GAT-LI

GAT-LI applies graph attention networks to classify functional brain networks and interpret feature importance to investigate connectivity differences in Autism Spectrum Disorder (ASD).

Key Features:

• Graph Attention Network Model (GAT2): GAT-LI employs GAT2, which uses graph attention layers and an attention pooling layer to learn node representations and derive graph-level representations.
• Classification Performance: GAT2 was evaluated on the ABIDE I database (1035 subjects) and on a synthetic graph dataset (4000 samples), demonstrating superior classification performance compared to comparative models.
• Interpreting Stage: GAT-LI uses GNNExplainer to elucidate feature importance within the trained GAT2 model, and comparative experiments showed GNNExplainer outperformed Saliency Map and DeepLIFT in interpretative accuracy.
• Feature Importance Identification: The interpreting stage identifies key features contributing to classification between ASD individuals and healthy controls (HC), highlighting brain connectivity differences.

Scientific Applications:

• Neuropsychiatric Disorder Analysis: Analysis of functional brain network variations in neuropsychiatric disorders such as ASD.
• Biomedical Graph Data Interpretation: Classification and interpretation of graph-structured biomedical data beyond ASD.

Methodology:

Two-stage process: a Graph Learning Stage where GAT2 learns node and graph representations, and an Interpreting Stage that employs GNNExplainer to determine feature importance.