A novel approach for ASD recognition based on graph attention networks

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Early Autism Spectrum Disorder (ASD) detection is crucial for improving quality of life.
• Identifying ASD using functional connectivity (FC) from fMRI data is challenging due to data heterogeneity.
• Current deep learning methods for ASD identification lack interpretability.

Purpose of the Study:

• To propose a novel, interpretable deep learning framework for ASD recognition using graph attention networks.
• To address the heterogeneity of fMRI data across different sites.
• To improve the interpretability of deep learning models in ASD diagnosis.

Main Methods:

• Utilized graph attention networks with regions of interest (ROIs) as nodes.
• Extracted node features from BOLD signals using wavelet decomposition, mean, and variance.
• Employed self-attention mechanisms for long-range dependency capture and node-selection pooling for ROI importance.

Main Results:

• The proposed framework achieved superior performance compared to recent studies on fMRI data from children (under 12) in the Autism Brain Imaging Data Exchange datasets.
• Demonstrated high correspondence between detected ROI importance and existing research findings.
• The model provided good interpretability regarding the contribution of different brain regions to ASD prediction.

Conclusions:

• The novel graph attention network-based approach offers a promising, interpretable solution for ASD identification from fMRI data.
• The method effectively handles data heterogeneity and enhances diagnostic accuracy.
• The interpretability of the model aids in understanding the neural underpinnings of ASD.