Preprocessing strategy influences graph‐based exploration of altered functional networks in major depression

Abstract

Resting‐state fMRI studies have gained widespread use in exploratory studies of neuropsychiatric disorders. Graph metrics derived from whole brain functional connectivity studies have been used to reveal disease‐related variations in many neuropsychiatric disorders including major depression (MDD). These techniques show promise in developing diagnostics for these often difficult to identify disorders. However, the analysis of resting‐state datasets is increasingly beset by a myriad of approaches and methods, each with underlying assumptions. Choosing the most appropriate preprocessing parameters a priori is difficult. Nevertheless, the specific methodological choice influences graph‐theoretical network topologies as well as regional metrics. The aim of this study was to systematically compare different preprocessing strategies by evaluating their influence on group differences between healthy participants (HC) and depressive patients. We thus investigated the effects of common preprocessing variants, including global mean‐signal regression (GMR), temporal filtering, detrending, and network sparsity on group differences between brain networks of HC and MDD patients measured by global and nodal graph theoretical metrics. Occurrence of group differences in global metrics was absent in the majority of tested preprocessing variants, but in local graph metrics it is sparse, variable, and highly dependent on the combination of preprocessing variant and sparsity threshold. Sparsity thresholds between 16 and 22% were shown to have the greatest potential to reveal differences between HC and MDD patients in global and local network metrics. Our study offers an overview of consequences of methodological decisions and which neurobiological characteristics of MDD they implicate, adding further caution to this rapidly growing field. Hum Brain Mapp 37:1422‐1442, 2016. © 2016 Wiley Periodicals, Inc.

Highlights

• Groups can be differentiated based on local graph metrics

• Group differences are influenced by preprocessing method

• Finding group differences depends strongly on the network density

• Conclusions regarding functional connectomic differences in major depression should be mindful of the influence of preprocessing strategies.

INTRODUCTION

Resting‐state functional magnetic resonance imaging (rs‐fMRI) measures the activity of the brain in the absence of an explicit task [Biswal et al., 1995]. Because of its simple experimental setting and brief acquisition time, rs‐fMRI holds promising potential as a clinical diagnostic tool [Cole et al., 2010]. Analysis of resting‐state time series reveals brain areas with temporally coherent fluctuations in signal amplitudes. Such correlations define functionally connected networks that have been observed in many neuroscientific studies [Fox et al., 2006; Smith et al., 2009].

In a clinical setting, rs‐fMRI measurements are becoming more common and show potential to be used to diagnose patients based on their patterns of functional connectivity [Fox and Greicius, 2010; Greicius, 2008]. Furthermore, rs‐fMRI has the potential to discriminate between disease subtypes as well as predict disease progression and treatment response [Orrù et al., 2012].

For such a scope of applications, understanding the influence of decisions during processing of datasets is crucial. For the purpose of clinical translation and evaluation, development of data mining techniques (e.g., classification or regression), which are optimized for the separation of clinical groups or detection of relevant features, will be necessary [Walter, 2013]. To provide the ground for diagnostic or predictive trials, such investigations rely on prior work about general effects of methodological choices on resulting brain features.

With the fast development and recent advances in rs‐fMRI research, it has become increasingly difficult to ignore the effect of preprocessing on outcomes of data analysis [Saad et al., 2012]. Furthermore, some of these preprocessing steps have been shown to significantly impact between‐group results [Yang et al., 2014]. However, studies so far investigated solely healthy subjects. Accordingly, the results of these studies may not be of direct relevance to clinical questions. Because of the lack of a gold standard and consensus in the field, disambiguating differences in results due to either true (the cohort) or spurious (preprocessing) effects is an important challenge.

Previous studies have investigated effects of temporal filtering, autocorrelation correction, global signal regression, head motion regression, correlation metric, order of analysis steps, test–retest reliability, utilized software packages and toolboxes on outcome of data analysis with the aim to optimize selection of analytical pipelines [Aurich et al., 2015; Braun et al., 2012; Carp, 2012a; Guo et al., 2012; Hallquist et al., 2013; Lund et al., 2005; Saad et al., 2012; Weissenbacher et al., 2009].

Decision Trees in Resting‐state fMRI Analysis

One of the current contentious issues in resting‐state analysis concerns global mean regression (GMR). The global mean signal is the sum of resting‐state fluctuations and noise (physiological and non‐physiological artifacts). Fluctuations in the global BOLD signal are assumed to have nuisance effects and are frequently regressed out. However, due to its unknown contributions on directionality of correlations, its use has become less frequent in studies investigating basic systems neuroscience questions. In contrast, studies focusing on clinical questions still frequently employ this technique. One motivation may be the absence of group differences in uncorrected datasets [Carp, 2012a], thus suggesting important influence of GMR on group effects [see e.g., Yang et al., 2014b].

Murphy et al. argued that regression of a global mean signal can introduce spurious anti‐correlations, which then influence connectivity measures [Murphy et al., 2009]. This effect was extended by the finding that GMR biases correlations differently in different regions depending on the underlying true interregional correlation structure, thereby potentially reducing the sensitivity of correlation measures. GMR can alter local and long‐range correlations, potentially spreading underlying group differences to regions where those effects did not occur. Saad et al. conclude that GMR creates rather than reveals relationships between brain regions and therefore should not be applied [Saad et al., 2012].

Nevertheless, GMR is still a contentious issue. The global signal is assumed to have non‐neuronal contributions, which may inflate inter‐subject variability and in principle reduce the significance of true group differences where they occur. Signal fluctuations caused by non‐neuronal sources may introduce correlations and cause an overestimation of FC strengths [Weissenbacher et al., 2009]. A study comparing preprocessing strategies, whilst confirming the results of Murphy et al. also conclude that GMR may be beneficial for rs‐fMRI analysis, because it greatly suppresses false correlations and increases specificity of functional connectivity results [Weissenbacher et al., 2009].

Braun et al. investigated test–retest (TRT) reliability of graph metrics of functional networks [Braun et al., 2012]. More specifically, regarding length of time series, regional activity estimation, global mean regression, frequency range, and network density. It was found that TRT reliability was moderate and highly dependent on parameters used for construction of networks and network densities. For most measures and densities, TRT reliability was not affected by GMR application. Liang et al. performed a study on the effect of correlation metric and preprocessing factor (GMR and filtering band) on resting‐state brain networks [Liang et al., 2012]. They found that global signal correction has a broad influence on global topological properties of brain networks. In contrast to the TRT reliability analysis undertaken by Braun et al., [2012], nonglobal mean regressed networks outperformed those with global signal removal. Confirmative evidence for including a reduced TRT reliability in ROI‐based correlation analyses with global signal regression was also provided by [Guo et al., 2012].

Another methodological concern is the width of the filtering band. Fluctuations in different frequency bands are considered to arise from a diversity of generative processes, each with specific properties and physiological functions [Buzsáki and Draguhn, 2004]. Liang et al. compared topological properties between slow‐4 (0.027–0.073 Hz) and slow‐5 (0.01–0.027 Hz) filtered, nonglobal mean regressed networks and found that the graph‐theoretical metrics small‐worldness and global efficiency were significantly higher in slow‐4 band than in slow‐5. However, slow‐4 band was more reliable for both global and regional network metrics [Liang et al., 2012]. Braun et al. found increased TRT reliability in broader frequency bands (0.0083–0.15 Hz) compared to a “standard” frequency band of 0.04–0.08 Hz, which is close to the slow‐4 frequency band [Braun et al., 2012].

In contrast, detrending is an issue, which has not been in the focus of methodological rs‐fMRI comparisons lately. It is often applied as a standard preprocessing step to regress out nonphysiological artifacts originating from the scanning procedure [Tanabe et al., 2002]. Because a trend caused by nonphysiological artifacts cannot be quantified during the fMRI measurement, its magnitude is unknown. A linear drift can be regressed out of the data and omission of detrending in the preprocessing pipeline reduces the data quality. Reasons brought up against its real usefulness include the improved scanner stability of up to date systems and the potential to alter original data in the case of suboptimal filters. Because its effects have not been included in the other studies in healthy populations, how the removal of a systematic drift in the time series impacts on higher order organization of brain networks or local graph metrics remains an open question.

