A detailed outline of the participant inclusion, the task, as well as fMRI data acquisition and preprocessing is described elsewhere [17, 18]. Below are summaries for each section.
Participants
Patients with BPD or Cluster-C personality disorders were recruited from mental health clinics at two sites in the Netherlands (Maastricht, Heerlen) and three sites in Germany (Freiburg, Lübeck, Hamburg), see “Brain parcellation and subject inclusion” of the results for the subject inclusion. Non-patient controls (NPC) were recruited from the general population at each site. Participants had to be hetero- or bisexual women aged 18–65 years, and were assessed by trained interviewers according to the DSM-IV criteria using the Structural Clinical Interview (SCID) II and I [20]. All participants provided written informed consent. The study was approved by the local ethical committees.
Task
Participants performed an adapted version of an emotion regulation paradigm, which involved the presentation of pictures that were preceded by a safe (emotion regulation) or look instruction. During the safe trials, participants were asked to imagine themselves as being in a safe situation. This regulation strategy was based on a central element of schema therapy, one of the main therapies for BPD [17]. There were 4 categories of picture stimuli (negative, neutral, positive, and erotic), and the task consisted of 96 trials divided into 4 runs of 24 trials each (see Fig. 1a for an outline of a trial). As part of the scanning session, participants also underwent resting-state scans before and after the task.
fMRI data
Acquisition and preprocessing
Functional and structural MRI data were acquired with 3 T scanners. The Functional images were acquired with a T2*-weighted echo planar imaging (EPI) sequence, with the following parameters: TR = 2000 ms, TE = 27 ms, flip angle = 90°, FoV = 192 × 192 mm, voxel size = 3 × 3 × 3 mm, and matrix = 64 × 64. In Maastricht 240 images and in Freiburg and Lübeck 252 images were collected. The number of interleaved axial slices in one volume was 32 in Maastricht and 34 in Freiburg and Lübeck. In Maastricht and Freiburg, the T2*-weighted slices were adjusted with a negative tilt of 30°, with the goal of minimizing susceptibility and distortion artifacts within the amygdala. The anatomical images were acquired with a T1-weighted sequence, with the following parameters: TR = 2250 ms, TE = 2.6 ms, flip angle = 9°, Field of View (FoV) = 256 × 256 mm, voxel size 1 × 1 × 1 mm. In total, 192 images were obtained in Maastricht, 160 in Freiburg, and 170 in Lübeck. The preprocessed images from a previous study were used (see [17] for details of the preprocessing pipeline) and transformed from BrainVoyager format into nifti format, using Neuroelf (www.neuroelf.net).
Analyses
Network estimation
Phasic brain connectivity networks were estimated by performing a whole-brain psychophysiological interaction (PPI) analysis. PPI analyses aim to identify brain regions, where the time-series (physiological signal) connectivity is moderated by a task condition (“the psychological variable”) [21]. Traditionally, this approach has consisted of predefining a source (or seed) region and then estimating its connectivity with other (target) regions. We extended this rationale to a whole-brain network method (see [15, 16] for similar approaches). Here, each brain node is once considered to be the source, while the other nodes are the target. The PPI terms were estimated separately for each task regressor, but since PPI analyses are in general less powerful compared to estimating activation (i.e., the main effect of a task regressor) [22], we opted for a summary model that collapsed the different emotional valences (negative, neutral, positive), and included the general look and safe condition. Each model contained the time-series of the source region, the task regressors (safe and look), the interaction terms (PPIs) and the confound regressors (motion parameters, cue presentation, and ratings) (Fig. 1b). This procedure resulted in two different brain connectivity matrices for each run, consisting of contrast estimates for the safe and look condition. The connectivity matrices were then made symmetric by averaging corresponding (a-b, and b-a) parameter estimates, and were subsequently averaged across the four different runs. Subsequently the absolute values of the connectivity matrices were taken, and the connectivity matrices were additionally thresholded at 5%, (i.e., retaining only the strongest 5% of the connection).
