Structural and functional covariance architecture of major depressive disorder: A meta-analytic structural equation modeling approach to primary neuroimaging analysis

<h2>Abstract</h2> Neuroimaging studies of major depressive disorder (MDD) report widespread disease-attributed abnormalities of brain structure and function. However, reports from mass univariate-driven studies are inconsistent. The objective of this study was to determine if a neuroimaging-based biomarker of MDD, which can reliably distinguish patients from healthy controls, can be generated using multivariate measures. Multivariate modeling of MDD was achieved through generation of a meta-analytic node-and-edge network model of MDD in which disease impacted brain regions (nodes) and their covariances (edges) were quantified with structural equation modeling (SEM). SEM assessment and voxel-based morphometry (VBM) analysis in primary datasets served to test our hypothesis that multivariate analyses of MDD provide improved signal over mass univariate methods. Brain areas reliably impacted by MDD (nodes) and their covariances (edges) were informed by previously published coordinate-based meta-analysis activation/anatomical likelihood estimation (CBMA-ALE) by our group. Meta-analytic model was then fit in primary structural (T1) magnetic resonance imaging (MRI) data and resting-state functional MRI (rs-fMRI) data. Primary datasets were derived from two previously recruited cohorts. Outcome measures (testing for differences between MDD and controls) from standardized SEM included: a) model goodness of fit assessment, and b) individual edge strength. SEM measures were assessed in heterogeneous MDD patient groups, and subsequently re-tested in 7 clinical subgroups of MDD patients. Meta-analytically generated MDD network model yielded 9 nodes with 6 edges among the regions. Model goodness of fit in meta-analytic datasets were good to exceptional. Model goodness of fit in regionally sampled gray matter density in primary T1 data was exceptional in clinical subgroups of MDD, poor in clinically heterogeneous subgroups of MDD, and poor in healthy control subjects. VBM analysis of the same T1 datasets yielded sparse results. Model goodness did not distinguish MDD from controls in regionally sampled primary rs-fMRI. These findings support our hypothesis of improved multivariate signal in MDD compared to findings derived from mass univariate analyses, however this effect was only detectable in T1 data (groupwise). Improved SEM goodness of fit in clinical subgroups of MDD patients supports our hypothesis of detectable neuroimaging effects of clinical heterogeneity in MDD.