[Paper Review] Learning the Hierarchical Organization in Brain Network for Brain Disorder Diagnosis

Jingfeng Tang, Peng Cao+4 more|arXiv (Cornell University)|Mar 10, 2026

Functional Brain Connectivity Studies0 citations

TL;DR

BrainHO learns hierarchical, learnable subgraph representations from fMRI connectivity without fixed atlases, achieving state-of-the-art accuracy on ABIDE and REST-meta-MDD and revealing interpretable disease-related sub-networks.

ABSTRACT

Brain network analysis based on functional Magnetic Resonance Imaging (fMRI) is pivotal for diagnosing brain disorders. Existing approaches typically rely on predefined functional sub-networks to construct sub-network associations. However, we identified many cross-network interaction patterns with high Pearson correlations that this strict, prior-based organization fails to capture. To overcome this limitation, we propose the Brain Hierarchical Organization Learning (BrainHO) to learn inherently hierarchical brain network dependencies based on their intrinsic features rather than predefined sub-network labels. Specifically, we design a hierarchical attention mechanism that allows the model to aggregate nodes into a hierarchical organization, effectively capturing intricate connectivity patterns at the subgraph level. To ensure diverse, complementary, and stable organizations, we incorporate an orthogonality constraint loss, alongside a hierarchical consistency constraint strategy, to refine node-level features using high-level graph semantics. Extensive experiments on the publicly available ABIDE and REST-meta-MDD datasets demonstrate that BrainHO not only achieves state-of-the-art classification performance but also uncovers interpretable, clinically significant biomarkers by precisely localizing disease-related sub-networks.

Motivation & Objective

• Motivate the limitations of fixed atlas-based subnetwork analysis in brain disorder diagnosis.
• Introduce BrainHO to learn intrinsic hierarchical brain organization from PCC-based graphs.
• Develop a hierarchical attention mechanism with subgraph tokens and an orthogonality constraint to ensure diverse, stable subgraphs.
• Enforce hierarchical consistency to refine node features and improve disease discrimination.
• Demonstrate state-of-the-art performance and provide interpretable biomarkers on ABIDE and REST-meta-MDD.

Proposed method

• Project PCC-based brain connectivity into node tokens.
• Apply a hierarchical attention mechanism across node, subgraph, and graph levels with learnable subgraph tokens.
• Use Sparsemax attention for node-to-subgraph aggregation to promote sparsity and interpretability.
• Compute a graph-to-graph attention to obtain a global brain representation for classification.
• Impose a subgraph orthogonality loss to encourage diverse subgraphs.
• Incorporate a hierarchical consistency loss via an auxiliary node-level classifier and a teacher-student KL divergence to refine node features.

Experimental results

Research questions

• RQ1 Can a learnable hierarchical organization model capture cross-network interactions beyond predefined atlases?
• RQ2 Does enforcing orthogonality among subgraphs and hierarchical consistency improve accuracy and interpretability?
• RQ3 How does BrainHO perform compared to atlas-based and raw signal-based baselines on large ASD and MDD rs-fMRI datasets?
• RQ4 What clinically meaningful biomarkers emerge from the learned sub-networks?

Key findings

• BrainHO achieves state-of-the-art accuracy and AUC on ABIDE and REST-meta-MDD using only static PCC connectivity.
• BrainHO attains accuracy 69.68% and AUC 73.80% on ABIDE, with high sensitivity (73.11%).
• On REST-meta-MDD, BrainHO reaches 64.71% accuracy and 67.43% sensitivity, with robust generalization.
• Ablation shows removing orthogonality or hierarchical consistency damages performance, validating their contributions.
• Learned sub-networks align with known functional networks and reveal cross-network interactions, supporting interpretability and clinical relevance.