Multi-site harmonization for magnetoencephalography spectral power data

A known issue with multi-site studies is the presence of site-specific effects that may confound effects of interest. These effects may be additive, multiplicative, or both. Numerous strategies have been developed and tested on microarray data from multiple batches, structural magnetic resonance imaging (MRI), diffusion tensor imaging (DTI), and functional MRI (fMRI). Multi-site magnetoencephalography (MEG) data represent a unique problem, however. The major MEG platforms differ substantially in sensor geometry, sensor layout, and noise-cancellation strategy, all of which may affect the distribution of the data. Another factor to consider in harmonization is retention of the relationship between the data and any covariates of interest. These relationships may be nonlinear, and individual sites may differ in the distribution of covariates. In this report, we test several previously developed methods for harmonization on a set of 16 open access datasets. We investigated ComBat, which uses empirical Bayes to improve model estimation; GAM-ComBat (Neuroharmonize), which extends ComBat to incorporate generalized additive modeling of the covariates of interest; CovBat (with the GAM extension), which performs a second round of ComBat harmonization to harmonize the covariance; and RELIEF, a matrix factorization technique. We found that overall, GAM-ComBat was the best choice for harmonizing the data while retaining the nonlinear dependence of the data on covariates of interest such as age. We demonstrate that harmonization of MEG data is possible and should be an integral part of any multi-site study.