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

Abstract

It is a challenge to centralize medical datasets due to privacy, security, and storage issues for major depressive disorder (MDD). Federated Learning offers a solution for collaborative training without centralized storage. Nonetheless, it often overlooks the issue of data heterogeneity across sites. We propose a Federated Graph Learning-based Cross-Network Layer Feature Alignment (FGLFA) model for MDD identification. Specifically, it trains and tests Graph Sampling and Aggregation (GraphSAGE) networks separately for each site to extract graph-structured features from site-specific data. The GraphSAGE network integrates the residual connections (RCs), thereby mitigating gradient vanishing during model training and accelerating convergence speed. The feature alignment module aligns cross-network layer features across sites to minimize the discrepancies in feature distributions between sites. The experimental results show that the FGLFA model achieves an average ACC of 65.1% and F1-score of 70.9% across three sites. This demonstrates consistent advantages over mainstream federated paradigms, reducing variance by 23%. The proposed method improves the accuracy of MDD identification, and provides a more efficient tool for early diagnosis and treatment of brain diseases.