Source-linked AI summary
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
TL;DR
Prior brain network methods have not unified high-order correlations with short- and long-range dependencies, and captured information can be redundant. IBAHGT combines information bottleneck-guided adaptive hypergraph convolution, Transformer encoding, and node-level adaptive fusion; it achieves state-of-the-art performance on two datasets.
Problem
Prior methods lack a unified integration of high-order correlations and short- and long-range dependencies, while information redundancy remains in the captured information.
Source: §1 · p. 2Method
IBAHGT combines HIB-guided adaptive hypergraph message passing, Transformer attention for short- and long-range dependencies, and information bottleneck-guided node-level fusion.
Source: §1 · p. 2Results
IBAHGT achieves state-of-the-art performance on two datasets.
Source: §5 · p. 9Takeaways & Limitations
The analysis identifies important ROIs and their fusion weights for the classification tasks.
Source: §4.4 · p. 8The study focuses on functional brain networks constructed from fMRI data; the influence of DTI-based structural networks remains to be explored.
Source: §D.1 · p. 18
Abstract
from arXiv · showhide
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.
1 Introduction
Brain networks contain high-order correlations and short- and long-range dependencies, but prior methods do not integrate these forms of information according to each ROI’s characteristics and may retain redundancy. IBAHGT combines information bottleneck-guided hypergraph learning, Transformer attention, and node-level adaptive fusion for brain disease diagnosis.
1 Introduction
- Graph methods capture pairwise dependencies but overlook collaborative interactions among multiple ROIs, whereas hypergraphs can represent high-order correlations.The introduction also notes that message-passing GNNs and HGNNs struggle with long-range dependencies, while Transformer attention can capture them.Source: §1 · p. 2
- Prior methods fail to integrate high-order correlations and long-range dependencies according to ROI-specific characteristics, while information redundancy can dilute disease-related discriminative information.The introduction argues that condensing information during high-order feature extraction and multi-source fusion is important for efficient representations.Source: §1 · p. 2
- IBAHGT unifies high-order correlations with short- and long-range dependencies, using an information bottleneck-guided adaptive hypergraph Transformer for brain disease diagnosis.The paper describes this as a unified framework for high-precision diagnosis.Source: §1 · p. 2
- HIB adaptively learns message-passing weights between nodes and hyperedges to capture maximally informative, minimally redundant high-order information.The approach applies the information bottleneck principle to optimize hypergraph message passing.Source: §1 · p. 2
- Node-level adaptive fusion learns weights for each ROI to integrate high-order and global information while reducing redundancy in the fused representation.The paper describes the resulting representations as efficient.Source: §1 · p. 2
2 Related work
Related work covers graph, hypergraph, and Transformer approaches to brain network analysis, alongside information bottleneck methods. The paper positions its approach against limitations in jointly analyzing dependency types and describes IB as target-relevant compression.
2.1 Deep Learning-Based Brain Network Analysis
- Graph methods overlook widespread high-order correlations, hypergraph methods model them, and Transformer methods capture long-range dependencies, but prior approaches do not provide a unified analysis of these dependency types.Source: §2.1 · p. 3
2.2 Information Bottleneck
- The Information Bottleneck encourages representations to retain information relevant to the target while excluding extraneous information; BrainIB applied this principle to brain network analysis.Source: §2.2 · p. 3
3 Method
IBAHGT combines adaptive hypergraph convolution, Transformer self-attention, and node-level fusion to represent high-order correlations and short- and long-range dependencies in brain networks. Information bottleneck objectives guide the extraction and fusion of task-relevant, minimally redundant information.
