acceptodds
Under review as a conference paper at ICLR 2027

Brain Disease Detection Based on Bottom-Up Hierarchical Graph Representation Learning

Abstract

Current research on brain disorders focuses on the functional connectivity networks of regions of interest (ROIs), but less attention has been paid to explicit brain functional cluster recognition, which neuroscience has identified as playing a crucial role in brain disease detection. In this paper, we propose a bottom-up hierarchical graph representation learning method for brain disease detection. Bottom-level brain graph representation with ROI node importance is designed to suppress noise for effective recognition of top-level brain functional clusters. Top-level brain graph representation with clustering consistency constraints explicitly characterizes functional clusters and interaction patterns. We demonstrate through theoretical analysis that bottom-level gating conditionally bounds perturbations in spectral alignment, while top-level clustering consistency constraints yield a conditional upper bound on the continuous normalized cut objective. Our model is validated on the publicly available neuroimaging datasets ADHD-200, OpenNeuro-ds002424, and OpenNeuro-ds000030, demonstrating significant advantages in brain disease detection. Code is available at https://anonymous.4open.science/r/Bottom-Up-Hierarchical-Graph-Representation-Learning-0670.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain Disease Detection, Hierarchical Graph Representation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6NqxOpenReview ↗
Rating
1
Confidence
4
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a bottom-up hierarchical graph representation learning framework for fMRI-based brain disease detection. The method first uses Transformer-based gating to estimate ROI importance and reweight node features, then learns soft assignments from ROIs to functional clusters and aggregates them into a coarsened cluster-level graph for downstream brain disease classification. A clustering consistency module further encourages the learned assignment subspace to preserve low-frequency Laplacian eigenvectors, for which the authors provide conditional theoretical connections to the continuous normalized-cut objective. The model is evaluated on multiple brain benchmarks, with additional ablations on the hierarchical components, the number of clusters, clustering stability, and computational complexity.

Critical strengths
  • The paper provides a biologically motivated hierarchical representation of brain functional organization. Explicitly learning ROI-to-cluster assignments makes the learned representation more interpretable than standard brain GNNs, and the authors provide theoretical and empirical analyses to support the design.
Critical weaknesses
  • The methodological novelty is not sufficiently established against recent high-order brain graph methods. Since the main contribution relies on explicitly modeling functional clusters and their higher-order interactions, comparisons with recent high-order approaches such as MuHL [1] and IBAHGT [2] are important for demonstrating the advantage of the proposed formulation.

[1] Sim, Jaeyoon, et al. "Learning Multi-Scale Hypergraph for High-Order Brain Connectivity Analysis." ICML (2026).

[2] Feng, Jingxi, et al. "Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis." NeurIPS(2026).

  • The empirical evidence is not sufficiently convincing. Several strong recent baselines are missing, and the reported p-values test performance against 0.5 rather than improvements over competing methods. Moreover, the large and inconsistent variances make it difficult to determine whether the reported gains are statistically reliable.
Questions for authors
  • How does the proposed method compare with recent high-order brain graph methods such as MuHL and IBAHGT? Such comparisons would materially affect my assessment of the methodological novelty.
  • The CCM formulation requires STSS^T S to be invertible, but the learned soft assignment does not guarantee full column rank. How often does SS approach rank deficiency in practice, and how is the inverse handled numerically during training?
Reviewer_kwnQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a bottom-up hierarchical graph representation learning framework for detecting brain diseases. The method combines a Transformer-based ROI importance-extraction module for noise suppression with a top-level graph-clustering module that explicitly learns functional clusters and their interactions via spectral-consistency constraints. The paper provides theoretical analyses connecting adaptive gating, spectral alignment, and the normalised-cut objective under specific assumptions. Experiments on three public neuroimaging datasets demonstrate improved classification performance over several existing baselines, supported by ablation studies and clustering analyses.

