Exploring interpretable graph convolutional networks for autism spectrum disorder diagnosis.

Area of Science:

• Neuroscience
• Computational Biology
• Machine Learning

Background:

• Identifying reliable biomarkers for autism spectrum disorder (ASD) is crucial for understanding its etiology, enabling earlier diagnosis, and developing targeted treatments.
• High-dimensional brain network data presents challenges in learning effective node representations and clean graph structures.
• Jointly modeling node representation, structure learning, and graph classification is essential for robust ASD analysis.

Purpose of the Study:

• To develop an interpretable graph convolution network (GCN) model for the joint learning of node features and clean structures in brain networks.
• To enable automatic brain network classification and interpretation for autism spectrum disorder (ASD).
• To identify salient brain regions and subnetworks as potential biomarkers for ASD.

Main Methods:

• Proposed FSL-BrainNet, an end-to-end trainable and interpretable framework.
• Utilized graph convolution networks (GCNs) for joint node feature and structure learning.
• Applied the model to brain network data for classification and biomarker identification.

Main Results:

• FSL-BrainNet achieved improved prediction performance on the ABIDE dataset compared to state-of-the-art methods.
• The model successfully identified a compact set of highly suggestive biomarkers for ASD.
• Identified biomarkers included relevant brain regions and subnetworks associated with ASD.

Conclusions:

• The proposed node feature and structure learning approach enables simultaneous selection of important brain regions.
• The model effectively identifies subnetworks relevant to autism spectrum disorder (ASD).
• This framework facilitates both accurate classification and biomarker discovery for ASD.