Spatial-Temporal Dynamic Hypergraph Information Bottleneck for Brain Network Classification.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Graph Theory

Background:

• Graph Neural Networks (GNNs) are used for brain network classification but face challenges with noisy signals and static network structures.
• Existing GNNs often overlook high-order topological features, limiting their ability to capture complex brain dynamics.

Purpose of the Study:

• To propose an adaptive unsupervised Spatial-Temporal Dynamic Hypergraph Information Bottleneck (ST-DHIB) framework for optimizing brain networks.
• To address noise contamination and static network limitations in brain network analysis.
• To capture higher-order spatial-temporal associations in brain activity.

Main Methods:

• Utilized Graph Information Bottleneck (GIB) for graph structure purification and dynamic signal processing.
• Employed Hypergraph Neural Network (HGNN) and Bi-LSTM to model higher-order spatial-temporal dependencies.
• Developed an adaptive, unsupervised framework for dynamic brain network optimization.

Main Results:

• The ST-DHIB framework effectively purifies noisy brain network signals.
• The approach successfully captures high-order topological features and dynamic brain changes.
• Experiments demonstrated the framework's advancement and generalization capabilities on patient-specific and cross-patient data.

Conclusions:

• The proposed ST-DHIB framework offers a significant advancement in brain network classification by addressing key limitations of existing GNN methods.
• This adaptive, dynamic approach enhances the analysis of complex brain activity, paving the way for improved clinical neuro-medicine applications.