Much of the preceding analyses have focused on the influence of preprocessing on specific changes in functional connectivity. However, the manner in which these influences propagate into system metrics that employ network theory have attracted little attention [Fornito et al., 2013]. In addition to the standard preprocessing questions, the use of graph theoretical metrics also introduces the selection of network density. During network construction, the density is set by selection of a sparsity threshold. Thresholding of correlation matrices at different levels results in different networks, which reveal different topological properties and which may obscure putative group differences [Rubinov et al., 2009]. This decision is a trade off between a sparse network (low threshold) and a highly linked one (high threshold). Densely connected networks may include weaker, potentially spurious connections and would thus converge to a featureless state, in which group differences may be obscured because inter‐subject variability is decreased [Lord et al., 2012; Rubinov et al., 2009]. On the other hand, a low sparsity may increase TRT reliability [Braun et al., 2012]. Highly interconnected networks do not show small‐worldness properties [Bassett et al., 2009; Zalesky et al., 2010]. Therefore, choosing a sparsity threshold inherently determines network properties. Threshold‐independent network assessment can be done by using an area under the curve measure, which requires recalculation of graph metrics for all of the tested thresholds. So far, studies on graph metrics rarely base their analysis on a single threshold level.

Clinical Exploitation

While most studies investigated effects of preprocessing on reliability or consistence of network metrics, comparably little research translated these insights into clinical applications. A very prominent application of rs‐fMRI is the comparison of different patient populations. Here, some effects, such as the global signal, may differ systematically and a different systemic effect of preprocessing strategies on group effects is hence quite plausible. In principle, one may assume that methods, which lead to higher consistency in healthy populations, might also support better group distinction. However, optimizing preprocessing for maximum group differentiation may offer alternative advantages, namely maximizing the sensitivity of the pipeline. This would be the case if the end goal was not the interpretation of individual results, but rather a reliable and specific accuracy in separation of two groups in a classification framework.

The objective of the present study is to characterize the influence of preprocessing methods on graph theoretical metrics of resting‐state brain networks in healthy controls and patients suffering from major depressive disorder (MDD). MDD is an affective disorder, which can be described as a disorder of stress response and adaptation. It is a common, recurrent, frequently severe, disorder linked to a diminished quality of life, as well as increased medical morbidity, and mortality rates [Ferrari et al., 2013]. Its unresolved diagnostic and therapeutic challenges concur to one of the most burdening illnesses for both the patients and health care systems. Identification of subgroups, which differ, for example, in their treatment response, risk for side effects, or disease course is thus crucial. Importantly promising rs‐fMRI‐based biomarkers have been reported for this disease lately [Sundermann et al., 2014], suggesting clinical exploitation during diagnosis [Lord et al., 2012] as well as for prediction of treatment efficacy [McGrath et al., 2013] or side effects [Metzger et al., 2013].

We focus on two of the most commonly discussed preprocessing issues—global mean regression and temporal filtering—for a systematic investigation. They are studied in combination with application of detrending. This choice allows for comparison between detrended versus nondetrended, broad versus relatively narrow filter bands, and global mean regressed versus non‐global mean regressed preprocessing choices, yielding twelve preprocessing variants in total. The effect of network sparsity is also investigated. Of note, exploration regarding which preprocessing pipeline results in the highest count of between‐group differences will not be a question in this study, because this can only be seen as a synonym for an optimal preprocessing strategy, if the goal would be to deploy techniques of data mining to develop a biomarker.

METHODS

Subject Cohort

We reanalyzed data described in Lord et al., because of the high diagnostic accuracy of over 90% using support vector classification [Lord et al., 2012]. This data set consists of 21 medicated patients suffering from an acute MDD episode (mean age: 37.9 ± 11.4, HAM‐D: 15.8 ± 4.8) and 22 healthy control subjects (mean age: 34.6 ± 6.2, HAM‐D: 0), matched with respect to age and gender. Exclusion criteria were major medical illness, history of seizures, medication with glutamate modulating drugs or benzodiazepines, prior electro‐convulsive therapy (ECT) treatments, pregnancy, atypical forms of depression, any additional psychiatric disorder, history of substance abuse or dependence, as well as contraindications against MRI.

Data Acquisition

The functional Magnetic Resonance Imaging data were acquired in Magdeburg, Germany, on a 3 Tesla Siemens MAGNETOM Trio scanner (Siemens, Erlangen, Germany) with an eight‐channel phased‐array head coil. Subjects were instructed to lie still, with eyes closed, and to let their mind wander without explicitly engaging in any specific cognitive task. Motion was minimized using soft pads fitted over the ears and participants were given earplugs to minimize noise. During 10 min of resting‐state scan, the scanner room was not illuminated. Functional time series of 488 time points were acquired with an echo‐planar imaging sequence. The following acquisition parameters were used: echo time = 25 ms, field of view = 22 cm, acquisition matrix = 44 × 44, isometric voxel size = 5 × 5 × 5 mm³. Twenty‐six contiguous axial slices covered the entire brain with a repetition time of 1250 ms (flip angle = 70°). The first 13 acquisitions were discarded to reach steady state and limit T1 effects.

Data Preprocessing

To check for motion unrelated spikes, data quality in terms of the two quality parameters percent signal change (PSC) and a correlation coefficient (r _qq), quantifying the difference of the sample distribution to the normal distribution, was measured using the aqua tool [Stöcker et al., 2005]. R _qq is sensitive to changes near the mean of the estimated probability distribution of the data and detects spike artifacts spreading over the whole volume well. An r _qq = 1 reflects optimal data quality, while a temporally localized PSC exceeding 5% cannot be entirely explained by the BOLD effect and as such is likely an artifact.

Because the dataset has an average PSC of 2.2% (min = 1.7, max = 3.5) and an average r _qq of 0.92 (min = 0.75, max = 0.96), spikes in the dataset could be excluded.

The time series were preprocessed using spm8 as executed in the DPARSF toolbox version 2.0 [Group, 2009; Song et al., 2011; Yan and Zang, 2010]. The initial functional time series were slice‐time acquisition corrected, realigned, and normalized to the EPI template. Twelve preprocessing variants were tested, including all possible combinations of detrending (yes or no), filtering (broad (0.01–0.08 Hz), slow‐4 (0.027–0.073 Hz), slow‐5 (0.01–0.027 Hz) frequency band), and global mean regression (yes or no) (Fig. 1). Time series from cerebrospinal fluid, white matter as well as six‐parameter rigid body motion regressors were regressed out of every dataset. The resulting volumes have 475 timepoints and were parcellated into 95 regions of interest (ROIs) using a modified version of the AAL atlas containing a higher level of parcellation for the cingulate cortex and insular cortex with bilateral regions stretching across the midline of the brain (Table AI) [Tzourio‐Mazoyer et al., 2002]. Potential between‐group differences in head motion were excluded (HC = 0.14, MDD =0.17, P = 0.069) by comparing voxel‐specific mean frame‐wise displacement indices (FD) between groups, as introduced by [Power et al., 2012] and implemented in DPARSF. Smoothing has not been performed, as it would introduce spurious high correlations between an index node (region) and its immediate neighbors [Fornito et al., 2010].

Figure 1.

Course of applied preprocessing steps. The initial dataset has been slice‐time corrected, realigned and normalized. The 12 preprocessing variants differ in application of detrending, filtering and global mean removal, but all variants had nuisance covariates removed. In the naming scheme, “D” and “nD” indicate detrending and no detrending, respectively. The second index refers to the three filtering bands (b = broad, 4 = slow 4 and 5 = slow 5) and the third index reflects choice (“G”) or omission of global mean regression (“nG”).

Time Course Extraction

Using DPARSF toolbox, time courses of 95 predefined ROIs were extracted from every of the 12 datasets. Correlation matrices with pair‐wise Pearson correlation coefficients were created.

Comparison of Raw Correlation Matrices

Jaccard coefficients were used to measure the similarity of raw correlation matrices between groups after preprocessing and before graph networks were constructed. Hence, raw correlation matrices were averaged per group, reduced to the upper right triangles excluding the main diagonal, thresholded, and binarized. Jaccard coefficients measure similarity of two sets A and B by quantifying the number of samples that are shared between the two sets according to the formula: $jaccard\left(A,B\right)=\frac{\left|A \cup B\right|}{\left| A \cap B\right|}$. The Jaccard coefficient ranges from [0,1], where 0 and 1 indicate disjoint and identical sets, respectively [Jaccard, 1912]. The Jaccard coefficient is proportional to the similarity of the compared variants or groups. Four binarization thresholds were tested, 0.1, 0.3, 0.5, and 0.7. If the binarization threshold is for example 0.7, only the correlation coefficients >0.7 will be compared. Because the location of correlation coefficients in the connectivity matrix is not arbitrary but is determined by the fixed order of nodes, a location‐restrictive analysis of the Jaccard measure was implemented. A shared sample of the correlation matrices represents a correlation coefficient (or edge), which survived the binarization threshold and occurs in both averaged group connectivity matrices at the same location. The Jaccard coefficient was computed for every combination of binarization threshold and preprocessing variant.

