority. Within the past three years, several widely recognized strong baselines for dynamic brain network modeling have emerged. For instance, STAGIN(Kim et al., NeurIPS 2021) reported approximately 96.8% accuracy on HCP seven-task classification, and BolT (Bedel et al., ICCV 2023) achieved around 97.2%. In the realm of static feature engineering, the work of Sripada et al. (High-accuracy machine learning techniques for functional connectome fingerprinting and cognitive state decoding, Nature Communications or related conference versions) has even reported classification accuracies exceeding 99%. The 94.7% accuracy reported for Brain-TM, in the absence of direct comparison with the aforementioned state-of-the-art methods, is far from sufficient to support its claim of “superior performance.” Major Weakness 2: Circular Reasoning in Brain State Clustering Analysis and Lack of Independent Quantitative Validation. The authors perform K-Means clustering on the intermediate features output by the Mamba module, determine the number of clusters K=7 using the elbow method and silhouette coefficient, and claim that the seven clusters correspond strongly to the seven HCP tasks. However, the model’s classification head itself contains seven neurons (corresponding to the seven tasks). After supervised training with cross-entropy loss, the feature space is inevitably partitioned into seven linearly separable subspaces. Performing K=7 clustering on this feature space and finding alignment with task labels is an inevitable mathematical consequence of model convergence, rather than an objective discovery of naturally emergent brain state numbers. Furthermore, the authors report no quantitative metrics—such as Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), or Cluster Purity—to assess clustering quality against ground-truth labels, relying solely on visual inspection of diagonal intensity in a heatmap, which is statistically insufficient. No sensitivity analysis is provided to examine whether the state-task correspondence remains stable if the number of clusters is changed to, for instance, K=5 or K=10.Major Weakness 3: Complete Absence of Generalizability Validation; the Model May Be Severely Overfitted to HCP Data Acquisition Characteristics. All experiments are conducted exclusively on the HCP Young Adult dataset, with performance evaluated via internal five-fold cross-validation. HCP data are renowned for their high temporal resolution (TR = 0.72 s), multi-band multi-echo acquisition sequences, and advanced ICA-FIX denoising, resulting in substantially higher signal-to-noise ratios than those found in most clinical or large-scale developmental cohorts (e.g., ABCD, UK Biobank, PNC). The authors provide no model transfer testing on any external dataset, nor do they design any cross-site or cross-acquisition-parameter validation experiments. Consequently, the reported 94.7% accuracy and the proposed “hub-driven” brain state mechanisms may well reflect idiosyncratic features specific to the HCP acquisition protocol rather than generalizable cognitive neuroscience principles. Major Weakness 4: Severe Imbalance Between Model Complexity and Marginal Performance Gain. Ablation experiments in Table 2 demonstrate that when the Hierarchical Transformer module is removed (i.e., retaining only GIN and Mamba, directly modeling temporal sequences across the 200 brain regions), classification accuracy remains at 93.73%, falling short of the full model (94.71%) by less than one percentage point. This negligible difference raises fundamental questions about the necessity of the hierarchical aggregation mechanism: introducing complex anatomically masked Cross-Attention and the 17-network inductive bias yields minimal practical performance improvement while substantially increasing computational overhead and model complexity. The authors attribute the primary value of the hierarchical module to “enhanced interpretability,” but they do not demonstrate whether similar hub localization results could be obtained from the flattened model via gradient or attention analysis. If comparable interpretations can be derived from the flattened model, then the current architecture exhibits clear signs of over-engineering.
- 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.
(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?
