acceptodds
Under review as a conference paper at ICLR 2027

HiBrain: Structure-Calibrated Higher-Order Hypergraph Learning for Brain Network Analysis

Abstract

Pairwise connectivity does not fully characterize joint multi-region dependence, while higher-order network construction alone does not establish subject-specific structural support. We propose HiBrain, a structure-calibrated multi-order hypergraph framework for brain disorder classification. HiBrain assesses third- and fourth-order cumulant increments relative to fitted lower-order models and calibrates candidate hyperedge support using individual gray-matter evidence. It integrates these relations through joint-evidence sparse attention and adjacent-order propagation, while conditional innovation admission controls additional higher-order contributions to a multi-order base prediction. On ABIDE-II, REST-MDD, and SRPBS-SCZ for autism spectrum disorder (ASD), major depressive disorder (MDD), and schizophrenia (SCZ) classification, respectively, HiBrain achieved the highest mean accuracy among 12 evaluated methods, exceeding the strongest baseline by 4.54, 6.94, and 6.47 percentage points. Within the same framework, it exceeded the strongest attention alternative in accuracy by 3.04, 3.04, and 2.99 points, respectively. External evaluation on ABIDE-I showed a 4.01-point accuracy gain over the strongest evaluated baseline. Model-derived regional importance showed greater cerebellar emphasis at higher orders in ASD and MDD, and posterior cortical and cerebellar emphasis at higher orders in SCZ. Code is available at https://anonymous.4open.science/r/HiBrain-47DF.

Primary area:
applications to neuroscience & cognitive science
Keywords:
resting-state fMRI, higher-order functional connectivity, hypergraph attention, structure-function integration, brain disorder diagnosis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_e5CwOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents HiBrain, a structure-calibrated multi-order hypergraph framework for brain disorder classification. It models second-, third-, and fourth-order functional interactions, with higher-order cumulants defined relative to lower-order dependencies and further calibrated using subject-specific gray-matter information. The resulting relations are then integrated through evidence-aware ROI–hyperedge attention and cross-order propagation, while a conditional admission mechanism selectively regulates the contribution of higher-order information to the final prediction. Experiments on ASD, MDD, and SCZ datasets, with external validation on ASD, demonstrate consistent improvements over the evaluated graph- and attention-based baselines.

Critical strengths
  • The paper presents a well-motivated approach to higher-order brain network modeling. Conditioning higher-order cumulants on lower-order relations and combining them with subject-specific structural information provides a meaningful extension beyond simply constructing hyperedges, and the extensive experiments consistently support the proposed framework.
Critical weaknesses
  • The interpretation of gray-matter information as structural support for functional hyperedges is not fully convincing. ROI-wise gray-matter measurements characterize regional morphology but do not directly establish anatomical connectivity among the ROIs forming a hyperedge. Therefore, the proposed calibration may be better understood as multimodal feature conditioning rather than explicit structural validation of higher-order relations.
  • The evaluation protocol is not strictly independent of the held-out data. The temporal block count is determined using scan lengths from the entire loaded cohort, including validation/test subjects and even the external ABIDE-I cohort, and this choice affects temporal views and candidate-selection statistics. In addition, the main results use a single model seed and do not report statistical tests against the strongest baselines.
Questions for authors
  • What evidence supports interpreting gray-matter-based calibration as structural support for a functional hyperedge, rather than as an additional morphological feature used for multimodal conditioning?
  • How does HiBrain compare with recent higher-order brain network methods such as HyperDiag [1] and MuHL [2]? Such comparisons would help clarify the empirical advantage of the proposed higher-order modeling strategy.

[1] Ma, Yulan, et al. "HyperDiag: Temporal–Regional Hypergraph Learning via Topology-Enhanced State Propagation for Brain Disease Diagnosis." AAAI (2026).

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

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

HiBrain constructs third- and fourth-order functional relations from cumulant statistics residualized against lower-order features, calibrates their support using individual gray-matter measurements, and combines them through sparse ROI–hyperedge attention and adjacent-order propagation. A conditional admission branch adds a bounded correction to the multi-order predictor. The paper evaluates ASD, MDD, and SCZ classification and includes transfer from ABIDE-II to an ABIDE-I cohort.