To also assess the effect of preprocessing on similarity of raw correlation matrices of healthy subjects only, Jaccard coefficients were computed for every combination of preprocessing variant per subject and afterward averaged over subjects to yield a more realistic estimate.

Network Construction and Computation of Global and Local Graph Metrics

Based on correlation matrices, graph networks with 95 nodes were constructed on single subject level. The graphs were rendered sparse by recursively removing edges, beginning with the weakest weights and progressing until a certain percentage of edges remained. To investigate the influence of different sparsity threshold levels on network properties, 16 sparsity thresholds were tested, starting from 10% in increasing steps of 2–40%. Where removal of an edge resulted in a disconnected graph, the respective edge was retained, even in case of low weight. From these weighted, undirected graphs, network metrics were derived using functions from the BCT toolbox [Rubinov and Sporns, 2010]. The seven global metrics: normalized clustering coefficient (cc_norm), clustering coefficient of the networks (cc_real), normalized characteristic path length (cpl), characteristic path length of the networks (cpl_real), small‐worldness index (SWI), assortativity (alpha), and global efficiency (E_glob), as well as seven local graph metrics degree, strength, average path length, betweenness centrality (BCI), participation index (PI), local efficiency (E_loc), and normalized local efficiency (LEGE) were computed. To control for putative differences in overall connection strength, each network was normalized to reference random networks obtained through a resampling algorithm [Maslov and Sneppen, 2002], with the effect that the metrics clustering coefficient and characteristic pathlength could be normalized to the reference networks [Lord et al., 2012; Rubinov and Sporns, 2010].

Statistical Analysis of Graph Metrics

For between‐group comparison of global graph metrics, two‐sample t tests with a significance level of α = 0.05 were performed using MATLAB. These t tests were calculated between groups for all 12 combinations of preprocessing variants, 16 sparsity thresholds, and the 5 global metrics alpha, cc, cpl, SWI and E_glob. Two sample t tests using the seven local graph metrics degree, strength, E_loc, PI, bci, LEGE, and average pathlength, were corrected for multiple comparisons regarding the number of nodes with FDR (P value < 0.05) using the multi‐test package of the R toolbox [R Core Team, 2012].

Robustness of Group Effects

The robustness of graph results was assessed by consistency of group differences in local graph metrics over sparsity thresholds. Afterwards, the robustness between methods was explored by measuring their similarity. This part of the analysis was performed in 5 steps Fig. AI

1. Binary matrices for every preprocessing variant and sparsity threshold were created. These binary matrices are of size 95 × 7 (# ROIs x # local graph metrics). They are termed “binary‐threshold‐matrices” and contain 1, if the respective pair of ROI and local graph metric survived FDR correction at a significance level of P = 0.05, i.e. if that pair shows a significant group difference between MDD patients and healthy controls. Otherwise, it contains 0. These binary matrices represent a convenient summary of group comparisons in the means of local graph metrics.

2. Weighted matrices were computed for every preprocessing variant. These matrices are termed “area‐under‐the‐curve‐matrices,” because they summarize results across all sparsity thresholds. The weighted matrices represent added binarized matrices for all sparsity thresholds and hence contain values in the range of [0,16].

3. The robustness was determined by again creating binary matrices for every preprocessing variant. These “robustness‐matrices” contain 1 if the respective tuple consisting of ROI and graph metric is considered to be robust over sparsity thresholds; otherwise they contain 0. Here, a tuple is defined to be robust, if the group difference is either significant or insignificant in a strong majority of cases. As there is no consensus in the literature regarding the value of this majority criterion, it was set a priori at 87.5% (14 of 16 cases).

4. The similarity of preprocessing variants in terms of their robustness to delineate control subjects and patients was assessed by comparing “robustness‐matrices” using Jaccard coefficients.

5. Jaccard coefficients represent similarity between preprocessing variants, thus a dissimilarity matrix containing values calculated as 1‐Jaccard, was created. The similarity of preprocessing variants was visualized via hierarchical clustering in a dendrogram, where the length of each U‐shaped line represents the Euclidean distance between two connected samples based on the dissimilarity matrix. The linkage order has been determined using the aforementioned matrix by internal MATLAB functions.

Correlation of Disease Severity to Local Metrics

In a post‐hoc analysis, correlations between disease severity and nodal graph metrics were investigated. The HAMD score of the MDD patients was correlated with their BCI in the left hippocampus and the left caudate nucleus for the 12 preprocessing variants and 16 network sparsities.

RESULTS

Effect of Preprocessing on Group Differences in Raw Correlation Matrices

To assess the effect of preprocessing variant on group discrimination, the similarity of averaged correlation maps of healthy controls and MDD patients for each preprocessing variant was quantified using Jaccard coefficients. Table 1 shows Jaccard coefficients in the range of 0.56–0.99, reflecting an overall high similarity of correlation matrices for all tested binarization thresholds between groups.

Table 1.

Effect of preprocessing on raw correlation matrices

Preprocessing variant	t = 0.1	t = 0.3	t = 0.5	t = 0.7
D_b_G	0.78	0.73	0.73	0.76
D_b_nG	0.99	0.90	0.72	0.67
D_4_G	0.78	0.73	0.73	0.76
D_4_nG	0.99	0.90	0.72	0.67
D_5_G	0.74	0.72	0.64	0.71
D_5_nG	0.96	0.82	0.68	0.68
nD_b_G	0.77	0.75	0.71	0.74
nD_b_nG	0.99	0.90	0.72	0.68
nD_4_G	0.77	0.71	0.73	0.74
nD_4_nG	0.99	0.88	0.56	0.58
nD_5_G	0.74	0.70	0.64	0.72
nD_5_nG	0.96	0.81	0.67	0.68

Similarity between averaged correlation matrices of healthy controls and MDD patients were quantified using Jaccard coefficients, which are listed for all 12 preprocessing variants. For abbreviations of preprocessing variants please refer to Figure 1. A Jaccard coefficient close to 1 represents a high similarity between groups. The Jaccard coefficients were calculated on binarized correlation matrices and the four used binarization thresholds t are listed. A t of 0.7 indicates that correlation coefficients >0.7 were compared.

Variants with GMR show a stable similarity (Jaccard = 0.64–0.78) over binarization thresholds. Without GMR, similarity varies over binarization thresholds (Jaccard = 0.56–0.99) and similarity between groups anticorrelates with binarization threshold uncovering high between‐group variance in the higher correlation coefficients. Variant nD_4_nG yielded the lowest Jaccard coefficients, reflecting the greatest dissimilarity between groups.

When comparing raw correlation maps regarding effect of preprocessing choice from healthy subjects only, it was consistently found that similarity of correlation maps anticorrelates with binarization threshold.

The effect of detrended versus non‐detrended processing on raw connectivity matrices was minimal (Jaccard = 0.54–0.93, Table 2). Likewise, when switching the choice of temporal filter between broad, slow‐4, and slow‐5, connectivity matrices were also similar (Jaccard = 0.34–1, Table 3). In contrast, global mean regression severely affected similarity of raw correlation matrices (Jaccard = 0.14–0.47, Table 4).

Table 2.

Effect of detrending on similarity of raw correlation matrices of healthy subjects

Comparison detrending	t = 0.1	t = 0.3	t = 0.5	t = 0.7
D_b_G versus nD_b_G	0.67	0.63	0.63	0.61
D_b_nG versus nD_b_nG	0.93	0.82	0.66	0.54
D_4_G versus nD_4_G	0.70	0.61	0.57	0.55
D_4_nG versus nD_4_nG	0.87	0.77	0.67	0.57
D_5_G versus nD_5_G	0.69	0.61	0.56	0.56
D_5_nG versus nD_5_nG	0.86	0.77	0.67	0.57
Mean	0.79	0.70	0.63	0.57

Six comparisons of preprocessing variants and four binarization thresholds t are listed. For abbreviations of preprocessing variants please refer to Figure 1. A Jaccard coefficient close to 1 reflects high similarity between the compared preprocessing variants.

Table 3.

