Modeling functional connectivity changes during an auditory language task using line graph neural networks.

Area of Science:

• Neuroscience
• Machine Learning
• Graph Theory

Background:

• Functional connectivity (FC) describes correlated activation between brain regions, often modeled as graphs.
• Graph Neural Networks (GNNs) analyze these FC graphs but traditionally focus on node (region) data.
• Existing GNNs struggle to fully characterize the crucial edge attributes representing inter-regional functional correlation.

Purpose of the Study:

• To investigate the efficacy of Line GNNs in analyzing functional connectivity (FC) graphs.
• To compare the performance of Line GNNs against traditional GNNs for predicting task-associated FC changes.
• To evaluate GNN performance across two distinct neuroimaging datasets.

Main Methods:

• Implemented two GNN architectures: GraphSAGE and Graph Convolutional Network (GCN).
• Trained both traditional and Line GNN versions of these architectures.
• Utilized two datasets: Human Connectome Project (HCP) with 205 participants and a smaller dataset with 12 participants.

Main Results:

• Line GNNs demonstrated superior performance over traditional GNNs in predicting FC changes on both datasets.
• On the HCP dataset, Line GraphSAGE achieved an 18% lower mean squared error than traditional GraphSAGE (p < 0.0001).
• Line GNNs showed statistically significant improvements with minimal overfitting on the second dataset.

Conclusions:

• Line GNNs offer a promising advancement for analyzing functional connectivity in brain networks.
• The edge-centric approach of Line GNNs effectively captures complex inter-regional relationships.
• This methodology enhances the prediction of task-related brain network dynamics.