acceptodds
Under review as a conference paper at ICLR 2027

HyperEvo: History-Aware Discrete Hypergraph Topology Evolution for Enhanced Brain Disease Classification

Abstract

Functional brain networks exhibit complex high-order interactions and temporally evolving patterns, and their abnormal reorganization may provide important information for brain disease classification. However, existing methods often rely on predefined or fixed topologies, which may fail to capture disease-associated reorganization of dynamic node-hyperedge relationships across temporal states. Moreover, historical functional states are rarely exploited to guide topology evolution, limiting the ability to characterize the temporal evolution of functional brain networks. To address these problems, we propose HyperEvo, a history-aware discrete hypergraph topology Evolution framework that improves brain disease classification by modeling dynamic functional brain networks. First, we construct window-level sparse hypergraphs from resting-state functional magnetic resonance imaging (rs-fMRI) signals and learn high-order functional representations through hypergraph convolution. Then, we aggregate preceding-window information through a history-aware topology evidence mechanism and perform budget-constrained discrete topology evolution by explicitly adding and deleting node-hyperedge memberships. Finally, we employ a multi-window kernel attention mechanism to integrate functional representations across temporal states and capture discriminative spatiotemporal patterns in dynamic functional brain networks. Extensive experiments on the ADNI and PPMI datasets demonstrate that HyperEvo achieves superior performance across multiple brain disease classification tasks. Further analyses show that the learned dynamic hypergraphs highlight interpretable brain regions and high-order functional connectivity patterns associated with disease-related functional alterations.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Functional Brain Network, Hypergraph Learning, Dynamic Brain Network, Topology Evolution, fMRI, Brain Disease Classification

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

This paper proposes HyperEvo to address the limitations of fixed hypergraph topologies and insufficient temporal modeling in brain network analysis. Specifically, it introduces a history-aware mechanism to guide discrete node–hyperedge membership updates and employs multi-window attention to integrate temporal features. Experiments on ADNI and PPMI datasets demonstrate improved brain disease classification performance and promising interpretability.

Critical strengths
  1. Experiments across eight classification tasks on ADNI and PPMI, together with ablation and parameter sensitivity analyses, demonstrate the effectiveness of the proposed framework.

  2. The visualization of discriminative hyperedges and brain regions provides insights into disease-related functional network alterations.

Critical weaknesses
  1. The methodological novelty of the proposed framework is limited. Modeling dynamic brain activity using historical states has been extensively explored [1,2]. Moreover, hypergraph-based approaches have been widely adopted for dynamic brain network analysis [3], and recent studies have further investigated adaptive optimization of hypergraph nodes and hyperedges [4]. Therefore, the authors should more clearly distinguish the proposed method from existing studies and highlight its unique contributions.

[1] Asynchronous Functional Brain Network Construction with Spatiotemporal Transformer for MCI Classification. IEEE TMI, 2024.

[2] Dynamic Graph Representation Learning for Spatio-Temporal Neuroimaging Analysis. IEEE TCYB, 2025.

[3] HyperDiag: Temporal–Regional Hypergraph Learning via Topology-Enhanced State Propagation for Brain Disease Diagnosis. AAAI, 2026.

[4] Learning High-Order Relationships with Hypergraph Attention-Based Spatio-Temporal Aggregation for Brain Disease Analysis. IEEE JBHI, 2026.

  1. The framework relies on sparse reconstruction to initialize hypergraphs. The authors should investigate whether different initialization strategies significantly influence the learned topology and classification performance.

  2. The experiments are conducted on ADNI and PPMI without independent external validation. Additional cross-site or cross-dataset experiments would strengthen the generalization claims.

Questions for authors
  1. The history-aware attention and discrete topology editing mechanisms may introduce additional computational overhead. The authors should report model complexity, parameter counts, and runtime comparisons.

  2. The identified discriminative hyperedges and brain regions are mainly supported by visualization and literature-based interpretation. Further statistical analysis would improve the reliability of the neurological findings.

  3. The three-class classification results on PPMI differ between Tables 1 and 2. For example, the complete model achieves 62.17% accuracy in Table 1 but 62.21% in Table 2.

  4. Although parameter sensitivity experiments are provided, the selected addition and deletion budgets vary across classification tasks. The authors should further explain how these budgets are determined and whether they can generalize across datasets without task-specific tuning.

  5. In Table 2, the same module is referred to as “w/o WKAF” and “w/o WLAF” for the ADNI and PPMI datasets, respectively.

  6. In Table 2, the terms “NC” and “CN” are used interchangeably to refer to normal controls. The authors should carefully proofread the entire manuscript to avoid such minor errors.

