acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Hypergraph Learning with Functional Alignment for Brain Disorder Diagnosis

Abstract

Graph Neural Networks (GNNs) have shown promise in brain disorder diagnosis by modeling neurophysiological signals as graphs, such as Electroencephalogram (EEG) and functional Near-Infrared Spectroscopy (fNIRS). However, existing brain graph learning methods remain limited by hierarchical anatomical heterogeneity, manifested as inconsistencies for alterations caused by brain disorders across different anatomical levels. Specifically, this heterogeneity can be analyzed from two perspectives: (1) Brain High-order interaction knowledge hierarchy heterogeneity, referring to inconsistencies in interactions among brain elements across anatomical levels. The inadequate capture of level-specific knowledge leads to unsuitable modeling. (2) Brain element functional role hierarchy heterogeneity, referring to inconsistency in the functional roles of the same element when viewed from different levels. This inconsistency leads to functional knowledge conflict, thereby impairing GNN cross-level generalizability. To address these challenges, we propose **HHRA**, a framework that addresses dual hierarchical heterogeneity. First, it introduces **H**ierarchical **H**igh-order Knowledge Capturing (HHKC), which enables the joint learning on multiple anatomical levels. Specifically, by constructing both pairwise and hypergraph relations across levels, HHKC effectively enables capturing level-specific high-order interaction knowledge. Second, it introduces Hierarchical Functional **R**ole **A**lignment (HFRA), which promotes the coordination across multiple anatomical levels. Specifically, it estimates and aligns the functional importance of brain elements across levels for consistent knowledge and a cross-level generalizable GNN. Extensive experiments on EEG and fNIRS datasets validate the superiority of HHRA. The code is available at <https://anonymous.4open.science/r/HHRA-6656/>.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Network, Medical Artificial Intelligence

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_9oCMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates methods for classifying brain disorders using electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS), with the aim of integrating interactions among different anatomical structures while analyzing the working mechanisms of the same functional channel from different structural perspectives. The proposed HHRA model consists of two modules: the HHKC module constructs a pairing map, a supermap of adjacent channels, and an induced anatomical supermap based on phase synchronization values at the channel level. From these three perspectives, information is transmitted between channel nodes via shared weights, and subsequent classification relies on the aggregation of these representations. HFRA integrates the variability of fractional Laplacian features with Fournette curvature into a single node role index and adjusts these indices according to different perspectives. The evaluation included binary classification tasks on three published EEG datasets and one proprietary fNIRS dataset, with results covering AUROC/AUPRC, view combination elimination, calibration elimination, and hyperparameter analysis. The complete model outperformed baseline models on most metrics, but the additional improvement provided by HFRA was relatively limited and did not necessarily constitute a significant advantage.

Critical strengths
  1. HHKC can simultaneously leverage higher-order relationships between each pair of connections, neighboring channels, and perceptual interactions between brain regions to represent different structural generalization biases. Appendix Table 3 clearly illustrates the division of brain regions from these three perspectives; while these three perspectives share information propagation parameters, they retain their respective classification representations (§2.1). This enables the method to achieve feasible architectural designs and provides a concrete interface for investigating the role of anatomical prior knowledge in learning brain signals.

  2. Table 1 reports AUROC values of 80.65, 92.22, and 88.04 for this method on the ADFTD, APAVA, and M-fNIRS datasets, respectively, demonstrating superior performance across most metrics. Table 2 also shows that the three-view model without HFRA outperforms the listed single-view and two-view configurations on the ADFTD and M-fNIRS datasets. These results confirm the practical value of incorporating structural views and provide strong empirical support for the adoption of this method. However, these results provide stronger support for multi-view modeling than for specific mechanisms of functional role alignment.

Critical weaknesses
  1. When using unit node and hyperedge weights, Equation (16) can be simplified to $ \kappa(v_i,e)=|e|-d_i+1, $ where (di)(d_i) denotes the number of adjacent hyperedges. The first perspective considers edges only when (Sij>0)(S_{ij}>0). If all non-diagonal PLV values are positive, a complete graph is formed, yielding $ \kappa^{(1)}(v_i)=4-K,\qquad \exp(-\beta\kappa^{(1)}(v_i))=\exp!\bigl(\beta(K-4)\bigr). $ For a 53-channel M-fNIRS dataset, the multiplier generated by the default value (β=1)(\beta=1) is approximately (1.9×1021)(1.9\times10^{21}). Figure 3 also evaluates (β=3,5,7)(\beta=3,5,7); when (β=3)(\beta=3), the multiplier is (e147)(e^{147}), which would cause an overflow if the exponential function were computed directly in float32. The paper does not describe methods for curvature normalization, exponential truncation, logarithmic domain handling, or higher-precision computations. Therefore, these equations cannot fully account for the stable results reported. Table 3 also omits channels 25, 28, and 29; if the implementation excludes these channels from the regional super-edge, the treatment of zero-degree cases in Equations (9) and (17) must be explicitly clarified.

