acceptodds
Under review as a conference paper at ICLR 2027

Revisiting Graph Neural Networks for Brain Disorders: A Neuroscience-Informed Perspective

Abstract

Brain disorders exhibit abnormal neural activity patterns and present significant challenges for accurate and objective diagnosis. Boasting a strong capacity for learning relationships, Graph Neural Networks (GNNs) are increasingly adopted to model multi-channel neural signals such as Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) for the diagnosis of brain disorders. However, existing methods suffer from two fundamental limitations from a neuroscientific perspective on brain activity analysis. First, existing methods are path-agnostic and fail to capture the directed propagation flow of neural signals and overlook essential causal mechanisms of brain activity. Second, multi-scale modeling introduces message-passing conflict that prevents effective semantic unification across hierarchical representations of brain regions and the entire brain network, and leads to unstable feature fusion. To address these challenges, we propose **CPHU**, a neuroscience-informed and path-sensitive framework for recognizing brain disorders. CPHU introduces **C**ausal **P**ath Signal Encoding (CPSE) to extract path-aware node representations via predefined and learnable propagation operators, and **H**ierarchical Brain Semantic **U**nification (HBSU) to resolve message-passing conflicts by aligning semantic features across intra-regional, global, and fused graphs. Extensive experiments demonstrate the effectiveness and superiority of CPHU in brain disorder diagnosis. The code is available at <https://anonymous.4open.science/r/CPHU-3686/>.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Electroencephalography, Brain Disorder

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GUy6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper clearly identifies two critical challenges in applying GNNs to brain disorder diagnosis from a neurological perspective. To address these challenges, the paper proposes CPHU, which is an effective automated diagnostic tool for brain disorders. CPHU introduces CPSE to extract path-aware node representation, and HBSU to resolve message-passing conflicts. Compared with existing methods, CPHU achieves superior performance on binary classification tasks across two public EEG datasets, ADFTD and APAVA, and a private fNIRS dataset, M-fNIRS.

Critical strengths
  1. CPHU uses both regional information and whole-brain information to model. Specifically, CPSE encodes signal propagation in the regional and whole-brain graphs, and HBSU can aligns semantic representations across these two scales.

  2. The HBSU chooses the most discriminative feature dimensions to align semantic consistency. This strategy focuses the consistency constraints on features that are more sensitive to the classification objective, while allowing the remaining dimensions to retain information specific to each view.

Critical weaknesses
  1. Directed signal propagation and cross-scale brain modeling have been studied before. The paper does not adequately discuss closely related methods (Shen et al., 2025; Gao et al., 2024). It also lacks comparisons with some relevant methods.

  2. CPHU proposes Causal Path Signal Encoding, but the causal paths are not clear. Neither the graph propagation matrix nor the signal flow embedding can demonstrate causal influence between brain regions.

  3. Several methodological details remain unclear. For example, the voting mechanism for topology-aware pooling strategy is unclear, and Eq.(14) appears , but in this paper, is not be defined.

  4. Evaluating CPHU on only three datasets provides limited evidence of its generalizability, and using only two evaluation metrics provides limited evidence to support its performance.

  5. No experiments directly show that the model learns neural signal pathways and resolves message-passing conflicts.

  6. The hyperparameter study is inadequate. Some parameters may affect performance, such as propagation iterations P or the smoothing hops K.

  7. This paper is not report key training settings, including the optimizer, learning rate, batch size, number of epochs, and so on.

  8. The description of the private M-fNIRS dataset is incomplete. The sampling rate, recording duration, and full preprocessing procedure are not reported.

Questions for authors
  1. Brain datasets from other modalities, such as fMRI, can also be used to study neural signal pathways. Why were only EEG and fNIRS selected? Can the proposed method be applied to fMRI datasets?

  2. Which properties of the graphs constructed using phase coherence make them suitable for modeling neural signal pathways and resolving message-passing conflicts?