Critical strengths

-The paper addresses the important problem of explicitly modelling functional brain clusters rather than relying solely on implicit representations of ROI interactions. This provides a meaningful motivation for the proposed hierarchical architecture.

-The proposed method achieves the best reported accuracy and F1-score across all three neuroimaging datasets, with particularly substantial improvements on OpenNeuro-ds002424. These results suggest that the framework can improve brain disease classification.

-The paper provides theoretical results connecting ROI-level gating, spectral alignment, and the normalised-cut objective under specific assumptions. Although conditional, these analyses provide mathematical motivation for the proposed architecture beyond empirical observations alone.

Critical weaknesses

-The proposed framework combines existing techniques, including Transformer-based feature gating, differentiable graph pooling, and spectral regularisation. The paper does not sufficiently distinguish its contributions from established hierarchical graph learning methods, raising concerns about its novelty for a top-tier conference.

-The theoretical guarantees rely on assumptions such as regular graph structure and full-rank cluster assignments, which may not hold for real brain connectivity networks. The paper does not empirically verify these assumptions, limiting the practical relevance of the theoretical results.

-Although explicit functional cluster recognition is a central contribution, the reported connectivity separation is modest, and modularity values are close to zero. The lack of comparisons with established functional brain networks or clustering baselines makes it difficult to assess whether the learned clusters are neurologically meaningful.

-The reported p-values are based on one-sample tests against 0.5 rather than direct comparisons with competing methods, making the claimed performance improvements difficult to assess statistically. Furthermore, the ablation results are inconsistent across datasets, and additional comparisons with established graph pooling methods are needed to demonstrate the effectiveness of the proposed components.

Questions for authors

-What are the key methodological differences between the proposed framework and existing hierarchical graph pooling methods, particularly those incorporating spectral clustering or normalised-cut regularisation? How much of the reported improvement can be attributed specifically to the proposed spectral consistency constraint?

-To what extent do the brain connectivity graphs in the evaluated datasets satisfy the regularity and full-rank assumptions required by the theoretical analysis? Can the authors provide empirical evidence that these theoretical guarantees apply to the actual datasets?

-Given the modest intra-cluster versus inter-cluster connectivity separation and near-zero modularity reported in Section 5.4, what evidence supports the claim that the learned clusters represent meaningful brain functional communities rather than task-dependent latent representations?

Reviewer_a5r9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper combines ROI feature gating, soft graph coarsening, and spectral reconstruction regularization for disease classification. Conditional theoretical bounds accompany experiments on three neuroimaging datasets.

Critical strengths
  1. Mean accuracy and F1 improve across all three datasets, with a substantial ds002424 gain (Table 1).
  2. BL/TL ablations and inference perturbations are informative; the theory appropriately distinguishes spectral guarantees from biological validity.
Critical weaknesses
  1. Participant-independent evaluation is unestablished. Appendix D uses a sample-level random split and notes that graphs need not represent distinct participants; Appendix B acknowledges repeated observations. Participant-grouped results are necessary to interpret the strongest ds002424 gain (pp. 17, 21).
  2. NIEM attends across all graphs in a minibatch (Appendix E, pp. 25–26), making predictions potentially dependent on companion graphs. No batch-composition analysis is provided; single-graph perturbation tests do not validate this implementation’s behavior.
  3. Explicit clustering is insufficiently differentiated. BrainNetTF already uses soft clustering (manuscript ref. [8]) but is absent from Table 1. BL/TL ablations remove whole blocks; they neither isolate CCM nor compare matched clustering readouts.
  4. The full training objective is unspecified. Table 6 lists clustering, orthogonality, and balance weights without defining their losses; Section 5.1 introduces an undefined auxiliary classifier. These omissions obscure what was optimized and prevent reproduction from the manuscript (pp. 6, 26).
Questions for authors
  1. Can you report participant-grouped results and compare graph-local attention with varying inference batch compositions?
  2. Please define the complete objective and isolate CCM against BrainNetTF-style clustering under matched training.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.