Graph Embedding Using Infomax for ASD Classification and Brain Functional Difference Detection.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder with characteristic brain changes.
• Functional magnetic resonance imaging (fMRI) is used to study these changes, but high dimensionality and low signal-to-noise ratio pose challenges for robust analysis.
• Developing informative brain representations is crucial for classifying ASD and detecting functional differences.

Purpose of the Study:

• To develop a novel pipeline using Graph Neural Networks (GNNs) for analyzing whole-brain fMRI data.
• To investigate the utility of mutual information (MI) loss (Infomax) for enhancing graph embeddings in ASD research.
• To improve classification performance and identify functional differences between individuals with ASD and healthy controls (HC).

Main Methods:

• Modeled whole-brain fMRI data as a graph to preserve geometrical and temporal information.
• Employed a GNN encoder, classifier, and discriminator pipeline.
• Simultaneously optimized graph-level classification loss and Infomax (mutual information maximization) for graph embedding.

Main Results:

• Infomax graph embedding acted as a regularization term, significantly improving classification performance.
• Identified separable nodal representations between ASD and HC groups in key brain regions (prefrontal cortex, cingulate cortex, visual areas).
• The proposed pipeline facilitated more robust ASD classification models and detected functional differences, suggesting potential new biomarkers.

Conclusions:

• The GNN-based approach with Infomax significantly enhances the analysis of fMRI data for ASD.
• This method improves the robustness of ASD classification and aids in detecting functional brain differences.
• The identified separable nodal representations offer potential for discovering novel ASD biomarkers.