  3. In Eq. (14), what value of k was used for each dataset, and how was it chosen?

  4. The paper addresses neural signal pathway modeling and cross-scale modeling.Could you add methods specifically designed for these two problems as baseline? Such as DSDirGCN-AM (Shen et al., 2025) and LAG (Gao et al., 2024).

  5. STAGIN, MDGL, and ALTER are methods proposed based on fMRI. How were these three baselines adapted to the EEG datasets?

REFERENCES Shen, K., She, Q., Yang, X., Gao, Y., and Fan, Y. Dynamic sparse directed graph convolutional network with attention mechanisms for EEG emotion recognition. Neurocomputing, 658:131749, 2025.

Gao, D., Zhang, H., Li, P., Tang, T., Liu, S., Zhou, Z., Ying, S., Zhu, Y., and Zhang, Y. A local-ascending-global learning strategy for brain-computer interface. In Proceedings of the AAAI Conference on Artificial Intelligence, 38(9):10039–10047, 2024.

Reviewer_7Q3qOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CPHU, a neuroscience-informed graph neural network framework for diagnosing brain disorders using EEG and fNIRS signals. The method addresses two challenges: conventional GNNs' inability to explicitly capture neural signal propagation paths and potential conflicts when integrating regional and whole-brain representations. CPHU introduces Causal Path Signal Encoding (CPSE), which combines predefined and learnable propagation operators to extract multi-hop path-aware node representations, and Hierarchical Brain Semantic Unification (HBSU), which uses adaptive feature fusion and semantic consistency constraints to align intra-regional, global, and fused graph representations. Experiments on two public EEG datasets and one private fNIRS dataset report competitive classification performance against several existing baselines. We present ablation studies and hyperparameter sensitivity analyses to assess the contributions of the proposed components.

Critical strengths

-The paper addresses relevant challenges in brain signal analysis by combining multi-hop propagation features with regional and whole-brain graph representations. This provides a strong motivation for developing a unified graph-learning framework.

-CPSE incorporates fixed and learnable propagation operators to enrich node representations, while HBSU introduces adaptive fusion and semantic alignment across different graph views. Together, they offer a coherent approach to integrating structural and semantic information across spatial scales.

-The proposed method achieves the best reported performance on both metrics for M-fNIRS, as well as the highest AUPRC on APAVA and ADFTD. Evaluations across EEG and fNIRS datasets, together with component ablations, provide encouraging evidence of the framework's predictive utility.

Critical weaknesses

-Although causal path modelling is presented as a central contribution, the proposed propagation operators are derived from normalised graph Laplacians and learned attention weights, which do not inherently establish causal or directional neural interactions. In particular, the fixed operator is symmetric for an undirected adjacency matrix, and the paper provides no causal discovery or directionality validation. Consequently, the claim that CPSE captures neural signal propagation mechanisms is not sufficiently supported.

  • CPSE primarily combines random feature initialisation, iterative graph propagation, and learnable attention, while HBSU employs established mechanisms such as gated fusion, gradient-based feature selection, and representation alignment. The paper does not adequately establish how these components differ fundamentally from existing multi-hop graph encodings and multi-view consistency learning approaches.

-The paper introduces message-passing conflict as a central problem but does not formally characterise or directly measure this phenomenon. The proposed Jensen-Shannon divergence constraints encourage similarity across representations, but it remains unclear whether they specifically resolve message-passing conflicts or simply act as additional regularisation. Moreover, applying divergence across all node pairs may suppress meaningful regional differences, and the validity of applying Jensen-Shannon divergence directly to learned feature embeddings needs clarification.

-CPHU does not achieve the best AUROC on APAVA, where STAGIN and ALTER outperform it, and its ADFTD improvements over MDGL are relatively small compared with the reported standard deviations. The private M-fNIRS dataset uses sample-level rather than subject-independent splitting, which can introduce subject-related leakage if multiple samples come from the same participant. Furthermore, the lack of paired statistical significance tests and limited module-specific ablations weaken the evidence for consistent superiority.

