Federated Brain Graph Evolution Prediction Using Decentralized Connectivity Datasets With Temporally-Varying...

Area of Science:

• Neuroscience
• Machine Learning
• Medical Informatics

Background:

• Predicting brain connectivity evolution is crucial for early disease diagnosis and clinical decisions.
• Decentralized longitudinal datasets with missing timepoints pose significant challenges for traditional machine learning.
• Federated learning (FL) offers a privacy-preserving approach for collaborative learning across multiple institutions.

Purpose of the Study:

• To develop the first federated learning framework for predicting time-dependent brain connectivity evolution from non-IID decentralized longitudinal data with missing acquisition timepoints.
• To enhance predictive performance for hospitals with incomplete data by leveraging data from other institutions without compromising privacy.
• To address the limitations of existing FL methods in handling longitudinal data with varying acquisition schedules.

Main Methods:

• Introduction of 4D-FED-GNN+, a novel longitudinal federated Graph Neural Network (GNN) framework.
• Uni-mode operation: acts as a graph self-encoder when the next timepoint is missing locally.
• Dual-mode operation: acts as both a graph generator and self-encoder when local follow-up data is available.
• Dual federation strategy: GNN layer-wise weight aggregation and pairwise GNN weight exchange.
• 4D-FED-GNN++ variant: federates based on hospital ordering derived from incomplete sequential patterns to aid poorly-conditioned hospitals.

Main Results:

• Comprehensive experiments on real longitudinal datasets demonstrate significant performance improvements.
• 4D-FED-GNN+ and 4D-FED-GNN++ substantially outperform benchmark methods in predicting brain connectivity evolution.
• The proposed methods effectively handle decentralized data with missing timepoints and varying acquisition sequences.
• Improved predictive accuracy for hospitals with challenging data conditions (e.g., consecutive or intermediate missing timepoints).

Conclusions:

• The proposed 4D-FED-GNN+ and 4D-FED-GNN++ frameworks represent a significant advancement in federated learning for neuroscience.
• These methods enable accurate prediction of brain connectivity evolution from complex, decentralized longitudinal datasets.
• The framework holds promise for advancing early disease diagnosis and personalized clinical decision-making in neurology.
• The open-source code facilitates further research and application of these privacy-preserving federated learning techniques.