Preserving specificity in federated graph learning for fMRI-based neurological disorder identification

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Resting-state functional magnetic resonance imaging (rs-fMRI) is crucial for understanding brain disorders.
• Graph neural networks (GNNs) are increasingly used for fMRI analysis but require large, centralized datasets, posing privacy and logistical challenges.
• Existing federated learning (FL) methods for fMRI often overlook site-specific demographic variations.

Purpose of the Study:

• To propose a novel specificity-aware federated graph learning (SFGL) framework for rs-fMRI analysis.
• To enable automated brain disorder identification while preserving data privacy and accounting for site-specific factors.
• To enhance the performance of federated learning models in neuroimaging research.

Main Methods:

• Developed a federated learning framework with a shared and a personalized branch at each client site.
• The shared branch utilizes a spatio-temporal attention graph isomorphism network for dynamic fMRI representation learning.
• The personalized branch integrates demographic data (age, gender, education) and functional connectivity for site-specific preservation.
• Fused representations from both branches for classification tasks.

Main Results:

• The SFGL framework demonstrated superior performance compared to state-of-the-art approaches.
• Experimental results on two fMRI datasets with 1218 subjects validated the framework's effectiveness.
• The proposed method successfully balances knowledge sharing with the preservation of site-specific characteristics.

Conclusions:

• SFGL offers an effective solution for privacy-preserving, multi-site rs-fMRI analysis.
• The framework enhances brain disorder identification by incorporating site-specific demographic information.
• This approach advances the application of federated learning in neuroimaging research.