Joint Dynamic Brain Network Estimation and Graph Representation Learning for the Recognition of Neurological...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Graph Neural Networks (GNNs) show promise in recognizing neurological disorders by analyzing brain networks.
• Existing GNN approaches have limitations: non-learnable inputs, static interaction assumptions, and poor interpretability/robustness across disorders.

Purpose of the Study:

• To address limitations of current GNNs for neurological disorder recognition.
• To propose a novel model, EBIGNN, for inferring dynamic Effective Connectivity (dEC) in brain networks.
• To enhance interpretability and robustness in brain disorder analysis.

Main Methods:

• Developed the Effective Brain Inference Graph Neural Network (EBIGNN).
• EBIGNN infers dynamic Effective Connectivity (dEC) within an end-to-end framework.
• Model is trained with direct feedback from downstream tasks, allowing flexible learning of relevant graph structures.

Main Results:

• EBIGNN demonstrated superior performance on three public datasets compared to state-of-the-art methods.
• The model provides interpretable insights into temporal evolution and altered connectivity patterns.
• Findings align with existing neuroimaging biomarkers, indicating clinical robustness.

Conclusions:

• EBIGNN offers a flexible, interpretable, and robust approach to neurological disorder recognition.
• The model's ability to infer dynamic brain connectivity advances the understanding of brain disorders.
• EBIGNN shows significant potential for clinical applications in neuroimaging analysis.