To estimate the tonic brain connectivity an approach was followed as described by [15, 23]. For each time-series the task-related variance was “removed” by regressing a model containing task regressors and confounds (motion parameter, cue presentation, and ratings) on the time-series, and subsequently using the residual time-series for further analyses (Fig. 1b). The connectivity matrix was estimated by applying a graphical lasso [24] to the residuals of all nodes with the function graphicallasso.m (https://statweb.stanford.edu/~tibs/glasso) for a range of regularization parameters, lambda. The optimal regularization parameter lambda was estimated for each network (each run and each subject) by minimizing the Bayesian Information Criterion [25, 26]. These connectivity matrices were then averaged across the four different runs. The absolute value of the connectivity matrix was taken and entered in a graph theoretical analysis.
As a control to the connectivity analyses, the task-related activity per region was also estimated by a basic regression analyses of the task condition (safe and look, and safe > look) on the time-series of each region.
Functional module assignment
Each of the resulting brain nodes was assigned to a higher level functional module as proposed in [19]. The functional module assignment was performed as follows: the term related to each functional module was entered in Neurosynth [27], an automated meta-analysis tool; “Emotion”, “Motivation”, “Cognitive Control”, “Default Mode”, resulting in four (“association test”) images. The cortical region related to the term “emotion” that showed overlap with the cognitive control regions was assigned to the latter. Finally, the correspondence of the brain regions of the above-described parcellation and the four functional modules was estimated by testing the overlap of each region with any voxel of the resulting Neurosynth maps, see Fig. 4.
Network measures
The resulting phasic and tonic networks (absolute weighted graphs), where used to estimative several network measures [6], using the Brain Connectivity Toolbox (https://sites.google.com/site/bctnet/). For each node, the strength (the sum of connectivity values), local efficiency (average inverse shortest path length of a node), and the participation coefficient (diversity of intermodular connections of individual nodes) [28] were estimated (Fig. 1c). These measures were selected to capture some of the essential properties of network nodes (i.e., integration and segregation) which are suggested to be related to personality [29].
Classification: linear support vector machine
To estimate the predictive accuracy in classifying borderline personality disorder, a linear support vector machine (SVM) function implemented in Matlab (The MathWorks, Inc) was applied to the brain network measures and the task activity contrast estimates as a set of features. The advantage of using the network measures as features is that it yields a reduced set (here 121 features) of variables for classification, compared to using the element-wise connectivity (here 7260), and can thus also be regarded as a dimensionality reduction method while allowing inference on the brain node level.
The different network measures (strength, participation coefficient, local efficiency) for the phasic (safe > look, and the safe + look contrast) and tonic connectivity, as well as the main effect of the task (also safe > look, and the safe + look contrast) where entered as feature models (resulting in 11 different models). Before the classification procedure, the input feature data was “corrected” for the different sites by performing a regression analysis with dummy coded site regressors and then using the resulting residuals.
Each group was compared to each other group (one vs. one classification) in a nested-cross-validation procedure, where the balanced accuracy (the average of the sensitivity and specificity) served as the key outcome measure (effect size) of the classification performance. The data was divided in tenfold validation and test/train data. In the inner loop of the nested cross-validation, the test/train data was further divided in a fivefold cross-validation train and test data to estimate the balanced accuracy per model. The feature model with the highest balanced accuracy was then fitted to all test/training data of the inner loop, and the model was evaluated on the performance of the validation data of the outer loop [see Fig. 1 for an illustration of the procedure]. This process was repeated 200 times to obtain a stable estimate of the mean balanced accuracy and a 95% confidence interval.
The SVM classification was then repeated 1000 times with randomly permuted labels [25], to obtain a permutation null distribution. The p-value was defined as the number of times the null distribution showed a balanced accuracy higher than the average balanced accuracy of the validation data divided by the total number of permutations. Follow-up analyses were subsequently performed to aid the functional interpretation of the best performing model of the inner loop: (1) the SVM model weights were averaged across folds and repetitions resulting in an averaged model of feature weights, indicating the relative (by ranking [30]) contribution of each brain region in the classification (2) groups were tested on the difference of the network measures per functional module (emotion, motivation, cognitive control, default mode). The Matlab code for the network and SVM prediction analyses is available at https://github.com/henkcremers/NetworkAnalysis.