Graph Neural Network for Interpreting Task-fMRI Biomarkers.

Area of Science:

• Neuroscience
• Machine Learning
• Biomarker Discovery

Background:

• Autism Spectrum Disorder (ASD) diagnosis and treatment benefit from identifying reliable biomarkers.
• Graph Neural Networks (GNNs) offer a promising approach for analyzing brain network data from functional Magnetic Resonance Imaging (fMRI).
• Existing feature interpretation methods for GNNs can introduce data distribution errors by replacing feature values.

Purpose of the Study:

• To develop a robust 2-stage pipeline for identifying ASD biomarkers using GNNs.
• To enable reliable interpretation of GNN features without altering data distributions.
• To validate the pipeline's accuracy, robustness, and generalizability for biomarker discovery.

Main Methods:

• An inductive GNN was employed to embed task-fMRI brain network graphs for ASD identification.
• A novel 2-stage interpretation pipeline was developed to identify feature importance without feature replacement.
• The GNN's interpretation capabilities were compared against Random Forest, and robustness was tested using different atlases and parameters.

Main Results:

• The GNN achieved high accuracy in identifying ASD.
• The developed pipeline successfully identified brain regions and sub-graphs as evidence for ASD classification.
• The detected biomarkers showed associations with social behaviors and aligned with existing literature, suggesting potential for new discoveries.

Conclusions:

• The proposed 2-stage GNN pipeline provides a reliable method for ASD biomarker interpretation.
• The identified biomarkers are relevant to ASD pathophysiology and social behaviors.
• This approach is robust, generalizable to other graph interpretation tasks, and aids in understanding ASD.