ComBat Harmonization: Empirical Bayes versus fully Bayes approaches.

Area of Science:

• Neuroimaging analysis
• Statistical modeling
• Biostatistics

Background:

• Neuroimaging studies require large datasets, necessitating data harmonization to mitigate biases from varying acquisition protocols and scanner differences.
• ComBat, an empirical Bayes method, is widely used for harmonizing structural neuroimaging data but can underestimate uncertainty.
• Subtle neuroanatomical variations and small effects demand robust harmonization techniques.

Purpose of the Study:

• To introduce and evaluate a fully Bayesian ComBat method for neuroimaging data harmonization.
• To compare the performance of fully Bayesian ComBat against the established empirical Bayes ComBat.
• To explore the utility of the fully Bayesian approach for data augmentation and downstream statistical analysis.

Main Methods:

• Implementation of a fully Bayesian ComBat using Monte Carlo sampling for statistical inference.
• Comparison of fully Bayesian ComBat with empirical Bayes ComBat on neuroimaging datasets.
• Assessment of harmonization accuracy, preservation of biological information, and computational efficiency.

Main Results:

• Empirical Bayes ComBat was more computationally efficient and better removed scanner-specific information.
• Fully Bayesian ComBat demonstrated superior preservation of disease and age-related biological information.
• Fully Bayesian ComBat achieved more accurate harmonization for traveling subjects and enabled data augmentation for improved classifier performance.

Conclusions:

• Fully Bayesian ComBat offers a more principled approach to neuroimaging data harmonization, preserving biological variability.
• The generative capabilities of fully Bayesian ComBat can enhance diagnostic accuracy, particularly in limited data scenarios like Alzheimer's disease detection.
• Posterior distributions from fully Bayesian ComBat facilitate robust brain-wide uncertainty assessment and advanced statistical analyses.