This paper proposes the Brain-TM framework, which integrates a GIN, a Hierarchical Transformer, and the Mamba module at the architectural level, and devises a three-tiered progressive analysis pipeline for interpretability evaluation. These technical attempts exhibit a degree of novelty and engineering merit. However, the following critical shortcomings collectively constitute the core rationale for a Weak Reject recommendation:
1.Severely Insufficient Experimental Comparisons That Fail to Substantiate the Claim of “Superior Performance.” The authors compare their model only against static graph methods from the 2017–2021 period, such as GAT, GCN, and BrainNetCNN, without including any contemporaneous or earlier state-of-the-art baselines designed for dynamic functional connectivity (e.g., STAGIN, NeurIPS 2021, which reported approximately 96.8% accuracy on HCP seven-task classification; BolT, ICCV 2023, which achieved around 97.2%). They also fail to acknowledge the work of Sripada et al. , which attained >99% accuracy using static feature engineering. Against this backdrop, the reported 94.7% accuracy cannot be regarded as cutting-edge performance, and the fairness of the comparison is questionable.
2.Circular Reasoning in Brain State Clustering and a Lack of Quantitative Validation. The model’s classification head is preset with seven task categories. Conducting K=7 clustering on the feature space derived from such strongly supervised training and subsequently observing alignment with the task labels is an almost inevitable mathematical outcome rather than an objective discovery of naturally emergent brain states. The authors report no quantitative metrics for clustering quality—such as Adjusted Rand Index (ARI) or Normalized Mutual Information (NMI) —nor do they test the stability of the state–task correspondence under alternative values of K. Consequently, the conclusion that “the brain possesses seven natural latent states” lacks persuasiveness.
3.Complete Absence of Generalizability Validation. All experiments are conducted exclusively on the single-site HCP Young Adult dataset. No external data (e.g., ABCD, UK Biobank, PNC) or cross-site evaluation is employed to assess the model’s transferability. The high temporal resolution and specialized preprocessing pipeline characteristic of HCP data may lead the model to overfit acquisition-specific idiosyncrasies, thereby undermining the generalizability of the reported brain-state mechanisms to broader populations.
4.Mismatch Between the Marginal Performance Gain and the Complexity of the Hierarchical Aggregation Module. Ablation experiments reveal that removing the Hierarchical Transformer module reduces accuracy by less than one percentage point (to 93.73%), indicating that this component contributes only modestly to final classification performance. Although the authors emphasize its value for interpretability, they do not demonstrate whether comparable explanatory insights could be obtained from the flattened model via alternative attribution methods. The necessity of this module therefore remains insufficiently justified.
In summary, although the work possesses commendable aspects in terms of methodological integration and interpretability visualization, the core deficiencies in experimental design severely undermine the credibility of its conclusions and the overall contribution. I therefore recommend Weak Reject, and I encourage the authors to resubmit after incorporating fair comparisons with dynamic state-of-the-art methods, adding generalizability validation, and rectifying the circular reasoning issues in the clustering analysis.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Dear Reviewer, Area Chairs, and Program Chairs, We thank the meta-reviewer (MR) and Reviewer (R) 1&2&3s’ constructive reviews (e.g. Rx.x). We have synthesized our responses into 5 key thematic areas (T1-T5).
T1: Comparison with Dynamic Spatiotemporal Models (MR, R1, R3) R1.1: Brain-TM (94.71%) outperforms current dynamic spatiotemporal models like ST-GCN (93.42%) and Brain Network Transformer (94.37%). R3.1: Although STAGIN achieves higher accuracy with a finer atlas, we prioritize brain interpretability analysis. Unlike STAGIN, which relies on high-resolution atlases, we use the widely adopted, standard Schaefer 200 for ROI-to-17-network mapping, thereby achieving more robust biological interpretability and task-specific hub identification.
T2: Sliding Window Impacts, Clustering Validity, and Task-Driven Brain States (MR, R1, R2, R3) R1.4: For sliding window impacts, our sensitivity analysis shows moderate windows optimize accuracy by minimizing temporal blurring. Furthermore, varying window sizes yields distinct neurobiological insights (Fig. 2C): short windows emphasize rapid-response networks, longer windows highlight sustained integration. R2.2&R3.2: We clarify that our goal is to decode task related states, not claim ‘seven natural latent states’. And the clustering results depend on objective quantitative indicators, rather than visual heatmap inspection (“natural” and “heatmap” did not appear in the original manuscript). We combined Silhouette Score and Inertia Curve to explore K=1 to 15, with quantitative metrics peaking exactly at K=7 and delivering the optimal clustering partition.