Effect of filtering on similarity of raw correlation matrices of healthy subjects

Comparison filtering	t = 0.1	t = 0.3	t = 0.5	t = 0.7
Slow4 versus slow5
D_4_G versus D_5_G	0.89	0.88	0.88	0.89
D_4_nG versus D_5_nG	0.98	0.94	0.90	0.87
nD_4_G versus nD_5_G	0.67	0.63	0.62	0.61
nD_4_nG versus nD_5_nG	0.93	0.82	0.67	0.55
Mean	0.87	0.82	0.77	0.73
Broad versus slow4
D_b_G versus D_4_G	0.85	0.81	0.79	0.80
D_b_nG versus D_4_nG	0.95	0.91	0.86	0.83
nD_b_G versus nD_4_G	0.70	0.61	0.57	0.55
nD_b_nG versus nD_4_nG	0.87	0.77	0.67	0.57
Mean	0.84	0.78	0.72	0.69
Broad versus slow5
D_b_G versus D_5_G	0.49	0.41	0.38	0.37
D_b_nG versus D_5_nG	0.81	0.64	0.47	0.34
nD_b_G versus nD_5_G	1.00	1.00	1.00	1.00
nD_b_nG versus nD_5_nG	1.00	1.00	1.00	1.00
Mean	0.82	0.76	0.71	0.68

Twelve comparisons between preprocessing variants and four binarization thresholds t are listed. For abbreviations of preprocessing variants please refer to Figure 1. A Jaccard coefficient close to 1 reflects high similarity between the compared preprocessing variants.

Table 4.

Effect of global mean regression on similarity of raw correlation matrices of healthy subjects

Comparison GMR	t = 0.1	t = 0.3	t = 0.5	t = 0.7
D_b_G versus D_b_nG	0.37	0.21	0.15	0.17
D_4_G versus D_4_nG	0.37	0.21	0.15	0.17
D_5_G versus D_5_nG	0.46	0.32	0.25	0.25
nD_b_G versus nD_b_nG	0.37	0.20	0.14	0.17
nD_4_G versus nd_4_nG	0.36	0.19	0.14	0.16
nG_5_G versus nD_5_nG	0.47	0.33	0.25	0.25
Mean	0.40	0.24	0.18	0.19

Six comparisons of preprocessing variants and four binarization thresholds t are listed. For abbreviations of preprocessing variants please refer to Figure 1. A Jaccard coefficient close to 1 reflects high similarity between the compared preprocessing variants

Effect of Preprocessing on Group Differences in Whole Brain Topological Network Properties

Group differences based on topological network properties were only found in 0.5% (5 of the 960 combinations of 5 global graph metrics, 12 preprocessing variants, and 16 sparsity thresholds) of the tested comparisons (Table 5).

Table 5.

Significant differences between healthy controls and MDD patients in topological network properties (global graph metrics) at a significance level of P = 0.05 were found in the nD_4_G and nD_4_nG preprocessing variants

Preprocessing variant	Sparsity	Graph metric	P value
nD_4_G	10%	g_cc_real	0.046
nD_4_nG	16%	g_cc_real	0.035
nD_4_nG	18%	g_cc_real	0.032
nD_4_nG	20%	g_cc_real	0.038
nD_4_nG	22%	g_cc_real	0.043

The global clustering coefficient of the real network (g_cc_real) measures segregation and reflects the average diffusiveness of clustered connectivity around individual nodes.

Considering 960 independent tests, one would expect 48 results with a P value <0.05 (960 × 0.05), much more than observed.

MDD patients have a reduced global clustering coefficient (cc_real), when the data were slow‐4 filtered (variant nD_4_nG) and at sparsity thresholds 16–22%, as well as following GMR (variant nD_4_G) at a sparsity threshold of 10% (Fig. 2). Effect sizes of the group comparison MDD patients versus healthy controls in clustering coefficient are provided in Fig. A2 and Table AII.

Figure 2.

Group differences in the global clustering coefficient cc_real. Upper and lower plots depict preprocessing variants nD_4_G and nD_4_nG, respectively. Gray circles and black diamonds represent data from healthy controls and MDD patients, respectively. Error bars represent standard error of the mean and black stars indicate group differences at a significance level of P < 0.05.

Effect of Preprocessing on Group Differences in Nodal Graph Metrics

Of note, a group difference in local metrics is always a tuple consisting of a local graph metric and a region. A full list of significant group differences using all preprocessing strategies and thresholds is provided in the appendix in Table AII. Overall, the occurrence of group differences is sparse, variable, and highly dependent on the combination of preprocessing variant and sparsity threshold.

For every combination of tested sparsity threshold and preprocessing variant, the respective number of significant results varied between 1 and 6 (Table 6). Across the whole range of sparsity thresholds, preprocessing variant nD_4_nG was associated with most group differences, with a total of 30 (Table 6).

Table 6.

Overview of occurrence of significant differences (P < 0.05, FDR corrected) between healthy controls and MDD patients in local graph metrics

Sparsity thresholds (%)	Preprocessing variant												Σ
	D_b_G	D_b_nG	D_4_G	D_4_nG	D_5_G	D_5_nG	nD_b_G	nD_b_nG	nD_4_G	nD_4_nG	nD_5_G	nD_5_nG
10	2	0	2	0	0	0	0	1	2	1	1	0	9
12	2	0	2	0	3	0	0	0	0	3	0	0	10
14	3	0	3	0	2	0	1	0	0	2	0	0	11
16	2	0	2	0	0	0	0	0	0	3	1	4	12
18	4	1	4	1	1	0	0	1	0	4	1	2	19
20	1	0	1	0	0	0	1	0	0	3	0	2	8
22	2	0	2	0	0	1	0	0	0	4	1	2	12
24	0	0	0	0	0	0	0	0	0	2	1	0	3
26	0	6	0	6	0	0	0	0	0	2	1	0	15
28	1	0	1	0	0	0	0	0	0	4	1	0	7
30	0	1	0	1	0	0	0	0	0	0	1	0	3
32	0	0	0	0	0	1	0	0	0	2	1	0	4
34	0	0	0	0	0	3	0	0	0	0	2	0	5
36	0	0	0	0	0	0	0	0	0	0	2	0	2
38	0	5	0	5	0	0	0	0	0	0	2	0	12
40	0	0	0	0	0	0	0	1	0	0	2	0	3
Σ	17	13	17	13	6	5	2	3	2	30	17	10	135

Results for each combination of tested sparsity threshold and preprocessing variant are listed.

Across all preprocessing variants, a sparsity threshold of 18% was associated with most between‐group differences (Table 6).

Regarding localization of group differences, significant group differences occurred on only a relatively modest number (25 of 95) of ROIs (Table AII).

Conversely, every of the tested local metrics revealed group differences for at least one preprocessing variant and ROI (Table AII).

Because more than one local metric can indicate a group difference in an ROI, the distribution of the local metrics indicating group differences was analyzed.

The number of possible combinations of local metrics (binomial coefficient “n choose k,” where k = 7 and n is the range 1 to) showing a group difference in any region is 127.

Of these 127 possible combinations, only 13 combinations of local metrics were found (Fig. 3).

Figure 3.

Overview of graph metrics depicting a significant (P < 0.05, FDR corrected) difference between healthy controls and MDD patients in local graph metrics for each combination of tested sparsity threshold and preprocessing variant. The color code represents combination of local graph metrics showing a group difference.

To unravel this composition even more, the frequency of single or multiple metrics indicating group difference was further analyzed.

In 54% of all regions showing group differences, only one local metric indicated the group difference (Fig. 4), showing that different local metrics capture unique aspects of the group difference.

Figure 4.

Distribution of combinations (including percentages) of local graph metrics indicating a significant difference between healthy controls and MDD patients.

A direct comparison of global mean regressed versus nonglobal mean regressed preprocessing variants revealed considerable alterations in occurrence of group differences regarding affected ROIs (Fig. 5). Notably, however, the right thalamus, bilateral dorsal anterior cingulated cortex (dACC), and left inferior parietal gyrus revealed group differences independent of (i.e., with and without) global mean regression. The number of ROIs revealing group differences when GMR has been regressed and without GMR is 11 and 17, respectively.

Figure 5.

Regions with significant differences (P < 0.05, FDR corrected) in local graph metrics between healthy controls and MDD patients are depicted with respect to application (G, red bars) or omission of global mean regression (nG, black bars) in the 12 preprocessing variants. The count of occurrence of a group difference in the ROI is depicted on the x axis. Only three of 25 regions show group differences in both G and nG preprocessing variants.

Robustness of Group Differences Over Sparsity Thresholds

To assess robustness of group differences in nodal metrics over sparsity thresholds, hierarchical clustering was used. The resultant dendrogram shows tight clustering of preprocessing variants with GMR, which were framed by variants without GMR (Fig. 6). When considering the group differences, Variants D_b_G and D_4_G were the most similar, followed by D_b_nG and then D_4_nG. The arrangement gives evidence that local graph metrics have a high similarity in terms of their robustness of group differences over sparsity thresholds, if they only differ in one aspect of the preprocessing and additionally revealed that preprocessing variant nD_4_nG exhibited lowest similarity to the rest of the variants.

