acceptodds
Under review as a conference paper at ICLR 2027

HADBN: Learnable Hierarchical Aggregated Dynamic Brain Networks for Interpretable Brain Disorder Diagnosis

Abstract

The complex spatio-temporal correlations in brain networks make brain disorder diagnosis challenging. Many spatio-temporal models underutilize anatomical constraints from brain spatial topology, limiting their ability to capture spatial relationships and dynamic interactions. To address this, we propose a novel framework, Learnable Hierarchical Aggregated Dynamic Brain Networks (HADBN), which incorporates multi-level anatomical priors of brain spatial topology and jointly learns multi-scale spatial representations and time-varying connectivity patterns. Specifically, HADBN constructs spatio-temporal interaction graph neural networks (GNNs) with a three-level hierarchical brain architecture—local ROI, middle sub-network, and global network levels—guided by functional parcellation atlases. Unlike existing multi-level spatio-temporal models, HADBN introduces learnable F-matrix aggregation modules that perform data-driven pooling across spatial scales, replacing fixed anatomical aggregation. Moreover, it employs a feature-driven temporal attention mechanism that computes adaptive window weights from multi-scale features, thereby enabling principled spatio-temporal integration and cross-level information flow. Experiments show that HADBN consistently outperforms state-of-the-art methods under 10-fold cross-validation, while the learned aggregation weights offer interpretable insights into disorder-relevant brain networks and support biomarker discovery. These results establish HADBN as a robust and interpretable framework for brain disorder diagnosis.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spatio-temporal graph neural networks, Hierarchical aggregation, Dynamic brain networks, Brain disorder diagnosis

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_w2QyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limitations of existing dynamic brain network methods in modeling hierarchical spatial organization and temporal dependencies. To address these challenges, the authors propose HADBN, a learnable hierarchical graph learning framework for brain disorder diagnosis. Specifically, it introduces learnable F-matrices for multi-level brain network aggregation, dynamic attention-weighted graph convolutions for spatiotemporal modeling, and cross-layer attention for information integration. Experimental results demonstrate competitive classification performance and provide interpretable brain network representations.

Critical strengths
  1. The proposed framework incorporates brain atlas priors to construct multi-level representations, providing a reasonable foundation for learning brain network organization.

  2. The learnable F-matrices enable data-driven aggregation across different spatial scales, improving the flexibility of conventional fixed hierarchical pooling.

Critical weaknesses
  1. Hierarchical brain network modeling, learnable graph pooling, temporal attention, and cross-layer feature interaction have been extensively explored. The proposed framework mainly integrates existing techniques, and its unique methodological contributions require clearer justification.

  2. Numerous methods for hierarchical dynamic brain network analysis have been proposed. The authors should include comparisons with these methods to better demonstrate the effectiveness of the proposed framework.

  3. In Table 3, removing the F-weights only moderately reduces performance, particularly on ADHD-200. Further comparisons with fixed anatomical pooling and alternative learnable pooling methods are needed to demonstrate the necessity of the proposed F-matrices.

Questions for authors
  1. Section 3.1 states that the F-matrices are constrained to be non-negative and normalized to unit row sums. Under these conditions, the L1 norm of each row is constant, making the additional L1 sparsity penalty potentially ineffective.

  2. The authors select a window length of 40 TR and a stride of 10 TR based on temporal similarity statistics. However, higher adjacent-window similarity does not necessarily indicate better preservation of meaningful brain dynamics. The authors should further evaluate how different window settings affect classification performance.

  3. The framework incorporates age, sex, IQ, and handedness into the final classification features, but their independent contributions are not evaluated. The authors should provide modality ablations to distinguish improvements from brain network modeling and phenotypic information.

  4. On Page 9, “It alse identified” should be corrected to “It also identified”. The authors should carefully proofread the manuscript to avoid such errors.

Reviewer_wemSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Overall assessment: Integrating multiscale brain networks, learnable aggregation, and temporal modeling is a worthwhile research direction. The paper presents experiments on two disorder classification tasks and includes component ablations. However, the current manuscript has substantial weaknesses concerning its performance claims, the validity of its loss functions, the definition of its hierarchical mechanisms, and the evidence supporting generalization.