3 Method
- The method starts from ROI functional-connectivity features and stacks L information bottleneck-guided adaptive hypergraph Transformer layers to form final node embeddings.Each layer produces high-order and global representations before node-level adaptive fusion.Source: §3 · pp. 3–4
3.1.1 Adaptive Hypergraph Convolution
- KNN constructs a hypergraph over ROI nodes, and adaptive convolution uses weighted node-to-hyperedge and hyperedge-to-node message passing to produce high-order representations.The first stage learns node weights to form hyperedge features; the second learns hyperedge weights for transfer back to nodes.Source: §3.1.1 · p. 4
3.1.2 Optimization Based on the HIB Principle
- The HIB objective favors task-relevant hypergraph messages while penalizing redundant information, with γ1 balancing the two terms.The mutual-information terms are intractable, so the method uses a lower bound for task-relevant information and an upper bound for redundancy.Source: §3.1.2 · pp. 4–5
3.2 Transformer Encoder
- Transformer multi-head self-attention models short- and long-range dependencies between ROIs to form a global representation.This complements the high-order information captured by hypergraph convolution.Source: §3 · p. 4, §3.2 · p. 5
3.3.1 Node-Level Adaptive Fusion
- Node-level adaptive fusion learns an independent weight for each ROI to balance its high-order representation against Transformer-derived dependency information.The two representations are integrated into updated node features.Source: §3.3 · p. 5, §3.3.1 · p. 6
3.3.2 Optimization Based on the IB Principle
- The fusion objective retains task-relevant information in updated features while reducing redundancy between high-order and global representations.The trade-off hyperparameter γ2 balances these objectives.Source: §3.3.2 · p. 6
3.4 Information Bottleneck-Guided Adaptive Hypergraph Transformer Layer
- IBAHGT layers combine parallel adaptive hypergraph convolution and Transformer encoding, then fuse their representations into updated node features.After L layers, the node embeddings are concatenated and passed to an MLP for label prediction.Source: §3.4 · p. 6, §3.5 · p. 7
3.5 Loss Function
- Training combines supervised cross-entropy with separate losses for adaptive hypergraph convolution and node-level adaptive fusion.λ1 and λ2 balance the component losses.Source: §3.5 · p. 7
4 Experiments
Experiments compare IBAHGT with baselines, test its components, and examine disease-related hyperedges, ROI fusion weights, and hyperparameter sensitivity. Results favor IBAHGT on the evaluated datasets and show that both information integration and the IB-guided components matter.
4.1 Experiments Settings
- Experiments use ABIDE and an ADNI subset, with 7:1:2 train/evaluation/test splits and mean and standard deviation across 10 test runs.The ABIDE sample includes 871 subjects; the ADNI subset includes 66 patients with AD and 139 normal controls.Source: §4.1 · p. 7
4.2 Comparison Experiments
- IBAHGT significantly outperforms baselines on ABIDE (ASD vs. NC: 77.6%) and ADNI (AD vs. NC: 86.3%).Table 1 reports the comparison results.Source: §4.2 · p. 7, Table 1 · p. 8
4.3 Ablation Study
- The combined method outperforms the Transformer encoder; on ABIDE, replacing IBAHGConv with HGNN+ or AHGConv (without the HIB principle) decreases accuracy by 3.5% or 2.5%, respectively.ABIDE diagnostic accuracy also decreases with average fusion or NAFusion without the IB principle, compared with IBNAFusion.Source: §4.3 · p. 8
4.4 Interpretation Analysis
- Discriminative hyperedges show missing MFG.R and MTG.R connections in ASD, and missing PHG.R and HIP.R connections in AD.The analysis identifies significant hyperedges using t-tests.Source: §4.4 · p. 8, Figure 3 · p. 9
- Important ROIs have different fusion-weight patterns: MTG.R favors high-order correlations, while PCUN.L and HIP.R are more associated with long-range dependencies.The analysis identifies the top 10 ROIs by smallest p-values; the stated weight patterns are for ASD and AD, respectively.Source: §4.4 · pp. 8–9, Figure 4 · p. 9
4.5 Hyperparameter Sensitivity Analysis
- Classification accuracy on both datasets peaks at γ1 = 0.1 and γ2 = 0.1 as each hyperparameter increases and then decreases.The search tests each parameter at 0.0001, 0.001, 0.01, 0.1, and 1.Source: §4.5 · p. 9
5 Conclusion
IBAHGT unifies MIMR high-order correlations with short- and long-range dependencies and suppresses redundancy through adaptive convolution and node-level fusion. The conclusion reports state-of-the-art performance on two datasets, while the algorithm appendix notes a matrix-form implementation for faster computation.
5 Conclusion
- IBAHGT learns MIMR high-order correlations and short- and long-range dependencies within one framework, suppressing redundancy for efficient brain-disease representations.The conclusion reports state-of-the-art performance on two datasets.Source: §5 · p. 9
A Algorithm
- The algorithm appendix expresses adaptive hypergraph convolution and node-level adaptive fusion in matrix form to accelerate computation.Algorithm 1 takes initial node features and an incidence matrix and outputs a predicted class label.Source: A Algorithm · p. 14
34 end
- Figure 5 depicts Markovian chains for information bottleneck-guided adaptive hypergraph convolution and node-level adaptive fusion.Source: Figure 5 · p. 15
B Proof
The appendix presents proofs organized around three propositions and includes loss reformulations and bounds. The supplied proof passages also connect two propositions to Markovian chains for the model’s convolution and fusion components.