Questions for authors

-How do the proposed fixed and learnable propagation operators capture causal or directional neural signal flow rather than statistical dependencies or multi-hop graph diffusion? Can the authors provide quantitative evidence, such as comparisons with established directed connectivity measures, to support the claimed causal interpretation?

-How is message-passing conflict formally defined and measured in the proposed framework? Can the authors demonstrate that the semantic consistency constraints specifically mitigate this phenomenon rather than merely providing generic representation regularisation? Additionally, how do you compute Jensen-Shannon divergence on feature embeddings that may not represent valid probability distributions?

-For the private M-fNIRS dataset, can the authors clarify whether multiple samples originate from the same subjects and whether the sample-level splitting protocol could introduce subject-related data leakage? Can the authors also provide paired statistical comparisons against the strongest baselines to establish the significance of the reported performance improvements?

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

This paper proposes CPHU, a graph neural network framework for brain disorder diagnosis using EEG and fNIRS signals. The framework consists of two main components: Causal Path Signal Encoding (CPSE), which generates path-aware node representations using fixed and learnable propagation operators, and Hierarchical Brain Semantic Unification (HBSU), which combines regional and global graph representations through adaptive fusion and semantic alignment. The method is evaluated on three datasets against several baselines, with additional ablation studies and hyperparameter analyses.

Critical strengths
  1. The paper addresses a relevant problem in brain signal analysis.
  2. The proposed framework combines structural encoding and cross-view feature alignment in a unified architecture.
  3. The experiments cover multiple EEG/fNIRS datasets and include comparisons with several representative methods.
Critical weaknesses
  1. CPSE shares substantial similarities with Random Feature Propagation (RFP, Eliasof et al.,Graph Positional Encoding via Random Feature Propagation, ICML 2023). The paper neither discusses this closely related work nor clearly establishes the novelty of CPSE over it.
  2. The PLV matrix in Eq. (1) is symmetric and cannot identify propagation directions, while the normalized Laplacian in Eq. (6) also does not encode directionality. Although the attention matrix in Eq. (5) can be asymmetric, this does not establish physiological causality. The method supports multi-hop structural encoding but does not substantiate the claimed causal neural signal propagation.
  3. Eq. (14), as written, lacks a ground-truth label or true-class selection term. The stated loss does not depend on the ground-truth labels and does not correctly define the intended supervised classification objective.
  4. Eq. (16) aligns every pair of nodes across views rather than corresponding nodes. This objective encourages representations of different nodes to become similar, potentially conflicting with the goal of preserving region-specific information. The paper provides neither a justification for this design nor a comparison with corresponding-node alignment.
  5. The paper does not formally define or directly measure the claimed message-passing conflict. Improvements in classification performance alone do not demonstrate that this conflict is resolved. Moreover, the HBSU ablation in Section 4.3 removes the JS alignment losses rather than the entire HBSU module, which also includes adaptive fusion and pooling. The current ablations do not fully isolate the contributions of the proposed components.
  6. APAVA contains only 23 subjects, leaving approximately 2–3 subjects in each test split under the reported 8:1:1 protocol. 
  7. Contains several typos and incorrect references, such as "Section section 3.1" and referring readers to Appendix B for brain region definitions when they are provided in Appendix C.
Questions for authors
  1. What are the specific methodological differences between CPSE and RFP (Eliasof et al., ICML 2023)? Can the authors provide a direct comparison?
  2. What evidence supports interpreting the learned propagation paths as physiological causal pathways?
  3. Could the authors clarify the focal loss formulation in Eq. (14) and explain why all-pairs alignment is preferred over corresponding-node alignment in Eq. (16)?
  4. Can the authors provide more targeted ablations to verify the proposed mechanisms, particularly the individual contributions of HBSU components?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.