Fed-ComBat: A Generalized Federated Framework for Batch Effect Harmonization in Collaborative Studies.

Area of Science:

• Neuroscience
• Biostatistics
• Data Science

Background:

• Multi-center studies require data harmonization to address biases and ensure interoperability.
• Current state-of-the-art harmonization methods, like ComBat, rely on random effect modeling but often necessitate data centralization, posing privacy and governance risks.
• Decentralized data analysis is crucial for large-scale studies involving sensitive patient information.

Purpose of the Study:

• To introduce Fed-ComBat, a novel federated framework for batch effect harmonization on decentralized data.
• To enable nonlinear covariate effect preservation without data centralization or prior parametric assumptions.
• To evaluate Fed-ComBat's performance against existing centralized and distributed harmonization methods.

Main Methods:

• Development of Fed-ComBat, a federated learning framework for batch effect correction.
• Implementation of random effect modeling within a decentralized data setting.
• Validation using extensive simulated data and analysis of 7 real-world neuroimaging cohorts (healthy controls, PD, AD, ASD).

Main Results:

• Fed-ComBat achieves harmonization results comparable to centralized methods in both linear and nonlinear scenarios.
• Analysis of real neuroimaging data demonstrated comparable performance between centralized and federated models for harmonizing hippocampal thickness trajectories across the lifespan.
• Federated harmonization using Fed-ComBat aligns with existing literature findings for nonlinear models.

Conclusions:

• Fed-ComBat provides an effective and privacy-preserving solution for batch effect harmonization in decentralized multi-center neuroimaging studies.
• The framework successfully handles linear and nonlinear biases without compromising data privacy or requiring data centralization.
• Fed-ComBat represents a significant advancement for collaborative research in neuroscience, particularly for studies involving vulnerable populations and sensitive data.