Figure 6.

Hierarchical clustering of robustness measures based on group differences in local graph metrics. A robust group difference (local graph metric and ROI) was defined to occur over the majority of tested sparsity thresholds. Similar preprocessing variants are located nearby in the hierarchical clustering plot.

Correlation of Disease Severity to Local Metrics

As a post‐hoc analysis, correlations of disease severity measured by the HAMD score and the nodal graph metric BCI in both the left hippocampus and the left caudate nucleus for the 12 preprocessing variants and 16 network sparsities. Evidence for significant negative correlations of the left caudate nucleus and HAMD scores was found for four preprocessing variants across most of the tested sparsity thresholds (Fig. 7).

Figure 7.

Correlation between local graph metrics and disease severity. Pearson correlation coefficients between local betweenness centrality (BCI) in the left caudate nucleus and HAMD score in MDD patients are color‐coded and listed for every preprocessing variant and sparsity threshold. Significant correlations (P < 0.05) are marked with a black cross.

DISCUSSION

In this work, we systematically investigated the effects of common preprocessing variants, including global mean‐signal regression (GMR), filtering, detrending and network density on graph theoretical metrics describing global and nodal properties of brain networks. Our aim was to systematically compare these preprocessing strategies by evaluating their influence on group differences between HC and MDD, and to propose recommendations for research using graph metrics to study major depression in rs‐fMRI.

Effect of Preprocessing on Group Differences in Raw Correlation Matrices

Our first objective was to assess the influence of preprocessing variants on group discrimination based on raw correlation matrices. Correlation maps of healthy control subjects and MDD patients tested for every preprocessing variant separately were found to be highly similar between groups.

Prior studies also used Jaccard coefficients to measure similarity of matrices or maps. For example, Pedoia et al. used different preprocessing toolboxes and compared the similarity of the resulting images with the Jaccard index [Pedoia et al., 2013]. To estimate similarity of activation patterns across experimental tasks between pairs of brain regions, Crossley et al. used the Jaccard index [Crossley et al., 2013].

Effect of Preprocessing on Group Differences in Whole Brain Topological Network Properties

Group differences in global graph metrics were absent in the majority of tested preprocessing variants, making it difficult to discuss the nuances of preprocessing steps on whole brain group differences. These sparse findings highlight the sensitivity of different preprocessing variants to different effects in the data.

Interestingly, MDD patients had a significantly decreased global clustering coefficient, which was found in the nondetrended slow‐4 filtered datasets (preprocessing variants nD_4_nG and nD_5_G).

Not performing detrending seems to have a pivotal effect on revealing group differences in topological networks, although there is no hypothesis on the neuronal causes of linear trends and detrending is currently done in most analysis.

An imperfect separation of the high pass filter (detrending) and the second bandpass filter combined with the possibility that nuisance covariate regression after filtering might reintroduce previously suppressed frequencies [Hallquist et al., 2013] might be another reasonable explanation for this effect being a result of the entangled interdependency of detrending, filtering and nuisance regression.

Of note, Liang et al. found evidence for high test‐retest reliability of both binary and weighted graph networks constructed from slow‐4 filtered resting‐state time series in short‐ and long‐term scans [Liang et al., 2012].

The current finding that nondetrended, slow‐4 filtered, non global mean regressed time series appear to have robust and reliable group differences in network topography, stands out in the context of all other preprocessing variants and regard should be paid.

Investigations by Zhang et al. indicate differences between MDD patients and control subjects in global path length and global efficiency [Zhang et al., 2011]. However, comparison between results of the study by Zhang et al. and the present one should be interpreted with caution, because of differences in patient cohort, age range, network construction method, and statistical analysis.

Effect of Preprocessing on Group Differences in Nodal Graph Metrics

Group differences in local graph metrics occurred sparsely and were highly dependent on an intricate interplay of preprocessing method, sparsity threshold, ROI, and local metric. On the basis of the current discussions in the community and previous investigations, we expected to find largest differences in occurrence of group differences between G and nG preprocessing variants. These differences became obvious when two of the most common preprocessing variants (D_b_G or D_b_nG, respectively) were compared. Group differences resulting from both preprocessing variants did neither show a similar nor an opposing pattern; they rather showed more complex differences. Local graph metrics are complementary in identification of altered nodal properties and an alteration in one graph metric is not necessarily accompanied by alterations in other metrics in the same node. Group differences were found in nonoverlapping regions of interest and were highly dependent on GMR. This finding underlines the importance of a considerate decision with and without this analysis step due to the resulting conclusions on (neuro‐) pathologies characterizing a particular disorder [Yang et al., 2014].

Therefore, we recommend considering results with and without global mean regression.

In a group comparison between patients with Autism Spectrum Disorder (ASD) and healthy control subjects, global signal regression was shown to lead to a reversal in the direction of group correlation differences relative to other preprocessing approaches, with a higher incidence of both long‐range and local correlation differences that favor the ASD group [Gotts et al., 2013]. Furthermore, the strongest group differences under other preprocessing approaches were the ones most altered by global signal regression and locations showing group differences no longer agreed with those showing correlations with behavioral symptoms within the ASD group [Gotts et al., 2013].

Our finding of convergent effects in some regions regardless of GMR replicates observations by [Horn et al., 2010], who found correlations of functional connectivity between ACC and anterior insula and glutamate level in ACC in both GM regressed and uncorrected data. In this study, we also found that for some connections, including thalamus, ACC, and parietal cortex, consistency of group differences was preserved. The dorsal ACC has been reported to show very robust cortical thinning in MDD [van Tol et al., 2014] accompanied with cellular [Steiner et al., 2011], molecular, and functional abnormalities [Li et al., 2014]. Likewise, structural aberrances have been reported for thalamus and parietal cortex [Osoba et al., 2013], which may contribute to very robust group effects that are more insensitive to preprocessing choice. In a prior study, MDD patients exhibited increased nodal centralities, predominately in the caudate nucleus and default‐mode regions, including the hippocampus, inferior parietal, medial frontal, and parietal regions as well as reduced nodal centralities in the occipital, frontal (orbital part), and temporal regions [Zhang et al., 2011]. In addition to the preprocessing reasons outlined in this article, this discrepancy might be very well related to the fact that Zhang et al. [2011] recruited unmedicated (drug‐naive), first‐episode MDD patients with a broad age range, whereas patients in the current sample were medicated, and mostly experiencing a recurrent depressive episode.

However, given the previous elaborations, interpretations about (non‐) significance of group effects in GM corrected data need to remain very cautious, since the neuro‐pathological basis of local aberrations is generally poorly understood for most neuro‐psychiatric disorders and an interpretive model of how abnormalities in default‐mode network (DMN) modulate patho‐physiology of MDD is yet to be developed [Wang et al., 2012].

While recent work in schizophrenia even discusses the global signal as a pathological feature [Yang et al., 2014], the results of the current study may not allow for advise on GSR, but we can give a general recommendation to consider both options since they lead to significant, however different results.

To establish relations between the findings of the current study and previously published studies on MD, to disentangle meaningful from incidental differences related to preprocessing, and to tentatively develop a basis close to the “ground truth,” we compared our findings with the ones from other studies.

We here base our selection on work by Wang et al, who have reviewed 16 studies on resting‐state fMRI and major depression [Wang et al., 2012]. Most of these studies chose a seed‐based ROI analysis approach, one used Independent Component Analysis (ICA), and 5 used Regional Homogeneity (ReHo). Because none of them used graph‐theoretic measures, as done in our study, we consider them to be equally relevant for inclusion. In summary of the seed based analyses, the abnormal involvement of the cortico‐limbic mood regulation circuit appeared to be the most robustly emerging pattern from the pool of analyzed studies. Furthermore, regions within the DMN have been shown to exhibit abnormal (increased and decreased) resting connectivity in MDD, suggesting its pervasive involvement. Abnormal FC of the anterior cingulate Cortex (ACC) and limbic structures has been found in several studies and has been suggested as a potential biomarker of MDD [Anand et al., 2005; Cullen et al., 2009; Horn et al., 2010; Liu et al., 2010; Wang et al., 2012].

Sheline et al. found a positive correlation with Hamilton depression rating scales (HAMD) and dorso‐medial prefrontal cortex (DMPFC, part of the DMN) connectivity [Sheline et al., 2010]. A network including the amygdala as well as subgenual and pregenual cingulate cortex showed increased activity in MDD patients [Sheline et al., 2010]. The authors speculate that disruption of this affective network could lead to hallmark symptoms of MDD (e.g., lost appetite, libido, and sleep disturbances). Zhou et al showed that depressed patients exhibited widespread increased in FC with both the task‐positive network (TPN) and the task‐negative network (TNN) [Zhou et al., 2010].

