Brain network classification based on dynamic graph attention information bottleneck.

Area of Science:

• Neuroscience
• Graph Theory
• Machine Learning

Background:

• Graph neural networks (GNNs) excel at brain network classification by identifying static connections.
• Brain signals are susceptible to noise from physiological and external factors, complicating graph analysis.
• Existing methods often overlook real-time connectivity variations, relying on static topologies.

Purpose of the Study:

• To propose a novel framework, dynamic graph attention information bottleneck (DGAIB), for enhancing brain graph structure.
• To address challenges posed by noisy brain graphs and static topology assumptions.
• To improve information exchange and classification accuracy in brain network analysis.

Main Methods:

• Constructing raw brain graphs using the Spearman function.
• Applying a graph information bottleneck (GIB) to optimize connections by masking redundant embeddings.
• Utilizing a graph attention network (GAT) to enhance feature aggregation and regional information exchange.

Main Results:

• The DGAIB framework was evaluated for robustness and generalizability.
• Experiments included patient-specific EEG data (CHB-MIT) and cross-patient fMRI data (ABIDE-I).
• The framework demonstrated improved processing of dynamic brain network information.

Conclusions:

• The proposed DGAIB framework effectively enhances brain graph representations.
• It addresses limitations of noise and static structures in current brain network analysis.
• The approach shows promise for more accurate brain network classification tasks.