Autism spectrum disorder diagnosis using graph attention network based on spatial-constrained sparse functional brain...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Accurate diagnosis of autism spectrum disorder (ASD) in children is clinically significant.
• Graph neural networks (GNNs) show promise for disease diagnosis using functional brain networks (FBNs).
• Challenges exist in constructing optimal FBNs from resting-state fMRI data and understanding FBN structure's impact on GNN performance.

Purpose of the Study:

• To propose a novel method, Pearson's correlation-based Spatial Constraints Representation (PSCR), for estimating FBN structures.
• To evaluate the effectiveness of the PSCR method and its influence on GNN-based ASD diagnosis.
• To compare different FBN construction methods and classification frameworks for ASD detection.

Main Methods:

• Developed the Pearson's correlation-based Spatial Constraints Representation (PSCR) method to estimate FBN structures.
• Transformed estimated FBNs into brain graphs for input into a graph attention network (GAT).
• Conducted extensive experiments on the ABIDE I dataset (n=871) comparing various FBN construction and classification approaches.

Main Results:

• The PSCR method demonstrated superiority in estimating FBN structures for ASD diagnosis.
• Different FBN construction methods significantly influenced GNN-based classification performance.
• The proposed PSCR and GAT framework achieved a promising ASD classification accuracy of 72.40%.

Conclusions:

• The PSCR method offers a robust approach for FBN estimation in ASD diagnosis.
• The study highlights the critical role of FBN structure in GNN-based disease classification.
• The PSCR-GAT framework provides a promising solution for improving patient-control separation and future ASD diagnosis.