Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis

Jingxi Feng; Xudong Chen; Yifan Zhang; Heming Xu; Hongcheng Han; Xijing Wang; Dong Zhang; Shaoyi Du

cs.AISep 29, 2026

Information Bottleneck-Guided Adaptive Hypergraph Transformer for Brain Disease Diagnosis

Authors: Jingxi Feng, Xudong Chen, Yifan Zhang, Heming Xu, Hongcheng Han, Xijing Wang, Dong Zhang, Shaoyi Du

Abstract

Exploring high-order correlations and long-range dependencies in brain networks holds significant value for both neuroscience research and clinical diagnosis. However, previous studies have lacked a unified integration of high-order and long-range dependency information in brain networks, and there is substantial redundancy behind various types of information. These issues limit their effectiveness in the diagnosis of brain diseases. To address this, we propose an Information Bottleneck-Guided Adaptive HyperGraph Transformer (IBAHGT). By incorporating the information bottleneck (IB) principle, this approach enables adaptive learning of high-order correlations and both short- and long-range dependencies within a unified framework for brain network analysis, achieving high-precision brain disease diagnosis. IBAHGT consists of three key components: an information bottleneck-guided adaptive hypergraph convolution, which introduces a novel hypergraph information bottleneck (HIB) principle to adaptively learn hypergraph message-passing weights between nodes and hyperedges, optimizes information flow and captures high-order information in brain networks that is maximally informative and minimally redundant (MIMR). The Transformer encoder captures global information within brain networks through the attention mechanism, specifically modeling short- and long-range dependencies. An information bottleneck-guided node-level adaptive fusion employs the IB principle to learn independent weights for each node, facilitating the fine-grained integration of high-order information and global information to obtain an efficient representation for downstream tasks. Extensive experiments demonstrate that the proposed method outperforms current state-of-the-art methods and can identify biomarkers for clinical applications.

Figures & tables

Figure 1: The brain network contains extensive high-order correlations as well as short-range and long-range dependencies. Multi-source information fusion, guided by the information bottleneck, aims to optimize the fused representation by capturing the minimal sufficient information from high-order representation $\mathbf{X}_{h}$ and global representation $\mathbf{X}_{t}$ for predicting the classification label $Y$ .

Figure 2: The overall framework of the proposed IBAHGT.

Datasets (Tasks)	ABIDE (ASD vs. NC)				ADNI (AD vs. NC)
Method	ACC (%)	SPE (%)	SEN (%)	AUC (%)	ACC (%)	SPE (%)	SEN (%)	AUC (%)
BrainNetCNN	66.2 $\pm$ 1.9	74.3 $\pm$ 4.1	56.8 $\pm$ 3.1	68.3 $\pm$ 2.6	73.2 $\pm$ 2.7	79.3 $\pm$ 4.2	60.0 $\pm$ 5.8	66.3 $\pm$ 4.7
ContrastPool	63.0 $\pm$ 2.5	71.1 $\pm$ 8.3	53.5 $\pm$ 4.6	68.1 $\pm$ 7.7	70.7 $\pm$ 4.1	72.1 $\pm$ 5.7	67.7 $\pm$ 10.2	69.5 $\pm$ 3.9
BrainGNN	64.5 $\pm$ 2.7	72.6 $\pm$ 7.7	55.0 $\pm$ 5.5	69.7 $\pm$ 7.2	73.2 $\pm$ 4.1	74.3 $\pm$ 4.2	70.8 $\pm$ 12.3	71.3 $\pm$ 6.1
BrainIB	68.3 $\pm$ 1.9	76.4 $\pm$ 5.4	58.8 $\pm$ 5.0	70.2 $\pm$ 1.9	74.1 $\pm$ 2.5	78.6 $\pm$ 5.1	64.6 $\pm$ 11.5	69.0 $\pm$ 6.5
$\mbox{HGNN}^{+}$	68.9 $\pm$ 2.1	77.0 $\pm$ 4.4	59.3 $\pm$ 5.9	71.0 $\pm$ 1.9	74.6 $\pm$ 2.9	81.4 $\pm$ 4.2	60.0 $\pm$ 5.8	69.2 $\pm$ 6.8

Table 1: Experimental results of the comparison methods.