A comparison of the above mentioned regions identified to be involved in MDD pathology and the evidence of the current study (Table AII) revealed the greatest convergence when the preprocessing variants D_b_nG and D_4_nG were used. Here, the regions dorsal ACC, pregenual ACC, anterior insula, and inferior parietal lobe showed between‐group differences.

We approach the attempt to develop a basis for a “ground truth” with the following restraints: firstly the sample sizes per study are rather small, secondly diverging analysis approaches have been used, and thirdly medication history and clinical characteristics (e.g., comorbidities) between subjects are potentially confounding factors.

Robustness of Group Differences Over Sparsity Thresholds

With the objective to propose a recommendation for choosing a promising sparsity threshold, we investigated its effect on group differences.

Indeed, lower sparsity thresholds‐in the range of 10–24%‐favored occurrence of between‐group differences in nodal metrics for most of the variants. Another interesting fact is that the group differences based on global graph metrics occurred within the range of sparsity thresholds that also showed the majority of the group differences based on local graph metrics. Hence, sparsity thresholds between 16 and 22% were shown to have the greatest potential to reveal differences between healthy subjects and MDD patients on both levels of the network and thus it is recommended to explore networks on this sparsity range in the context of MDD.

Because of the reduction in group differences by inclusion of spurious connections observed at lower sparsities [Lord et al., 2012; Rubinov et al., 2009, and Fig. 1], findings observed at a lower sparsity are more likely to be robust, because the likelihood that they are also evident at higher sparsity thresholds is high and therefore can be interpreted with more confidence.

Using functional connectivity derived from resting‐state scalp EEG, Rubinov et al also found a greater between‐group separation for sparser networks in a study of schizophrenia [Rubinov et al., 2009]. Langer et al. approached the thresholding problem by determining the sparsity threshold on the basis of an average network across all subjects [Langer et al., 2012]. The average network was thresholded over a wide range of sparsity levels. The particular threshold that corresponded to a small‐world topology was chosen and subsequently applied to individual correlation matrices. Results based on this approach contain an inherent bias for finding small‐worldness characteristics. However, thresholding at a single level always meets with the criticism that the topological properties of the resulting graph may depend on the respective threshold value. As a general cautionary remark toward pathophysiological interpretations of specific graph metrics and their relationship to specific disease mechanisms, one has to consider that all of the tested graph metrics revealed group differences in at least some of the tested combinations despite robust correction for multiple comparisons. As evident in Table AII, not only regions, but also the individual nodal properties showing group differences systematically vary across the different preprocessing variants. Whether the impact of an illness is expressed in one graph metric or another—e.g., participation index or betweenness centrality—may thus be as much a question of the individual data handling, as it may be related to neurobiological principles.

Correlation of Disease Severity to Local Metrics

In a post‐hoc analysis, we checked for correlations of disease severity and local graph metrics based on previous evidence by [Zhang et al., 2011].

In contrast to the earlier findings, however, no evidence for a correlation of BCI in left hippocampus was found. Regarding effect of preprocessing on correlation with disease severity, evidence for a negative correlation of BCI in the left caudate and HAMD scores was solely found in preprocessing variants with global signal regression. Although we also found a significant correlation, these results differ from [Zhang et al., 2011] as the direction of the correlation is reversed. In addition to the already discussed methodological differences (Effect of Preprocessing on Group Differences in Nodal Graph Metrics section) between the study by [Zhang et al., 2011] and the present one, another confounding factor leading to this specific discrepancy might be the differing range of HAMD scores: the MDD patients investigated in Zhang et al. had HAMD scores in the range of 18–34 (higher disease severity), whereas our patients had lower HAMD scores (8–23, lower disease severity). A note of caution is due here because re‐testing significance of regions that previously revealed significant main effects can be considered as a result of double‐dipping. Also, may one question the suitability of the choice of regions with maximum group differences, which might inherently be limited in their within‐group variation, leaving limited variance to be explained by severity.

Limitations

Generalization of our results needs to proceed with caution, because investigations in the present study are limited to rs‐fMRI acquisitions at a magnetic field strength of 3 Tesla with comparably long data acquisitions. Most studies investigate data of shorter durations of 5–6 min. The influence we observed for detrending might not be relevant to other datasets, because thermal drift is specific to every scanner and also depends on the length of the scan itself and used sequence. A quantification of drift [Tanabe et al., 2002] can only be provided with extra efforts, for example by evaluating thermal fluctuations in simultaneously scanned heat phantoms.

More importantly, the effect of detrending on datasets that were further bandpass filtered needs to be seen in the context of the specific filter characteristics for the applied filter as implemented in DPARSF. Especially in case of different smoothness of roll off and cut off solutions for the passband of the filter used in other analysis software, this particular outcome will not necessarily be the same.

Several caveats of the present study should be kept in mind: First, we only had one dataset from each subject at hand and thus could not investigate repeated measures and test–retest reliability.

Second, our data were acquired during an eyes closed session: It has recently been shown that eyes open conditions lead to higher test–retest reliability, at least in healthy controls who were instructed to fixate a point on the screen [Patriat et al., 2013]. While this approach may be very well suited for studies investigating a ground truth neuroanatomical basis in motivated healthy controls, it would jeopardize a major benefit for psychiatric patient studies, namely it's low demand on the participant. Furthermore, one has to expect fundamental differences in participants' adherence to the task and as a consequence, differences in resting‐state activity, which is mainly due to differences in accuracy to constantly fixate e.g., a white cross on a black screen. Most prominent effects would be expected in the visual system, which however, as in the case of seasonal affective disorder [Borchardt et al., 2014] may be strongly involved in the pathology of even depressive disorders. While all our participants denied falling asleep on direct questioning, future investigations might rather incorporate explicit measures such as concurrent EEG recordings to avoid this problem and its associated risk of disease unrelated confounds in resting‐state activity.

Effects of motion on rs‐fMRI data have previously been described in depth (e.g., [Aurich et al., 2015; Glover et al., 2000; Lund et al., 2005; Power et al., 2012; Weissenbacher et al., 2009]). For example, Jo et al found that correlation estimates obtained after regression of the global signal are more affected by motion artifacts in a distance‐dependent manner [Jo et al., 2013]. In the present study, the relative contribution of small head motion could be minimized by the rather large voxel size. Because the dataset has very comparable voxel‐specific mean frame‐wise displacement indices (FD) [Power et al., 2012] between groups, the present work may be rather a reference of what else should be considered even if the FD is not likely to bias findings. Therefore, the conditions in terms of group constitution was identical for all preprocessing variants, while any higher order interaction of preprocessing and any potential confounder is possible and was not considered in our approach.

When using graph theory, network construction and all metrics are based on the number of vertices. Changing the number of ROIs or choosing overlapping ROIs always changes topological network properties and impacts nodal metrics. For example, Wang et al. examined the impact of parcellation scheme on brain networks and demonstrated that despite common small‐world topology, significant differences exist in nodal clustering coefficient and efficiency, and global characteristic path length, characteristic path length of the real network, and efficiency between two groups of parcellation schemes [Wang et al., 2009]. For the AAL atlas, nodal degree and efficiency were found to be positively correlated with regional size at several sparsity threshold levels [Zalesky et al., 2010].

Regarding the similarity measures, it would be interesting to compare the adjacency matrices likewise with other measures of similarity like the conformity measure as presented in [Chang et al., 2009].

Because publishing recommendations for rs‐fMRI preprocessing is challenging, a step beyond noting observed effects might be possible with the help of generative models or use of multimodal data [Fornito et al., 2013, 2015].

CONCLUSION

The primary conclusion of this study is the choice of preprocessing options has a striking effect on network properties in general and group differences in particular.

Here, evidence for a clear effect of global mean regression on correlation matrices, topological network properties as well as local graph metrics is presented and thereby this study again underscores the importance of decisions on data preprocessing strategies for subsequent results.

Differentiation between MDD patients and healthy controls based on local graph metrics is highly affected by preprocessing variant.

The evidence suggests that differences between healthy subjects and MDD patients based on local graph metrics can be found with most of the 12 studied preprocessing variants, but most importantly that finding group differences strongly depends on the chosen sparsity threshold. It seems that characteristic differences can be better revealed when the slow‐4 band is used for filtering and in general, if sparsity thresholds in the range of 10–24% are used.