Critical strengths

None

Critical weaknesses

1.The framework is evaluated only on ABIDE-I and ADHD-200, comprising 1,025 and 541 participants, respectively. The primary concern is not simply the sample size, but the limited scope of evaluation: both tasks involve established multisite neurodevelopmental disorder cohorts. These experiments provide insufficient evidence for broad applicability or robust generalization. Cross-validation on pooled multisite data does not establish performance on previously unseen sites. I recommend adding a third independent cohort. If additional data are unavailable, the authors should at least conduct leave-one-site-out or site-grouped outer validation on the existing datasets, report site-specific performance with confidence intervals, and assess the effects of acquisition protocols, age distributions, and diagnostic categories. Claims of broad applicability should be narrowed if the evaluation remains restricted to these two tasks. 2.The abstract states that HADBN “consistently outperforms state-of-the-art methods.” However, on ABIDE-I, HADBN achieves an accuracy of 68.14%, below BrainGNN’s 69.31%, and an AUC of 69.32%, below BrainGNN’s 69.52%. On ADHD-200, HADBN achieves the highest accuracy, but its AUC of 73.48% is lower than STIGR’s 78.11%. The results therefore do not support consistent superiority across datasets and metrics. The abstract, experimental discussion, and conclusion should be revised to distinguish clearly between metrics on which HADBN improves and those on which it does not. Performance differences should be assessed using paired comparisons on identical data splits, with appropriate confidence intervals. Rankings based solely on mean performance are insufficient to establish statistically significant superiority. 3.The introduction and related work repeatedly discuss static versus dynamic connectivity, fixed aggregation, and limited cross-scale interaction. This extended general background obscures the central questions: which specific limitation of existing hierarchical methods is addressed, what F-based aggregation changes, and why those changes should improve interpretability or generalization. Citation accuracy also requires attention. Section 3.1 attributes the Yeo atlas to Dickstein et al. (2006), but the corresponding reference concerns an ADHD neuroimaging meta-analysis. The reference list additionally contains duplicate entries. The background should be reorganized around the task requirements, a precisely defined methodological gap, and testable hypotheses. Repeated material should be consolidated, and a concise comparison table could distinguish HADBN from existing hierarchical models in terms of structural priors, aggregation, temporal modeling, and interaction direction. Each claimed contribution should map directly to a methodological component and supporting experiment. Atlas citations and duplicate references should be checked throughout. 4.In the presented attention formulation, queries originate from local representations, while keys and values originate from global representations. The output updates the local representations, explicitly implementing global-to-local information transfer. No corresponding attention operation that updates the global representations is shown, despite the conclusion claiming “bidirectional information flow.” Bottom-up aggregation alone does not establish bidirectional cross-attention. The authors should clarify whether “bidirectional” refers to the overall hierarchical computation or a bidirectional attention module. If reverse attention is implemented, its equations and update order should be provided; otherwise, the terminology should be revised consistently. Comparisons among no interaction, global-to-local interaction, local-to-global interaction, and bidirectional interaction would provide stronger support than removing the entire attention module alone. 5.The hierarchy combines 90 AAL regions, 17 intermediate networks, and seven global networks. However, the mapping from AAL regions to Yeo networks is insufficiently described, particularly for subcortical regions. It remains unclear how outputs from differently sized subsets are assembled, whether assignments can cross predefined network boundaries, and whether parameters are shared across time windows and participants. Consequently, the distinction between learning a hierarchical structure and learning aggregation weights within a fixed hierarchy is unresolved. The authors should provide the complete mapping, masking rules, tensor dimensions, and forward-pass pseudocode, explicitly separating fixed priors from learnable components. Comparisons with fixed mean aggregation and parameter-matched learnable pooling are needed. The paper should also compare its mechanism and performance with relevant hierarchical approaches discussed in the manuscript, including MHNet and BrainHGT, to establish what the proposed hierarchy contributes. 6.The overall objective combines the classification loss with four auxiliary components, while the F-related regularizer itself contains five terms. This introduces substantial tuning complexity, yet Table 3 provides no loss-by-loss ablation. More importantly, the manuscript states that each F-vector is constrained to be nonnegative and sum to one through ReLU and normalization. If these constraints hold exactly, its L1 norm is identically one, rendering the sparsity term in Equation (11) ineffective, while the row-sum penalty in Equation (12) is identically zero. This raises a fundamental question about the validity of the objective, beyond concerns about complexity alone. The authors should clarify whether these losses are computed before or after normalization, specify any numerical stabilization, and explain how all-zero vectors are handled. Starting from the classification-only objective, they should evaluate the incremental contribution of each auxiliary term and report weight sensitivity and gradient magnitudes. Constant or redundant penalties should be removed. If sparse aggregation is intended, the chosen mechanism must remain effective under the imposed constraints and be supported experimentally.