Proofs
- The proof appendix restates Propositions 1, 2, and 3; the proofs of Propositions 2 and 3 refer respectively to Markovian chains in Fig. 5(a) and Fig. 5(b).Source: B Proof · pp. 15–16
- The Proposition 1 proof passage indicates a cross-entropy loss reformulation and invokes a Nguyen, Wainright & Jordan bound.A lemma is stated for two random variables and a function of them.Source: B Proof · p. 15
- The Proposition 2 and 3 passages state claims for arbitrary distributions Q(Zl), Q(Xl_h), and Q(Xl), respectively.Source: B Proof · pp. 15–16
C.1 Analysis of Attention Maps and Discriminative Connections
Attention-map analysis highlights different within-network patterns for the two classification tasks. T-tests identify discriminative connections, with the reported ASD and AD connections differing by task.
Attention maps
- ASD classification has relatively high attention within the SMN, while AD attention scores are primarily higher within the DMN.The maps show group-level average attention.Source: §C.1 · p. 16
Discriminative connections
- For ASD vs. NC, significant discriminative connections include ACG.L–IPL.R and SFGdor.R; only connections with p-value < 0.01 are shown.Source: §C.1 · p. 16
- For AD vs. NC, a reported discriminative connection is MFG.L–SFGdor.L.Connections were identified by t-tests on attention maps.Source: §C.1 · p. 16
C.2 Hyperparameter Analysis
The hyperparameter analysis identifies k = 3 and L = 2 as the settings with the highest accuracy on both classification tasks. IBAHGT’s training time is slightly longer than that of Transformer-based methods, an increase the authors consider acceptable given the performance improvement.
C.2 Hyperparameter Analysis
- k = 3 produced the highest accuracy on both classification tasks.At k = 2 the hypergraph degenerates into a graph and loses high-order information; larger k increases aggregation and homogenizes node representations, reducing accuracy.Source: Figure 7 · p. 17, §C.2 · p. 17
C.3 Running Time
- IBAHGT training is slightly longer than Transformer-based methods, an increase the authors consider acceptable given its performance improvement.The adaptive hypergraph convolution and node-level fusion are implemented in matrix form to accelerate GPU training.Source: Table 3 · p. 17, §C.3 · p. 17
C.2 Hyperparameter Analysis
- L = 2 produced the highest accuracy on both classification tasks.Fewer layers were insufficient to capture complex dependencies, while more layers led to overfitting and declining performance.Source: §C.2 · p. 17
C.4 Baselines and Additional Comparative Experiments on ADHD-200 Dataet
On ADHD-200, IBAHGT outperforms baselines on the majority of metrics. The comparison includes CNN-, graph- or hypergraph-, and Transformer-based methods.
C.4 Baselines and Additional Comparative Experiments on ADHD-200 Dataet
- The comparison spans CNN-, graph- or hypergraph-, and Transformer-based methods.Experiments used the papers’ open-source code and their specified optimal parameters for a fair comparison.Source: §C.4 · pp. 17–18
- On ADHD-200, IBAHGT outperforms baselines on the majority of metrics.The dataset contains 357 ADHD patients and 573 normal controls from eight international imaging centers.Source: Table 4 · p. 18, §C.4 · p. 18
D.1 Limitations.
The study focuses on functional brain networks constructed from fMRI data; the influence of structural networks constructed from DTI remains for future exploration. For real-world diagnostic assistance, the authors say AI should serve only as decision support, with final decisions left to medical professionals.
D.1 Limitations.
- The study focuses on fMRI-based functional networks; structural DTI connectivity remains for future exploration.The authors propose future work integrating functional and structural brain networks to analyze ROI dependencies more comprehensively.Source: §D.1 · p. 18
D.2 Possible Societal Impacts.
- In real-world clinical settings, AI-based diagnostic methods should serve only as decision support, with final decisions made by medical professionals.The authors note that AI-assisted diagnosis may involve misdiagnoses that are difficult to avoid and can negatively affect patients and society.Source: §D.2 · pp. 18–19