Assuming that meta‐analytical disparity of a specific graph metric in a relevant region also originates from methodological variations during preprocessing, our set of intricate graph metric differences may suggest a focus on which preprocessing pipelines may be more likely to lead to the specific disparity. Until multisite and meta‐analytical data mining attempts in depression are disposable, we provide our observations as a guide for specific follow‐up investigations.

Most helpful recommendations at this point firstly provide a guide where MDD researchers interested in a particular region and possible graph metric could find the methodological parameters most likely to reveal differences in there, secondly underline the importance of potential additional analyses incorporating e.g., specific slow band‐filters and thirdly perform a systematic analysis of large scale between‐group differences by performing the comparisons suggested in our study.

More broadly, our investigation is in agreement with Carp et al. showing that the choice of preprocessing steps significantly affects both the results identified and their interpretation [Carp, 2012b]. Moreover, if different analysis strategies have varying or insignificant outcomes, all investigated methodological choices should be reported.

Table AI.

Abbreviated names of 95 regions of interest, their MNI coordinates, and class of brain region

#	Name	x	y	z	Class
1	Precentral_L	−38.65	−5.68	50.94	Primary
2	Precentral_R	41.37	−8.21	52.09	Primary
3	Frontal_Sup_L	−18.45	34.81	42.2	Association
4	Frontal_Sup_R	21.9	31.12	43.82	Association
5	Frontal_Sup_Orb_L	−16.56	47.32	−13.31	Paralimbic
6	Frontal_Sup_Orb_R	18.49	48.1	−14.02	Paralimbic
7	Frontal_Mid_L	−33.43	32.73	35.46	Association
8	Frontal_Mid_R	37.59	33.06	34.04	Association
9	Frontal_Mid_Orb_L	−30.65	50.43	−9.62	Paralimbic
10	Frontal_Mid_Orb_R	33.18	52.59	−10.73	Paralimbic
11	Frontal_Inf_Oper_L	−48.43	12.73	19.02	Association
12	Frontal_Inf_Oper_R	50.2	14.98	21.41	Association
13	Frontal_Inf_Tri_L	−45.58	29.91	13.99	Association
14	Frontal_Inf_Tri_R	50.33	30.16	14.17	Association
15	Frontal_Inf_Orb_L	−35.98	30.71	−12.11	Paralimbic
16	Frontal_Inf_Orb_R	41.22	32.23	−11.91	Paralimbic
17	Rolandic_Oper_L	−47.16	−8.48	13.95	Association
18	Rolandic_Oper_R	52.65	−6.25	14.63	Association
19	Supp_Motor_Area_L	−5.32	4.85	61.38	Association
20	Supp_Motor_Area_R	8.62	0.17	61.85	Association
21	Olfactory_L	−8.06	15.05	−11.46	Primary
22	Olfactory_R	10.43	15.91	−11.26	Primary
23	Frontal_Med_Orb_L	−5.17	49.17	−7.4	Association
24	Frontal_Med_Orb_R	8.16	50.84	−7.13	Association
25	Rectus_L	−5.08	54.06	−18.14	Paralimbic
26	Rectus_R	8.35	51.67	−18.04	Paralimbic
27	Hippocampus_L	−25.03	37.07	−10.13	Limbic
28	Hippocampus_R	29.23	35.64	−10.33	Limbic
29	ParaHippocampal_L	−21.17	6.65	−20.7	Paralimbic
30	ParaHippocampal_R	25.38	6.25	−20.47	Paralimbic
31	Amygdala_L	−23.27	35.4	−17.14	Limbic
32	Amygdala_R	27.32	37.01	−17.5	Limbic
33	Calcarine_L	−7.14	−14.92	6.44	Primary
34	Calcarine_R	15.99	−8.83	9.4	Primary
35	Cuneus_L	−5.93	−42.92	27.22	Association
36	Cuneus_R	13.51	−41.81	28.23	Association
37	Lingual_L	−14.62	−20.74	−4.63	Association
38	Lingual_R	16.29	−19.78	−3.87	Association
39	Occipital_Sup_L	−16.54	−15.95	28.17	Association
40	Occipital_Sup_R	24.29	−15.15	30.59	Association
41	Occipital_Mid_L	−32.39	−0.67	16.11	Association
42	Occipital_Mid_R	37.39	0.64	19.42	Association
43	Occipital_Inf_L	−36.36	−78.67	−7.84	Association
44	Occipital_Inf_R	38.16	−73.15	−7.61	Association
45	Fusiform_L	−31.16	−80.13	−20.23	Association
46	Fusiform_R	33.97	−79.36	−20.18	Association
47	Postcentral_L	−42.46	−67.56	48.92	Primary
48	Postcentral_R	41.43	−66.93	52.55	Primary
49	Parietal_Sup_L	−23.45	−84.26	58.96	Association
50	Parietal_Sup_R	26.11	−80.85	62.06	Association
51	Parietal_Inf_L	−42.8	−80.73	46.74	Association
52	Parietal_Inf_R	46.46	−79.7	49.54	Association
53	SupraMarginal_L	−55.79	−78.29	30.45	Association
54	SupraMarginal_R	57.61	−81.99	34.48	Association
55	Angular_L	−44.14	−40.3	35.59	Association
56	Angular_R	45.51	−39.1	38.63	Association
57	Precuneus_L	−7.24	−22.63	48.01	Association
58	Precuneus_R	9.98	−25.49	43.77	Association
59	Paracentral_Lobule_L	−7.63	−59.56	70.07	Association
60	Paracentral_Lobule_R	7.48	−59.18	68.09	Association
61	Caudate_L	−11.46	−45.82	9.24	Subcortical
62	Caudate_R	14.84	−46.29	9.42	Subcortical
63	Putamen_L	−23.91	−33.64	2.4	Subcortical
64	Putamen_R	27.78	−31.5	2.46	Subcortical
65	Pallidum_L	−17.75	−60.82	0.21	Subcortical
66	Pallidum_R	21.2	−59.98	0.23	Subcortical
67	Thalamus_L	−10.85	−56.07	7.98	Subcortical
68	Thalamus_R	13	−56.05	8.09	Subcortical
69	Heschl_L	−41.99	−25.36	9.98	Primary
70	Heschl_R	45.86	−31.59	10.41	Primary
71	Temporal_Sup_L	−53.16	11	7.13	Association
72	Temporal_Sup_R	58.15	12.07	6.8	Association
73	Temporal_Pole_Sup_L	−39.88	3.86	−20.18	Paralimbic
74	Temporal_Pole_Sup_R	48.25	4.91	−16.86	Paralimbic
75	Temporal_Mid_L	−55.52	−0.03	−2.2	Association
76	Temporal_Mid_R	57.47	0.18	−1.47	Association
77	Temporal_Pole_Mid_L	−36.32	−17.56	−34.08	Paralimbic
78	Temporal_Pole_Mid_R	44.22	−17.55	−32.23	Paralimbic
79	Temporal_Inf_L	−49.77	−18.88	−23.17	Association
80	Temporal_Inf_R	53.69	−17.15	−22.32	Association
81	Medial_Prefront_lower_L	−6.9	−20.68	16.9	Association
82	Medial_Prefront_lower_R	8.4	−21.78	15.3	Association
83	Medial_Prefront_upper_L	−4.8	15.14	44.9	Association
84	Medial_Prefront_upper_R	7.6	14.75	46.4	Association
85	Ant_Insula_L	−34.9	−33.8	0.3	Paralimbic
86	Ant_Insula_R	37.3	−37.23	−1.7	Paralimbic
87	Post_Insula_L	−37.9	14.59	8.2	Paralimbic
88	Post_Insula_R	38.9	14.55	7.2	Paralimbic
89	Rostral_ACC_bilateral	−2.3	−28.05	−5.7	Paralimbic
90	Pregenual_ACC_bilateral	0.4	−31.07	8	Paralimbic
91	Dorsal_ACC_bilateral	2.2	27	29.5	Paralimbic
92	Posterior_MCC_bilateral	0.2	−10	41	Paralimbic
93	23d_bilateral	0	−35.7	42.1	Paralimbic
94	dPCC_bilateral	−1.1	−43.2	26.8	Paralimbic
95	vPCC_bilateral	1.1	46.6	−12.7	Paralimbic

L and R stand for left and right, respectively.

Table AII.