Questions for authors

Are the F-vectors constrained to be nonnegative and sum to one before computing Equations (11) and (12)? If so, the L1 penalty is constant and the row-sum penalty is zero. How do these terms contribute to optimization?

Does Equation (3) operate on fixed Pearson-correlation inputs or learned representations? Which trainable parameters receive gradients from this loss?

What aspects of the hierarchy are learned beyond aggregation weights within predefined networks? Please clarify the AAL-to-Yeo mapping, parameter sharing across participants and windows, and whether reverse cross-attention is implemented in addition to the global-to-local update shown in Equations (6)–(7).

How should the claim of consistently outperforming existing methods be reconciled with BrainGNN’s higher ABIDE-I accuracy and AUC, and STIGR’s higher ADHD-200 AUC? Do paired comparisons support statistically reliable improvements on the metrics where HADBN performs best?

Do existing analyses establish performance on unseen sites and separate the contributions of imaging features from phenotypic information? These results would materially affect my assessment of generalization and architectural benefit. A third independent cohort would be a useful future extension, but is not essential to answering these questions.

How are regional importance scores computed from F-weights, and what evidence establishes their stability and fidelity to model predictions? Without such evidence, would the authors narrow the biomarker claims to candidate discriminative regions?

Reviewer_bEPCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a hierarchical spatiotemporal graph neural network framework HADBN for brain disorder diagnosis using resting-state fMRI. The method constructs three levels of brain representations, i.e., local regions of interest, intermediate subnetworks, and global functional networks. The key contribution is a learnable hierarchical aggregation mechanism based on F-matrices; it aims to combine anatomical priors with data-driven pooling. The framework also integrates dynamic attention-weighted graph convolution, temporal attention, cross-layer attention, and phenotypic information to learn discriminative representations for brain disorder classification. It is evaluated on ABIDE-I and ADHD-200.

Critical strengths

This paper addresses brain network representation learning by attempting to incorporate hierarchical anatomical organization into dynamic functional connectivity analysis. The combination of atlas-guided brain hierarchies with learnable aggregation provides a potentially useful way to balance prior anatomical knowledge and data-driven representations.

Critical weaknesses
  1. Methodological novelty is limited. The learnable F-matrix aggregation mechanism is presented as a central contribution, but its advantages over standard attention-based pooling, differentiable graph pooling, or fixed atlas-guided aggregation are not convincingly demonstrated. The paper also discusses closely related hierarchical approaches (e.g., MHNet, BrainHGT, and DHGFormer) but does not include direct experimental comparisons. Besides, there are potential inconsistencies in the mathematical formulation, which needs to be double-checked. For instance, Equations 6&7 construct queries from local-level representations and keys and values from global-level representations. This formulation allows local representations to incorporate global information, but does not explicitly implement the reverse direction. The claimed bidirectional interaction is therefore not adequately supported by the presented architecture. Appendix A introduces both an L1 sparsity penalty and a row-sum regularization term; if the constraints stated in Section 3.1 are enforced, the L1 norm of each aggregation vector is constant and the row-sum penalty is zero. This is another concern about the technical soundness.

  2. Experimental results do not substantiate the claimed state-of-the-art performance. The paper repeatedly claims that HADBN consistently outperforms existing methods, but the presented results (e.g. Table 2) do not always support this conclusion. This discrepancy is concerning because superior classification performance is a primary justification for the proposed architectural complexity. Besides, the evaluation is restricted to two multi-site datasets. The authors also state that baseline hyperparameters are taken from the original publications; this causes uncertainty about whether all methods received comparable optimization under the reported experimental protocol.

Questions for authors
  1. What is the (fundamental) methodological novelty of HADBN relative to existing hierarchical graph learning methods? Could you provide direct comparisons with MHNet, BrainHGT, DHGFormer, and standard learnable graph pooling mechanisms?

  2. How are the F-matrix normalization constraints and regularization terms implemented?

  3. Where is the bidirectional cross-level information exchange implemented? Could you identify the corresponding local-to-global mechanism and explain how it differs from conventional cross-attention?

  4. How robust are the results and learned biomarkers across imaging sites and experimental repetitions? Could you provide leave-one-site-out evaluation, clarify the hyperparameter selection procedure for both HADBN and the baselines, and quantify the stability of the learned aggregation weights across folds and random seeds?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.