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

This paper proposes HyperEvo, a history-aware dynamic hypergraph framework for brain disease classification. The method constructs window-specific sparse hypergraphs from rs-fMRI signals, aggregates information from previous temporal states, and adaptively modifies node–hyperedge memberships through discrete addition and deletion operations. The evolved representations are further integrated using window-level attention for downstream classification. Experiments on ADNI and PPMI report improved performance over several graph- and hypergraph-based baselines across multiple classification tasks.

Critical strengths
  • The explicit editing of node–hyperedge memberships is intuitively aligned with dynamic brain-network modeling. The ablation results also suggest that topology modification and temporal-history modeling contribute to classification performance.
Critical weaknesses
  • The claimed history-aware topology evolution appears only weakly differentiated from feature-level temporal modeling followed by heuristic hypergraph reconstruction. Historical information is aggregated through node representations, while topology updates are ultimately determined by cosine similarity and predefined Top-k addition/deletion budgets. It is therefore unclear what fundamentally new topology-learning principle is introduced beyond combining temporal feature aggregation with adaptive hypergraph editing.
  • The empirical evaluation is not sufficiently convincing. The experiments use only two relatively small cohorts, while the reported eight tasks are different label combinations from the same datasets. Moreover, several key hyperparameters, including window number, history length, topology-editing budgets, and sparsity coefficients, are selected separately for each task, but the paper does not clearly describe a nested validation protocol for these choices.
Questions for authors
  • Since historical information enters through node representations rather than previous incidence matrices, what distinguishes the proposed approach from temporal feature aggregation followed by adaptive hypergraph reconstruction?
Reviewer_K4SJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework for dynamic functional brain network analysis that explicitly updates hypergraph topology over time to improve brain disease classification. The proposed HyperEvo approach captures high-order interactions and temporal changes in functional brain states for improved diagnosis. HyperEvo first constructs window-specific sparse hypergraphs from resting state fMRI signals using LASSO-based reconstruction and extracts high-order representations through hypergraph convolution. It then combines adjacent window gating and multi-window historical attention to generate topology evidence. A straight-through estimator enables gradient based optimization of the discrete editing decisions. Finally, window-level attention integrates the evolved representations for classification. HyperEvo was evaluated on two public datasets, ADNI and PPMI, for eight binary and multiclass classification tasks. The experiments compare the proposed approach against conventional machine learning methods and several recent dynamic brain network models. Results are presented and analyzed.

Critical strengths

The explicit modeling of discrete hypergraph topology evolution using historical functional information is a strength of this paper. Unlike approaches that primarily update node representations or adjust the weights of existing hyperedges, HyperEvo allows node hyperedge memberships to be added and removed. This provides a potentially useful mechanism for capturing changes in high order brain connectivity. The combination of historical state aggregation, budget-constrained topology editing, and re-encoding on the evolved hypergraph forms a nice framework.

Critical weaknesses

My main concern is that this paper does not sufficiently establish that the proposed history-aware discrete topology evolution provides a fundamentally new or necessary modeling capability beyond existing adaptive graph and hypergraph learning methods. The novelty seems limited. Individual components are largely established techniques. The paper does not adequately distinguish its contribution from recent approaches to dynamic graph topology learning and temporal hypergraph adaptation (e.g., HyperDiag and spatiotemporal evolutionary graph learning methods). The current ablations mainly remove individual modules rather than comparing the proposed discrete editing mechanism against simpler, parameter-matched alternatives (e.g., continuous incidence learning, adaptive hyperedge weighting, or topology editing without historical guidance). It remains unclear whether the performance improvements specifically arise from discrete structural evolution or from the additional representation capacity and temporal modeling components. Besides, the datasets are relatively small, with 245 ADNI samples and 156 PPMI samples, and the evaluation relies on stratified 10 fold cross-validation without external validation. I have some concerns about possible selection bias and the robustness of the reported improvements.

Questions for authors
  1. What is the precise methodological novelty relative to existing dynamic graph and hypergraph topology learning approaches? How does HyperEvo differ conceptually and algorithmically from HyperDiag and spatiotemporal evolutionary graph learning?

  2. Is discrete topology editing necessary, or could similar improvements be obtained using continuous adaptive incidence matrices? Could you compare against parameter matched continuous topology learning, dynamic hyperedge reweighting, and history-independent Top-k editing?

  3. How are hyperparameter selection and cross-validation organized to prevent information leakage? Were these values selected using an independent validation set or nested cross-validation within each training fold? Were comparable tuning budgets provided to the baselines?

  4. Can your provide stronger evidence for the biological interpretability of the evolved hypergraphs? The current analyses identify disease associated regions and hyperedges, but do not quantify their consistency across subjects or validation folds.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.