Datasets (Tasks)	ABIDE (ASD vs. NC)				ADNI (AD vs. NC)
Method	ACC (%)	SPE (%)	SEN (%)	AUC (%)	ACC (%)	SPE (%)	SEN (%)	AUC (%)
TransEnoder	69.3 $\pm$ 0.9	77.9 $\pm$ 4.5	59.3 $\pm$ 4.2	72.7 $\pm$ 5.3	76.1 $\pm$ 2.8	77.9 $\pm$ 6.9	72.3 $\pm$ 10.4	75.6 $\pm$ 5.8
IBAHGconv	70.9 $\pm$ 1.3	81.7 $\pm$ 3.3	58.3 $\pm$ 2.9	73.4 $\pm$ 5.1	78.0 $\pm$ 2.2	82.9 $\pm$ 5.3	67.7 $\pm$ 10.2	77.3 $\pm$ 3.8
IBAHGT w/o IBAHGConv	74.1 $\pm$ 1.4	76.6 $\pm$ 4.9	71.3 $\pm$ 6.4	77.8 $\pm$ 3.4	81.9 $\pm$ 2.5	84.3 $\pm$ 3.6	76.9 $\pm$ 8.4	80.4 $\pm$ 2.7
IBAHGT w/o HIB	75.1 $\pm$ 1.5	78.1 $\pm$ 4.1	71.5 $\pm$ 6.5	79.2 $\pm$ 2.3	83.4 $\pm$ 2.4	84.3 $\pm$ 3.6	81.5 $\pm$ 7.9	81.0 $\pm$ 5.1
IBAHGT w/o IBNAFusion	73.0 $\pm$ 1.3	77.5 $\pm$ 4.6	67.8 $\pm$ 6.4	77.4 $\pm$ 3.5	81.0 $\pm$ 1.8	85.7 $\pm$ 3.9	70.8 $\pm$ 5.8	80.3 $\pm$ 3.9

Table 2: Experimental results of the ablation study.

Figure 3: Visualization of discriminative hyperedges for each classification. In each figure, the blue balls are the ROI contained in the hyperedge and the green balls are the center of the involved ROIs.

Figure 4: Visualization of important ROIs with their fusion weights (Color bar represents the magnitude of the fusion weights) and the hyperparameters influence.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: (a) The Markovian chain of information bottleneck-guided adaptive hypergraph convolution. (b) The Markovian chain of information bottleneck-guided node-level adaptive fusion.

Figure 6: Visualization of attention maps and discriminative connections. connections within the same neural system ( VN , AN , BLN , DMN , SMN , SN , MN , CCN ) are colored accordingly, while connections across different systems are colored gray.

Figure 7: The effect of varying hyperparameters. (a) and (c) are experiments on ABIDE dataset. (b) and (d) are experiments on ADNI dataset.

Method	Running Time on ABIDE(s)	Running Time on ADNI(s)
BrainNetCNN	$39.75\pm 1.03$	$8.04\pm 0.98$
ContrastPool	$334.89\pm 4.23$	$83.62\pm 2.50$
BrainGNN	$294.01\pm 1.11$	$74.97\pm 2.08$
BrainIB	$212.43\pm 1.32$	$51.66\pm 1.58$
$\mbox{HGNN}^{+}$	$36.55\pm 1.22$	$7.63\pm 3.22$
FC-HAT	$82.56\pm 0.69$	$22.76\pm 1.08$

Appendix

Table 3: Running time with different methods.

