Toward Integrating Federated Learning With Split Learning via Spatio-Temporal Graph Framework for Brain Disease...

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Functional Magnetic Resonance Imaging (fMRI) is crucial for mapping brain functional connectivity and predicting diseases.
• Challenges in fMRI data utilization include high costs, time-consuming data collection/labeling, and privacy concerns hindering multi-site data integration.

Purpose of the Study:

• To propose a novel framework, Federated learning and Split learning Spatio-temporal Graph (F G), to overcome limitations in fMRI data analysis.
• To enable privacy-preserving integration of multi-site fMRI data for enhanced brain disease prediction.

Main Methods:

• Developed an integrated Federated learning and Split learning Spatio-temporal Graph (F G) framework.
• Split a spatio-temporal model into client temporal and server spatial models using federated and split learning.
• Incorporated a time-aware mechanism and InceptionTime model for temporal feature extraction and a united graph convolutional network for spatial integration.

Main Results:

• The F G framework effectively utilizes multi-site fMRI data without compromising privacy.
• Demonstrated reduced risk of overfitting and improved learning from limited datasets.
• Significantly boosted spatio-temporal feature extraction for fMRI data analysis.
• Outperformed state-of-the-art methods on ABIDE and ADHD200 datasets.

Conclusions:

• The proposed F G framework offers a robust solution for privacy-preserving multi-site fMRI data analysis.
• The method enhances brain disease prediction by effectively leveraging spatio-temporal graph networks.
• Identified potential biomarkers for brain disease prediction through community discovery algorithms.