TE-HI-GCN

TE-HI-GCN applies hierarchical graph convolutional networks and transfer learning to analyze sparse brain network data for improving psychiatric disorder diagnosis.

Key Features:

• Ensemble framework: Combines multiple models into an ensemble to enhance classification performance on brain network data.
• Hierarchical Graph Convolutional Networks (GCNs): Uses hierarchical GCNs to model graph-structured brain networks and capture intrinsic correlations among subjects.
• Transfer Learning: Incorporates transfer learning to improve generalization across related psychiatric disorder domains and uncover cross-disorder correlations.
• Network embedding learning for high-dimensional and noisy data: Improves network embedding learning specifically to handle high-dimensional and noisy brain network data.
• Mitigation of limited labeled data and GCN depth limitation: Addresses limited labeled training data and depth limitations of traditional GCN models to improve accuracy and interpretability.

Scientific Applications:

• Autism Spectrum Disorder (ASD) diagnosis: Demonstrated approximately 27.93% improvement in accuracy and 31.38% improvement in AUC compared to traditional GCN models.
• Alzheimer's Disease (AD) diagnosis: Achieved about 16.86% higher accuracy and a 44.50% increase in AUC over conventional methods.
• Validation and biomarker evidence: Validated through experiments on the ADNI and ABIDE databases, with discovered subnetworks supporting biomarker evidence for ASD.

Methodology:

An ensemble of hierarchical graph convolutional networks combined with transfer learning to improve network embedding learning for sparse, high-dimensional, noisy brain network data, with mechanisms to mitigate limited labeled data and traditional GCN depth limitations.