    Furthermore, Equation (15) measures the diffusion-related variability in the learned representation; interpreting this as physiologically significant requires validation. The different roles played from different structural perspectives do not necessarily constitute a harmful conflict that should be eliminated. The paper lacks a comparative analysis with simpler consistency regularizers, as well as an explicit assessment of cross-level generalization. These issues directly affect the credibility of the paper’s most distinctive contribution.

  2. In Section 3.1, it is first noted that all datasets are partitioned independently by subject. Subsequently, the sample-level segmentation of M-fNIRS is explained. This is because the samples are not organized by subject. The appendix explicitly states that the dataset contains 1,090 subjects. If each sample corresponds to a single subject, then sample-level segmentation is feasible; otherwise, it must be explicitly ensured that there are no duplicate records or duplicate segmentations across different groups. However, in the current description, this point is not explicitly clarified, nor is evidence provided to address this omission.

    APAVA includes 23 subjects, while the TDBrain evaluation subset includes 50 subjects. Therefore, using an 8:1:1 split ratio may result in very small independent test queries. An excessive number of time windows does not increase the number of independent subjects, but it remains unclear whether the AUROC/AUPRC statistical units and the five seeds will affect subject-based segmentation methods. As for the cumulative effect of HFRA, more robust statistical evidence is needed to support it. After adding the 3D model, the AUROC for ADFTD increased by 1.14 points, the AUROC for M-fNIRS increased by 0.33 points, and the AUPRC for M-fNIRS decreased by 0.86 points (Table 2). These results indicate potential advantages, but they do not yet demonstrate consistent and universal benefits.

Questions for authors
  1. Please explain the exponential multiplier in the complete figure shown above and specify whether this implementation includes sparsification, curvature normalization, exponential clipping, logarithmic domain processing, or any high-precision computations not mentioned. Additionally, please provide the actual values (α) for the three views in a typical training batch, as well as the ranges for the curvature values, convolution exponents, classification loss, weighted alignment loss, and gradient norm. In Figure 3, does the value approach infinity when β is large? If there are discrepancies between the implementation and the formula, please provide a consistent definition and verify the results after making the necessary corrections.

  2. With the three-view architecture, classifier, training budget, and subject partitioning fixed, please compare the full HFRA with the non-aligned approach, diffusion-energy-only alignment, and the credible feature consistency regularizer. Please include a standard Laplace control group with (α=1)(\alpha=1) to evaluate fractional-order designs. Report paired comparisons and uncertainty estimates, rather than just individual means. Since the role index depends on the amplitude of the representation, also examine whether these gains primarily reflect changes in feature scale.

  3. Why should role differences of the same channel across different levels be aligned, rather than retained as complementary information? Please clarify whether this metric estimates model-specific importance or physiological functional importance, and provide corresponding validation—for example, the predicted relationship of its variation under controlled channel or regional perturbations. Please also provide an operational definition of “cross-level generalization.” If this refers to transfer to unseen levels or anatomical regions, appropriate holdout evaluations are required; if it refers to consistency across the three jointly trained views, please adjust the scope of the statement accordingly.

  4. Does each M-fNIRS sample correspond to a different, unique subject? Are there duplicate recordings or multiple segments from the same subject? Please provide the number of subjects, categories, and time windows in each partition for every dataset, and clarify whether these five seeds altered the initialization method, the partitioning method, or both. Are AUROC/AUPRC calculated at the time-window level or the subject level? If any subject appears in multiple partitions, the main comparisons must be rerun using a subject-independent partitioning method.

Reviewer_6SKAOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes HHRA, a hierarchical hypergraph framework for brain disorder diagnosis from EEG and fNIRS signals. The method constructs channel-level pairwise and hypergraph representations together with a region-induced hypergraph to capture interactions across multiple anatomical views. It further introduces a functional role alignment objective that combines fractional-Laplacian-based diffusion sensitivity and Forman–Ricci curvature to encourage consistent node roles across views. Experiments on four datasets report improvements over the evaluated time-series and graph-learning baselines.

