FGLFA: A Federated Graph Learning-based Cross-Network Layer Feature Alignment Model for Major Depressive Disorder...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Centralizing medical datasets for major depressive disorder (MDD) research is hindered by privacy, security, and storage concerns.
• Federated Learning (FL) enables collaborative model training without data centralization but often struggles with data heterogeneity across different sites.

Purpose of the Study:

• To propose a Federated Graph Learning-based Cross-Network Layer Feature Alignment (FGLFA) model for improved MDD identification.
• To address the challenge of data heterogeneity in federated learning for medical datasets.

Main Methods:

• The FGLFA model utilizes Graph Sampling and Aggregation (GraphSAGE) networks trained independently at each site to extract graph-structured features.
• Residual connections (RCs) are integrated into the GraphSAGE networks to mitigate gradient vanishing and accelerate convergence.
• A feature alignment module is employed to harmonize cross-network layer features, reducing distribution discrepancies between sites.

Main Results:

• The FGLFA model achieved an average accuracy (ACC) of 65.1% and an F1-score of 70.9% across three independent sites.
• The proposed method demonstrated consistent advantages over mainstream federated paradigms, reducing variance by 23% and enhancing MDD identification accuracy.

Conclusions:

• The FGLFA model offers an effective solution for privacy-preserving, collaborative MDD identification.
• This approach provides a more efficient tool for the early diagnosis and treatment of brain diseases, overcoming data heterogeneity issues.