mGNN-bw: Multi-Scale Graph Neural Network Based on Biased Random Walk Path Aggregation for ASD Diagnosis

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Autism spectrum disorder (ASD) diagnosis often relies on neuroimaging, with functional magnetic resonance imaging (fMRI) revealing altered functional connectivity.
• Existing graph neural networks (GNNs) struggle to capture long-range brain connectivity and individual variations crucial for ASD classification.

Purpose of the Study:

• To develop an advanced GNN model, mGNN-bw, that effectively captures long-range functional connectivity and individual brain network differences in ASD.
• To enhance the accuracy of computationally assisted diagnosis for ASD using neuroimaging data.

Main Methods:

• Proposed a novel multi-scale graph neural network (mGNN-bw) utilizing biased random walks and a co-optimization strategy.
• Integrated high-order brain networks (path encoding/aggregation) with low-order networks (Pearson correlation) for multi-scale feature representation.
• Employed node pooling scores to guide biased random walks for capturing long-range connectivity.

Main Results:

• Achieved superior classification performance on the ABIDE I dataset, reaching 74.8% accuracy with CC200 and 73.2% with AAL atlases.
• Outperformed existing GNN methods in differentiating individuals with ASD from typically developing controls.
• Identified key brain regions associated with ASD, including the frontal lobe, insula, cingulate, and calcarine, consistent with prior research.

Conclusions:

• The proposed mGNN-bw model offers a significant advancement in ASD diagnosis by effectively addressing limitations in capturing long-range connectivity and individual differences.
• This multi-scale approach demonstrates robust feature representation for improved classification accuracy in ASD.
• The findings support the potential of advanced computational methods for more precise clinical diagnosis of ASD.