GAT-LI: a graph attention network based learning and interpreting method for functional brain network classification.

Area of Science:

• Neuroscience
• Computational Biology
• Artificial Intelligence

Background:

• Autism spectrum disorders (ASD) present a spectrum of symptoms affecting brain connectivity.
• Graph neural networks (GNNs) show promise for analyzing neuropsychiatric disorders but face challenges in accuracy and interpretability.
• Understanding brain network alterations in ASD is crucial for diagnosis and treatment.

Purpose of the Study:

• To develop an accurate and interpretable GNN model for classifying functional brain networks in ASD.
• To identify specific brain connectivity patterns associated with ASD.
• To compare the proposed model's performance against existing methods.

Main Methods:

• Proposed GAT-LI, a two-stage method involving graph learning and interpretation.
• Developed GAT2, a graph attention network with attention pooling for brain network classification.
• Utilized GNNExplainer for interpreting the GAT2 model and identifying feature importance.

Main Results:

• The GAT2 model achieved superior classification performance on the ABIDE I database compared to other models.
• GNNExplainer demonstrated better interpretation performance than Saliency Map and DeepLIFT.
• Identified key features contributing to the classification of ASD versus healthy controls.

Conclusions:

• GAT-LI provides an effective framework for classifying functional brain networks and interpreting GNN models in ASD research.
• The method offers a valuable tool for analyzing graph data in other biomedical applications.
• This approach enhances our understanding of brain connectivity in ASD.