MV-ComBat and MV-CovBat: Multivariate Frameworks for Joint Harmonization of Multi-Metric Neuroimaging Data

Area of Science:

• Neuroimaging analysis
• Statistical genetics
• Biostatistics

Background:

• Neuroimaging data aggregation across sites and studies is common.
• Site- and scanner-related batch effects can obscure biological variation and introduce spurious associations.
• Existing ComBat methods are primarily univariate, failing to account for dependencies across multiple metrics and features.

Purpose of the Study:

• To develop and evaluate multivariate extensions of ComBat for harmonizing complex neuroimaging data.
• To address batch effects in means, variances, and cross-metric/cross-feature covariance.
• To improve biological signal preservation and accuracy in multi-site, multi-metric neuroimaging studies.

Main Methods:

• Proposed MV-ComBat, a multivariate extension of ComBat for joint harmonization of multiple metrics.
• Implemented MV-ComBat using empirical Bayes (EB) and Bayesian Markov Chain Monte Carlo (MCMC).
• Extended CovBat to MV-CovBat for latent-space harmonization of covariance-related batch effects across features and metrics.

Main Results:

• MV-ComBat effectively reduces batch effects across multiple metrics and features.
• EB implementation is robust to measurement error; MCMC better recovers cross-metric correlations.
• MV-CovBat further separates true biological variation from batch effects when independence assumptions are violated.
• Simulations show MV-ComBat improves correlation recovery and signal preservation over univariate ComBat.

Conclusions:

• MV-ComBat and MV-CovBat offer a flexible framework for harmonizing complex, multi-metric neuroimaging data.
• These methods are crucial for accurate analysis in large-scale, multi-site neuroimaging studies.
• The proposed methods enhance the reliability and interpretability of neuroimaging findings.