Critical strengths
  1. The paper distinguishes a higher-order representation from evidence that higher-order information is useful beyond lower-order features. Residualized statistics and the reference-state attention mechanism make this distinction part of the model rather than leaving it as a motivation.
  2. The evaluation covers three disorders and includes a separate ASD test cohort. The subject-level cross-fitting of the boundary predictor and the detailed account of candidate selection are useful reproducibility measures, although the scan-length exception discussed below limits the external-evaluation claim.
Critical weaknesses
  1. The external pipeline is not fully fixed using source-training data. Appendix A.2.2 states that the temporal block count depends on the shortest scan in the combined ABIDE-II and ABIDE-I collection, and that this can change reliability estimates and candidate selection. This is not diagnostic-label leakage, but it permits held-out data to influence source-side feature construction and needs to be separated from a source-frozen transfer result.
  2. The ablations do not establish the incremental benefit of fourth-order information under matched multimodal inputs. In Table 5, adding the fourth order without structural modules reduces ABIDE-II accuracy from 71.36 to 64.57, while the full model adds several GM-based paths. Without pairwise-plus-GM and second/third-order-plus-GM controls retaining the same structural branches, the improvement cannot be assigned specifically to structure-calibrated fourth-order relations.
Questions for authors
  1. Does adding the external cohort actually change the temporal block count for any reported run? Please report this explicitly and provide an external result with the block-count rule and all source-side feature construction fixed before ABIDE-I is loaded.
  2. How do pairwise-plus-GM and second/third-order-plus-GM variants perform when the structural encoders, direct GM fusion, and prediction head are retained? These are the comparisons needed to determine whether the fourth-order branch adds useful information.
  3. The higher-order token receives both a residualized cumulant and the original cumulant magnitude. Does the benefit remain with residual-only attributes, and how does it compare with original-statistic-only attributes? A lower-order surrogate-null control would further test whether the retained increments reflect more than estimation noise or limitations of the quadratic predictor.
  4. Because the base predictor already contains higher-order features, what does the additional conditional-admission branch reject in practice? Please report realized contributions before and after admission under corrupted higher-order inputs, and state whether the implementation satisfies the zero-preserving fusion condition required for exact fallback in Appendix A.6.3.
Reviewer_c91GOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

HiBrain is a multi-order hypergraph framework for classifying patients and healthy controls from resting-state fMRI and T1-derived gray-matter measurements. It represents third- and fourth-order functional relations using cumulant residuals relative to fitted lower-order models, then calibrates candidate hyperedge support using subject-specific gray-matter information. Its evidence-calibrated attention includes a zero-valued reference state that can absorb weight when candidate evidence is weak. A conditional admission branch adds a bounded correction to a base prediction that already incorporates multiple orders. The paper evaluates three internal cohorts and an external ASD cohort.

Critical strengths
  1. The attention mechanism addresses a concrete aggregation problem. A softmax over candidate hyperedges assigns the full attention mass even when every candidate is weakly supported. HiBrain’s zero-valued reference state lets the model reduce the total message sent back to a region. In the attention replacement experiments, the full attention module improves accuracy by 2.99–3.04 percentage points over the strongest alternative for each cohort. See Eqs. (3)–(4) and Table 3.

  2. The external ASD evaluation extends the evidence beyond internal cross-validation. HiBrain improves accuracy in all three internal cohorts and achieves 70.94% on the 499-participant ABIDE-I cohort, compared with 66.93% for SMGD, the strongest evaluated baseline. See Tables 2 and 4.

Critical weaknesses
  1. The ablations leave the gains from the core mechanisms unresolved. Gray-matter information enters through hyperedge calibration, structural propagation, and direct prediction fusion. Table 5 lacks a pairwise functional model with the same gray-matter encoder and direct fusion pathway. The w/o conditional admission variant also retains the extra correction branch and replaces hierarchical allocation with an unconstrained softmax. That comparison tests how the branch allocates weight across orders. The branch’s incremental benefit over the multi-order base prediction remains unmeasured.

  2. The external evaluation has an unquantified dependence on the target cohort. Appendix A.2.2 uses the shortest scan among all loaded subjects, including ABIDE-I subjects, to determine whether to use two or four temporal blocks. If including ABIDE-I changes the block count, it can also change reliability estimates and candidate selection on the source training data. The rule uses unlabeled scan lengths, and the manuscript leaves it unclear whether the block count changed in practice.

Questions for authors
  1. Under the same data splits, can you report results for a pairwise functional model with the same gray-matter encoder and direct fusion pathway and for the multi-order base model with the additional correction branch disabled? These comparisons would help isolate the contributions of higher-order functional modeling and conditional admission, respectively.

  2. Did including ABIDE-I change the temporal block count? How much of the accuracy and AUC improvement in Table 4 remains when the partitioning rule is determined exclusively from ABIDE-II training data and then frozen?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.