Federated Bayesian network learning from multi-site data.

Area of Science:

• Neuroscience
• Machine Learning
• Biostatistics

Background:

• Identifying functional connectivity biomarkers for major depressive disorder (MDD) is crucial for understanding the disorder and enabling early intervention.
• Multi-site neuroimaging data can enhance statistical power but face challenges due to inter-site heterogeneity and data sharing restrictions.

Purpose of the Study:

• To develop a method for learning Bayesian networks from multi-site resting-state functional magnetic resonance imaging (rs-fMRI) data, overcoming heterogeneity and data sharing barriers.
• To identify functional connectivity biomarkers for major depressive disorder (MDD).

Main Methods:

• Proposes NOTEARS-PFL, a federated joint estimator incorporating shared and site-specific information using sparse group lasso penalty.
• Utilizes the alternating direction method of multipliers for optimization, enabling local data processing and central network structure updates.
• Addresses data sharing constraints inherent in multi-site studies.

Main Results:

• NOTEARS-PFL demonstrates effectiveness and accuracy on synthetic and real-world multi-site rs-fMRI datasets.
• The method shows superior efficiency and precision compared to alternative approaches in identifying brain functional connectivity.
• Validated on resting-state functional magnetic resonance imaging (rs-fMRI) data from major depressive disorder (MDD) patients.

Conclusions:

• NOTEARS-PFL is a novel toolbox for learning heterogeneous brain functional connectivity in MDD patients from multi-site data.
• The method efficiently handles data sharing constraints, crucial for collaborative multi-site research.
• Comprehensive experiments confirm the excellent efficacy of NOTEARS-PFL for MDD research.