List of all significant (P < 0.05, FDR corrected) differences between healthy controls and MDD patients in nodal graph metrics (listed as tuples of local graph metric and abbreviated ROI name) for 12 preprocessing variants and 16 sparsity thresholds

Preprocessing variant	Sparsity (%)	Group difference (P < 0.05, FDR)
		metric	ROI
D_b_G	10	Eloc	Parietal_Inf_R
		Eloc	Parietal_Inf_R
	12	Strength	Rectus_L
		Degree	Rectus_L
	14	Strength	Rectus_L
		Degree	Rectus_L
		PI	Parietal_Inf_L
	16	Strength	Rectus_L
		PI	Parietal_Inf_L
	18	Strength	Rectus_L
		Degree	Rectus_L
		PI	Parietal_Inf_L
		Degree	Frontal_Inf Tri_R
	20	Degree	Frontal_Inf Tri_R
	22	Degree	Frontal_Inf Tri_R
		PI	Parietal_Inf_L
	28	PI	vPCC_bilateral
D_b_nG	18	PI	Parietal_Inf_L
	26	PI	Calcarine_L
		PI	Calcarine_R
		PI	Lingual_R
		PI	Lingual_L
		PI	Postcentral_L
		PI	Temporal_Sup_R
	30	PI	Calcarine_L
	38	PI	dorsal_ACC_bilateral
		PI	Angular_L
		PI	Ant_Insula_L
		PI	Ant_Insula_R
		PI	pgACC_bilateral
D_4_G	10	Eloc	Parietal_Inf_R
		Eloc	Parietal_Inf_R
	12	Strength	Rectus_L
		Degree	Rectus_L
	14	Strength	Rectus_L
		Degree	Rectus_L
		PI	Parietal_Inf_L
	16	Strength	Rectus_L
		PI	Parietal_Inf_L
	18	Strength	Rectus_L
		Degree	Rectus_L
		PI	Parietal_Inf_L
		Degree	Frontal_Inf Tri_R
	20	Degree	Frontal_Inf Tri_R
	22	Degree	Frontal_Inf Tri_R
		PI	Parietal_Inf_L
	28	PI	vPCC_bilateral
	18	PI	Parietal_Inf_L
D_4_nG	26	PI	Calcarine_L
		PI	Calcarine_R
		PI	Lingual_R
		PI	Lingual_L
		PI	Postcentral_L
		PI	Temporal_Sup_R
	30	PI	Calcarine_L
	38	PI	dorsal_ACC_bilateral
		PI	Angular_L
		PI	Ant_Insula_L
		PI	Ant_Insula_R
		PI	pgACC_bilateral
D_5_G	12	Pathlength	Rectus_L
		Pathlength	Olfactory_L
	14	Pathlength	Rectus_L
	18	PI	vPCC_bilateral
D_5_nG	22	PI	pgACC_bilateral
	32	PI	Calcarine_L
	34	PI	Calcarine_L
		PI	Calcarine_R
		PI	Lingual_L
nD_b_G	14	PI	Thalamus_R
	20	Degree	Frontal_Inf_Tri_R
nD_b_nG	10	bci	pgACC_bilateral
	18	PI	pgACC_bilateral
	40	PI	Angular_L
nD_4_G	10	Eloc	Heschl_R
		LEGE	Heschl_R
nD_4_nG	10	Degree	dorsal_ACC_bilateral
	12	Degree	dorsal_ACC_bilateral
		Strength	dorsal_ACC_bilateral
		BCI	dorsal_ACC_bilateral
	14	Degree	dorsal_ACC_bilateral
		Strength	dorsal_ACC_bilateral
	16	Degree	dorsal_ACC_bilateral
		Strength	dorsal_ACC_bilateral
		Degree	ParaHipp_L
	18	Degree	dorsal_ACC_bilateral
		Strength	dorsal_ACC_bilateral
		Degree	ParaHipp_L
		Eloc	Frontal_Mid_Orb_R
	20	Degree	dorsal_ACC_bilateral
		Degree	ParaHipp_L
		PI	Postcentral_L
	22	Degree	dorsal_ACC_bilateral
		Degree	ParaHipp_L
		PI	Postcentral_R
		PI	Postcentral_L
	24	Degree	dorsal_ACC_bilateral
		Degree	ParaHipp_L
	26	Degree	dorsal_ACC_bilateral
		PI	Angular_L
	28	Degree	dorsal_ACC_bilateral
		BCI	dorsal_ACC_bilateral
		PI	Angular_L
		PI	Frontal_Mid_Orb_R
	32	PI	Angular_L
		PI	Frontal_Mid_Orb_R
nD_5_G	10	Degree	dorsal_ACC_bilateral
	16	BCI	Pallidum_R
	18	BCI	Pallidum_R
	22	BCI	Pallidum_R
	24	BCI	Pallidum_R
	26	BCI	Pallidum_R
	28	BCI	Pallidum_R
	30	BCI	Pallidum_R
	32	BCI	Pallidum_R
	34	BCI	Pallidum_R
		LEGE	Parietal_Inf_L
	36	BCI	Pallidum_R
		LEGE	Parietal_Inf_L
	38	BCI	Pallidum_R
		LEGE	Parietal_Inf_L
	40	BCI	Pallidum_R
		LEGE	Parietal_Inf_L
nD_5_nG	16	Eloc	Thalamus_R
		Eloc	Pallidum_L
		LEGE	Thalamus_R
		LEGE	Pallidum_L
	18	Eloc	Thalamus_R
		LEGE	Thalamus_R
	20	Eloc	Thalamus_R
		LEGE	Thalamus_R
	22	Eloc	Thalamus_R
		LEGE	Thalamus_R

L and R stand for left and right, respectively.

Table AIII.

Effect size measured in Cohen's d for group comparisons (HC versus MDD) in global clustering coefficient (cc_real) for every tested preprocessing variant and sparsity threshold

	Effect size of group difference in global cc of real network
Variant	10%	12%	14%	16%	18%	20%	22%	24%	26%	28%	30%	32%	34%	36%	38%	40%
D_4_nG	0.15	0.41	0.47	0.38	0.39	0.30	0.23	0.20	0.19	0.18	0.16	0.17	0.14	0.14	0.15	0.14
nD_4_nG	0.35	0.41	0.42	0.64	0.65	0.63	0.61	0.58	0.54	0.49	0.49	0.45	0.45	0.46	0.44	0.44
D_b_nG	0.15	0.41	0.47	0.38	0.39	0.30	0.23	0.20	0.19	0.18	0.16	0.17	0.14	0.14	0.15	0.14
nD_b_nG	0.34	0.48	0.52	0.51	0.34	0.23	0.24	0.17	0.13	0.11	0.11	0.08	0.11	0.13	0.14	0.15
D_4_G	0.35	0.19	0.09	0.04	0.04	0.03	0.09	0.08	0.08	0.05	0.02	0.01	−0.03	−0.05	−0.06	−0.08
nD_4_G	0.60	0.47	0.27	0.22	0.18	0.24	0.22	0.16	0.14	0.08	0.06	0.05	0.04	0.01	0.01	0.00
D_5_nG	0.05	0.09	0.16	0.10	0.08	0.03	0.02	−0.02	−0.06	−0.11	−0.12	−0.12	−0.10	−0.08	−0.07	−0.08
nD_5_nG	0.09	0.28	0.11	−0.04	−0.07	−0.01	−0.12	−0.10	−0.12	−0.14	−0.12	−0.12	−0.14	−0.13	−0.12	−0.10
D_b_G	0.35	0.19	0.09	0.04	0.04	0.03	0.09	0.08	0.08	0.05	0.02	0.01	−0.03	−0.05	−0.06	−0.08
nD_b_G	0.38	0.18	−0.01	−0.03	−0.04	0.00	0.04	0.01	−0.03	−0.06	−0.07	−0.10	−0.10	−0.11	−0.13	−0.16
D_5_G	−0.15	−0.05	−0.04	−0.02	0.00	0.02	−0.02	−0.06	−0.08	−0.09	−0.14	−0.15	−0.16	−0.17	−0.17	−0.17
nD_5_G	−0.15	−0.05	0.03	0.06	−0.03	−0.02	−0.05	−0.08	−0.11	−0.15	−0.17	−0.17	−0.18	−0.18	−0.19	−0.18

The framed entries mark significant group differences (P < 0.05).

Figure A1.

This schematic drawing visualizes steps one to three of the robustness estimation to assess robustness of differences between healthy controls and MDD patients in local graph metrics regarding the whole range of tested sparsity thresholds.

Figure A2.

Effect sizes of the group comparison MDD patients versus healthy controls in the global graph metric clustering coefficient. The colored bars indicate the effect size of each preprocessing variant at a sparsity threshold of 18%. To demonstrate the effect of detrending on effect size, each pair of preprocessing variant that only differs in application (D) or omission (nD) of detrending is plotted next to each other in the same color.