Distributed Harmonization: Federated Clustered Batch Effect Adjustment and Generalization.

Area of Science:

• Medical data analysis
• Bioinformatics
• Machine learning in healthcare

Background:

• Independent and identically distributed (i.i.d.) data is crucial for reliable analysis.
• Multi-site medical data collection enhances diversity but introduces site-specific biases, violating i.i.d. assumptions.
• Existing harmonization methods like COMBAT struggle with newly integrated or unseen data sites, necessitating costly retraining.

Purpose of the Study:

• To develop a novel harmonization algorithm, Cluster ComBat, that effectively addresses site bias in multi-site medical data.
• To improve the usability and computational efficiency of data harmonization, particularly for scenarios involving new or unseen data sites.
• To leverage data clustering patterns for enhanced harmonization performance.

Main Methods:

• Development of the Cluster ComBat algorithm, integrating cluster analysis with existing harmonization techniques.
• Extensive simulations to evaluate the algorithm's performance under various conditions.
• Validation using real-world medical imaging data from the Alzheimer's Disease Neuroimaging Initiative (ADNI).

Main Results:

• The proposed Cluster ComBat algorithm demonstrates superior performance compared to existing methods.
• The approach effectively harmonizes site bias while preserving essential biological information.
• Cluster ComBat shows significant improvements in usability, especially for handling data from unknown or newly joined sites.

Conclusions:

• Cluster ComBat offers a more efficient and compatible solution for harmonizing multi-site medical data, particularly in dynamic data environments.
• The algorithm's ability to leverage data clusters enhances its effectiveness and reduces the need for retraining.
• This work provides a valuable tool for researchers working with distributed and diverse medical datasets.