Critical strengths
  • None
Critical weaknesses
  • The central functional-role alignment is insufficiently justified. The motivation acknowledges that brain elements can legitimately assume different roles across anatomical levels, yet HFRA explicitly minimizes these cross-level differences. Moreover, with unit node and hyperedge weights, the proposed Forman–Ricci term largely reduces to a function of hyperedge size and node degree, making its interpretation as a meaningful functional-role descriptor unclear.
  • The experimental protocol raises serious concerns about subject-level validity. APAVA and TDBrain contain only 23 and 50 evaluated subjects, leaving only a few subjects in an 8:1:1 test split, while it is unclear whether metrics are computed per subject or over thousands of segmented windows. More importantly, M-fNIRS is explicitly split at the sample level because subject identities are unavailable, which does not guarantee independence between training and test data.
Questions for authors
  • Why should functional roles be aligned across anatomical levels when the paper’s own motivation suggests that legitimate roles may differ across levels? What evidence shows that such differences represent harmful conflict rather than meaningful multi-scale information?
  • Since the main contribution is framed around multi-scale higher-order brain modeling, why are recent directly related methods such as MuHL [1], which explicitly learns multi-scale hypergraph structures, not included in the comparison? This comparison would materially affect my assessment of the empirical novelty of HHRA.

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

Reviewer_PaV5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates brain disorder diagnosis using GNNs and proposes a HHRA to address high-order interaction knowledge heterogeneity and functional role heterogeneity across different anatomical levels. Specifically, the proposed framework captures multi-level interaction information by constructing a channel-level pairwise graph, a channel-level hypergraph, and a brain-region-level hypergraph. It further introduces a cross-level functional role alignment mechanism by integrating fractional Laplacian diffusion sensitivity with Forman–Ricci curvature. Experiments on multiple EEG and fNIRS datasets demonstrate competitive performance, while the ablation studies provide preliminary evidence supporting the effectiveness of multi-level information integration.

Critical strengths
  1. The figures are visually appealing, creatively designed, and of high quality, although their structure and presentation would benefit from further clarification and refinement.

  2. This paper investigates hierarchical anatomical heterogeneity in brain disorder diagnosis by considering differences in high-order interaction knowledge and functional roles across anatomical levels. The motivation is reasonable and well-founded.

  3. The proposed method demonstrates strong performance on three EEG datasets and one fNIRS dataset, achieving the best results on seven of the eight reported evaluation metrics. Furthermore, the ablation studies provide preliminary evidence supporting the effectiveness of multi-level information integration. These results suggest that the proposed framework has empirical value for brain disorder classification, although the effectiveness of its core mechanisms requires further validation.

Critical weaknesses
  1. The research motivation of this paper is meaningful. However, the authors do not appear to clearly distinguish between multi-order interaction modeling and multi-level anatomical modeling. Specifically, HHKC constructs a channel-level pairwise graph, a channel-level hypergraph, and a brain-region-level hypergraph. The first two are based on the same set of channel nodes and differ primarily in how interactions are modeled, while the third still performs message passing over channel nodes. Therefore, it remains unclear whether these three graph structures genuinely represent distinct anatomical levels or merely reflect different graph construction strategies. Furthermore, the authors do not sufficiently justify why existing multi-level brain graph learning methods are unable to capture these interactions. This weakens the logical connection between the research motivation and the proposed methodological design.

  2. The paper argues that the same brain network element may assume different functional roles across anatomical levels and considers such differences a source of knowledge conflicts. However, HFRA directly encourages consistency by minimizing discrepancies in functional role scores across levels. In fact, differences in functional roles across anatomical levels may reflect meaningful functional specialization within brain networks rather than harmful knowledge conflicts.

  3. The paper constructs a functional role index by combining fractional Laplacian diffusion sensitivity with Forman–Ricci curvature. However, the relationship between diffusion response magnitude and actual functional roles in brain networks is not sufficiently justified. The authors should further clarify the theoretical motivation and provide empirical evidence supporting this relationship.

  4. When unit node and hyperedge weights are adopted in Eq. (16), the proposed curvature measure can be simplified to a statistic determined by hyperedge sizes and node hyperdegrees. Therefore, the authors should clarify its distinctive advantages over conventional structural statistics and conduct more fine-grained ablation studies to evaluate the individual contributions of diffusion sensitivity, curvature, and the role alignment constraint.

Questions for authors

See Weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.