T3: Cross-Validation Protocol and Atlas/Dataset Generalizability (MR, R1, R2, R3) R1.2&2.1: Our 5-fold CV is strictly subject-level: All sliding windows of an individual subject are assigned entirely to training or testing sets, preventing data leakage. Our hub identification shows high stability across all 5 CV folds, hub locations are consistent across test sets, the 94.71% accuracy reflects true cross-subject generalization. While our visualization used one fold due to space limits, cross-fold statistical consistency was rigorously verified. R2.3&R3.3: We acknowledge parcellations may cause minor boundary variations, but the Schaefer 200 was chosen as a widely adopted dynamic connectomics standard. Our node-to-network dual analysis mitigates parcellation bias and ensures neuroanatomically reliable findings. The HCP dataset is the neuroimaging community’s benchmark, supporting interpretation and direct literature comparison. Validating diverse datasets is added to our Future Work.
T4: Architectural Bias in Hub Identification and Hierarchical Necessity (MR, R2, R3) R2.4&R2.5: For architectural bias, we clarify our design uses prior knowledge to explore functional hubs, not enforce them. GIN embedding prevents BOLD variance amplification, and masks only limit affiliations. Thus, the identified hubs (posterior cingulate, prefrontal cortex) are strictly data-driven. Crucially, mean pooling reduces accuracy but still yields highly overlapping hubs via centrality metrics, while label-shuffling destroys these patterns entirely. Lastly, we have replaced all causal phrasing in the manuscript with correlational terms. R3.4: For Hierarchical Necessity, removing the module drops accuracy to 93.73%, and renders the model a black box with unstable attention weights. The 17-network bias is not over-engineering, but a vital constraint for stable, biological hub localization.
T5: Clarification on Ablation and GIN (R1, R2) R1.3: The Transformer baseline was grid-searched (lr: 1e-4~1e-3&dropout: 0.1~0.5). The performance gap is not due to suboptimal baseline configuration, mainly due to Mamba’s inherent superiority in capturing dynamic fMRI sequences. R2.6: We clarify GIN node features are derived from functional connectivity profiles. At window t, each node v is represented by the corresponding row of thresholded matrix A_t, and updated via Eq. (1).
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.
Although the proposed hierarchical spatiotemporal framework and its interpretability analysis are interesting, critical evaluation flaws necessitate a rebuttal. The authors should address: (1) unfair baseline comparisons by including recent dynamic spatiotemporal SOTA models; (2) circular reasoning and lack of quantitative metrics in the clustering-based interpretability analysis; (3) clarifications on subject-level cross-validation to rule out data leakage; (4) potential architecture bias in hub identification and the marginal performance gain of the hierarchical module; and (5) the lack of cross-dataset generalizability validation.
- 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 clarified the strict subject-level data splitting to rule out data leakage (R1, R2), explained that the clustering analysis was quantitatively driven (R2, R3), and justified the architectural design choices regarding baselines and the necessity of the hierarchical module for biological interpretability (R2, R3). Reviewers 1 and 2 explicitly upgraded their scores to Accept, recognizing the method’s innovation and the value of its built-in interpretability for the neuroimaging 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.
Reject
- Please justify your recommendation.
This is a borderline case. While the rebuttal addresses many concerns related to the model tuning and experimental setup, in my opinion, the main weakness of the original paper was the lack of appropriate spatio-temporal baselines. The authors provided such comparisons in the rebuttal, but this would constitute a new experiment, which not allowed during the rebuttal phase. On a side note, the performance gain over ST-GCN and BNT is minimal and unlikely to be statistically significant. Ignoring this new set of comparisons, I would lean towards rejection of the original paper due to the lack of appropriate baselines. Finally, there are other reviewer concerns that have not been addressed in the rebuttal (e.g., circular reasoning in the analysis, lack of generalizability, complexity vs performance).
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 rebuttal sufficiently addressed the reviewers’ questions regarding clarifying details of the experiments (validation and comparison with other methods).