Harmonization of derived measures

We have developed a statistical harmonization approach [29], building upon the COMBAT method that has been successfully used for over a decade to remove batch effects in genomic studies and recently adapted in neuroimaging research [111–114]. This approach is fully multivariate, utilizing hyper-parameters to define over-arching statistical priors, and has been successfully adopted in imaging. In order to model nonlinear effects of covariates (e.g., age), we have combined COMBAT with generalized additive models (GAMs) using spline functions. The resultant COMBAT-GAM general tool for harmonization can be applied to various forms of data, including ROIs and coefficients of structural covariance and functional connectivity networks. Preliminary results of statistical harmonization from the present consortium are shown in Fig. 5.

Preliminary results from the COORDINATE MDD datasets (EMBARC, Oxford, HMRRC, Stanford, STRADL) showing age trajectories in grey matter, white matter and ventricular volumes in MDD patients (colored blue) as compared to healthy controls (colored red)

Harmonization of resting state fMRI data is typically performed at the correlation matrix level. In particular, we have developed a functional connectivity covariance batch effect correction (FC-CovBAT) [115] that models second order moments of the upper-triangular elements of individual correlation matrices derived from the fMRI data. FC-CovBAT is an extension of COMBAT and CovBat methods [116] for structural imaging data. These methods statistically model the site/scanner differences not only in the means and variances of the multivariate correlation values, but also in the covariance structures between the multivariate correlation values from the FC data (Fig. 6).

Preliminary results showing functional connectivity of DMN (between anterior cingulate cortex and posterior cingulate cortex) in healthy controls (top panel) and MDD participants (bottom panel) from the COORDINATE MDD datasets (EMBARC, Oxford, SCU (HMRRC), SNAP (Stanford), STRADL)