Datasets (Tasks)	ADHD-200 (ADHD vs. NC)
Method	ACC (%)	SPE (%)	SEN (%)	AUC (%)
BrainNetCNN	62.3 $\pm$ 1.7	67.6 $\pm$ 2.7	53.5 $\pm$ 1.5	61.8 $\pm$ 1.4
ContrastPool	61.9 $\pm$ 1.6	69.0 $\pm$ 4.0	50.4 $\pm$ 4.2	60.6 $\pm$ 3.5
BrainGNN	62.5 $\pm$ 1.9	67.3 $\pm$ 3.3	54.7 $\pm$ 4.5	62.3 $\pm$ 2.1
BrainIB	63.5 $\pm$ 2.0	68.7 $\pm$ 2.8	55.2 $\pm$ 5.0	63.4 $\pm$ 2.2
$\mbox{HGNN}^{+}$	64.4 $\pm$ 1.0	68.9 $\pm$ 2.5	57.2 $\pm$ 5.8	64.1 $\pm$ 1.6

Appendix

Table 4: Comparison experiments results on the ADHD-200 dataset.

Explore similar work

Learning Multi-Scale Hypergraph for High-Order Brain Connectivity Analysis
Jun 2, 2026Jaeyoon Sim, Soojin Hwang, Seunghun Baek +2Human Brain Networks Functional Connectivity
Pohang University of Science and Technology (POSTECH), Pohang, South Korea · University of North Carolina at Chapel Hill, Chapel Hill, USA
Navigating Hierarchy: Hyperbolic Learning on Brain Graphs for Disorder Diagnosis
Jul 8, 2026Yapeng Li, Bo Jiang, Ziyan Zhang +2Human Brain Networks Hyperbolic Learning
School of Computer Science and Technology, Anhui University, 111 Jiulong Road, Hefei, 230601, China
SABER: A Semantic-Aligned Brain Network Analysis Framework via Multi-scale Hypergraphs
Jul 2, 2026Yidan Xu, Xiangmin Han, Rundong Xue +1Human Brain Networks Hierarchical Semantic Alignment
Hangzhou Dianzi University, China · Tsinghua University, China · Xi’an Jiaotong University, China

Jun 2, 2026, cs.LGJ/K move · Enter open · S save

Learning Multi-Scale Hypergraph for High-Order Brain Connectivity Analysis

Jaeyoon Sim, Soojin Hwang, Seunghun Baek, Guorong Wu+1

Pohang University of Science and Technology (POSTECH), Pohang, South Korea · University of North Carolina at Chapel Hill, Chapel Hill, USA

Human Brain Networks Functional Connectivity Hypergraphs Alzheimer

Open paper AskPDF arXiv

Understanding complex interactions between brain regions is critical for early neurodegenerative disease classification such as Alzheimer's Disease (AD) and Parkinson's Disease (PD). While graph-based models are widely used to analyze brain networks, most existing approaches primarily focus on pairwise interactions between directly connected nodes, limiting their ability to capture higher-order dependencies across multiple regions. Although hypergraph-based methods have been proposed to model higher-order relations, many rely on predefined hyperedges or restrict learning to hyperedge weights, reducing flexibility and limiting their capacity to capture multi-resolution structural patterns. In this regard, we introduce an adaptive multi-scale hyperedge learning framework, i.e., MuHL, which constructs hierarchical node features and dynamically learns high-order interactions through continuous hyperedge construction over multi-resolution graph signals. Extensive experiments on multiple brain network benchmarks demonstrate that MuHL consistently improves disease classification performance across different stages, and further identifies key regions of interest (ROIs) and their group-wise interactions from the learned hyperedges that are associated with disease progression, highlighting its potential as a powerful tool for brain network analysis in neurodegenerative disorders.

Abstract

Explore similar work

Learning Multi-Scale Hypergraph for High-Order Brain Connectivity Analysis

Navigating Hierarchy: Hyperbolic Learning on Brain Graphs for Disorder Diagnosis

SABER: A Semantic-Aligned Brain Network Analysis Framework via Multi-scale Hypergraphs