Contrastive Brain Network Learning via Hierarchical Signed Graph Pooling Model.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Brain networks are crucial for understanding brain function, development, and disease.
• Current graph learning methods struggle with signed network data, limited sample sizes, and lack of interpretability in brain network analysis.

Purpose of the Study:

• To develop an interpretable hierarchical signed graph representation learning (HSGPL) model for brain functional networks.
• To address limitations in existing graph learning techniques for clinical phenotype prediction and biomarker discovery.
• To enhance model performance through a novel data augmentation strategy for contrastive learning.

Main Methods:

• Proposed an interpretable hierarchical signed graph representation learning (HSGPL) model.
• Implemented a new data augmentation strategy for functional brain networks to support contrastive learning.
• Evaluated the framework on classification and regression tasks using Human Connectome Project (HCP) and Open Access Series of Imaging Studies (OASIS) datasets.

Main Results:

• The HSGPL model demonstrated superior performance compared to state-of-the-art techniques across various prediction tasks.
• The proposed data augmentation strategy improved model performance in contrastive learning.
• Graph saliency maps successfully detected and interpreted phenotypic biomarkers, showcasing model interpretability.

Conclusions:

• The HSGPL model offers a powerful and interpretable approach for analyzing signed brain functional networks.
• The framework facilitates the discovery of novel biomarkers for clinical phenotypes and neurodegenerative diseases.
• The study highlights the potential of interpretable graph representation learning in advancing neuroimaging research.