gms | German Medical Science

Introduction and research hypothesis: Differentiation between healthy control subjects and patients suffering from Major Depressive Disorder (MDD) based on resting-state functional MRI (rs-fMRI) data sets is performed with the aim to identify robust diagnostic markers. Due to high analytic flexibility in rs-fMRI data analysis, gold standards for preprocessing do not exist. Graph-theoretical metrics can be used to describe properties of brain networks, where nodes represent brain regions and edges represent connections in between. Because preprocessing variants impact global and nodal graph metrics, we investigate which combination of 12 preprocessing variants yields highest classification accuracies.

Material and Methods: Rs-fMRI data of 22 healthy subjects and 21 MDD patients was acquired at 3T and subjects were instructed to lie still with their eyes closed. After standard preprocessing steps using SPM8, 12 preprocessing variants were created. In the first stage the initial dataset was subjected to either linear detrending or not, in the second stage to either filtering at broad (0.01-0.08 Hz), slow-4 (0.027-0.073 Hz), or slow-5 (0.01-0.027 Hz) frequency band and in the third stage to either regression of a global signal (GMR) or not. Nuisance covariates for motion, white matter, and cerebrospinal fluid were regressed out using DPARSF [1] and volumes were parcellated into 95 regions of interest (ROI) using templates of the AAL atlas [2] with a higher level parcellation of bilateral cingulate and insular cortices. Time courses of all nodes (ROIs) were extracted and correlation coefficients were calculated for all pairs of nodes. Networks were constructed from distance penalized correlation matrices [3] and rendered sparse until the strongest edges remained with 16 sparsity thresholds between 10-40%. Local graph metrics Degree, Strength, Clustering Coefficient, Characteristic Pathlength, Local Efficiency, Participation Index, and Betweenness Centrality were calculated using Brain Connectivity Toolbox [4]. After FDR correction, nodes with a significant group difference (p=.05) for a metric were selected as features. Group classification was performed for every preprocessing variant and sparsity threshold (192 combinations) by support vector machines with both radial basis functions (RBF) and linear kernels using LibSVM toolbox [5].

Results: The number of selected features differed highly between preprocessing variants and sparsity thresholds. Compared to the RBF kernel, the linear kernel always yield lower classification accuracies. Only four methods allowed for classification in all of the tested sparsity thresholds, namely distance penalized variants without GMR and either broad or slow-4 filtered. Support vector classification of MDD patients and healthy controls can best be done with an accuracy of 98%, if the rs-fMRI time series have been solely filtered with a slow-4 frequency band and a sparsity threshold of 26% has been used to construct the graph network with 34 features mainly stemming from Characteristic Pathlength.

Discussion: It is assumed that the global mean signal contains information which is specific for subgroups, thereby capturing differences between control subjects and MDD patients. Detrending and filtering band variants seem to have a minor influence on classification accuracy, whereas the choice of kernel has an evident impact.