|
nif:isString
|
-
The trauma-exposed group in this study included survivors of a subway fire disaster in South Korea [9, 10]. The initial assessment was conducted for 38 individuals aged 18 to 50 years at approximately 1.65±0.74 months after the disaster (time 0). Among them, 30 completed the time 1 assessment (time 1), 25 completed up to the time 2 assessment (time 2), and 17 completed up to the time 3 assessment (time 3). The diagnosis and symptoms severity of PTSD were assessed using the Structural Clinical Interview for the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) [15] and Clinician-Administered Posttraumatic Stress Disorder Scale (CAPS) [16], respectively. Details on exclusion criteria for participation in this study are described in S1 Text. Age- and sex-matched healthy controls were initially enrolled as the trauma-unexposed group. Out of 36 healthy individuals from the original cohort [9], 29 age- and sex-matched healthy individuals, who completed two or more assessments, were included as the trauma-unexposed group in the present study [10] (Supplementary Table). Among them, all 29 completed up to the time 2 assessment and 21 completed up to the time 3 assessment. Serial assessment data of 30 trauma-exposed individuals and 29 trauma-unexposed individuals who undertook two or more assessments were eventually included in the final analysis. Both the trauma-exposed and unexposed groups underwent three waves of assessments including multimodal neuroimaging and comprehensive psychiatric evaluations over a 5-year period at approximately 1.3-year intervals: time 1 at 1.43±0.22 years, time 2 at 2.68±0.32 years, and time 3 at 3.91±0.33 years after the trauma. Between the trauma-exposed and unexposed groups, there were no differences in sex, age, education, and handedness (Table 1A).
Table data removed from full text. Table identifier and caption: 10.1371/journal.pone.0177847.t001 Demographic and clinical characteristics of participants. SD, standard deviation; NS, not significant at P = 0.05; PTSD, posttraumatic stress disorder; CAPS, Clinician-Administered Posttraumatic Stress Disorder Scale.
Written informed consent was obtained from all participants. The study protocol was approved by the Institutional Review Board of the Seoul National University Hospital and all procedures contributing to this work comply with the latest version of the Declaration of Helsinki.
Serial brain scans were performed using a Signa EXCITE 3.0T MRI system (GE Healthcare, Milwaukee, WI, USA). At each assessment, high-resolution T1-weighted sMRI data were obtained using a three-dimensional inversion recovery spoiled gradient-echo pulse sequence with the following parameters: echo time (TE) = 1.4 ms, repetition time (TR) = 5.7 ms, flip angle (FA) = 20°, matrix size = 256 x 256, field of view (FOV) = 22 cm, and slice thickness = 0.7 mm. Also, dMRI data were acquired using a dual spin-echo echo-planar imaging sequence with the following parameters: 25 diffusion directions, b values = 0 s/m2 for no diffusion weighting and 1,000 s/m2 for diffusion weighting, TE = 90 ms, TR = 10,000 ms, matrix = 256 x 256, FOV = 24 cm, and slice thickness = 3.5 mm.
Brain regions of the fear circuitry including the amygdala, OMPFC, hippocampus, insula, and thalamus were selected as ROIs for the extraction of brain structural features. The OMPFC was determined to consist of the ventromedial and adjacent orbitofrontal cortices [17, 18]. Detailed methods regarding the delineation and subsequent processing of ROI masks are described in S1 Text.
A total of 23 brain structural features were derived from the analysis of multimodal neuroimaging data including sMRI and dMRI data (Fig 1 and Supplementary Fig). For the five ROIs, structural features included four categories: (i) local features from sMRI data including the volume of the amygdala and the grey matter density of the other four ROIs; (ii) region-wise connectivity features from dMRI data including connection density and connection cost in each ROI; (iii) pair-wise connectivity features from dMRI data corresponding to tract strength between the amygdala and each of the other four ROIs; and (iv) network features from dMRI data including network efficiency for the network composed of all five ROIs and for each subnetwork composed of three ROIs.
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0177847.g001 Multimodal characteristics of the amygdala, OMPFC, hippocampus, insula, and thalamus assessed at each time point.A set of candidate brain structural features was derived from multimodal neuroimaging data analysis, which comprehensively characterized local, region-wise connectivity, pair-wise connectivity, and network features of the amygdala, orbitofrontal and ventromedial prefrontal cortex (OMPFC), hippocampus, insula, and thalamus. The volume of the amygdala was measured by manual segmentation, and the grey matter density of the OMPFC, hippocampus, insula, and thalamus was assessed by employing voxel-based morphometry (For details, refer to S1 Text). Grey matter density of each ROI was provided by averaging values over all voxels in the ROI.
Fiber tracts among the five ROIs were reconstructed using the streamlining method as implemented in the TrackVis software (http://trackvis.org/) (For details, refer to S1 Text). For each pair of the ROIs, reconstructed fiber tracts were counted, and then the number of fiber tracts was divided by the summed volume of interconnected ROIs to measure connection density [19], and it was multiplied by the average length of fiber tracts to compute connection cost [20]. Region-wise connection density and cost were obtained by averaging the pair-wise values for each ROI.
Fiber tracts between the amygdala and each of the other four ROIs were reconstructed using the probabilistic method as implemented in the FDT tool of the FSL software (http://fsl.fmrib.ox.ac.uk/) (For details, refer to S1 Text). The seed-based classification procedure [21] was adopted to assess relative tract strength between the amygdala and the other four ROIs. For each voxel of the amygdala, the count ratio of fiber tracts reaching only the specific ROI to fiber tracts reaching any of the four ROIs [22, 23] was computed. Relative tract strength was provided for each pair of the ROIs (amygdala-OMPFC, amygdala-hippocampus, amygdala-insula, and amygdala-thalamus) by averaging the ratio values over all voxels except those with the ratio value below 0.01.
Network efficiency measures how efficiently information is exchanged over a network [24]. The measure was computed for a structural network that was constructed based on the density of reconstructed fiber tracts among five bilateral ROIs (OMPFC-amygdala-hippocampus-insula-thalamus), and each of their subnetworks determined for three bilateral ROIs (OMPFC-amygdala-hippocampus, OMPFC-amygdala-insula, and OMPFC-amygdala-thalamus). Local feature values and region-wise and pair-wise connectivity feature values were summed or averaged over the two hemispheres to preclude preference for either hemisphere. In addition, all feature values were adjusted for age and sex, and amygdala volume was additionally adjusted for the intracranial volume. For each structural feature, the trauma-exposed individuals' values were converted into standardized Z-scores in terms of the standard deviation (SD) from the mean of the trauma-unexposed group. Resultantly, the feature values of the trauma-exposed group had a distribution with reference to the mean and SD values of the trauma-unexposed group, such that a negative Z-score of a trauma-exposed individual indicated that their adjusted feature value was below the mean of the trauma-unexposed group in SD units.
Classification between the trauma-exposed and unexposed groups was performed based on the 23 multimodal features (Fig 1 and Supplementary Fig) of the five ROIs. We used logistic regression with ridge estimators as a classification algorithm [25] as implemented in the Waikato Environment for Knowledge Analysis interface [26]. At each time point, seven different classification models with 4 to 10 brain structural features were constructed. The priority of the brain structural features among 23 was determined based on the rank of point-biserial correlation coefficients of individual features (Fig 2).
Figure data removed from full text. Figure identifier and caption: 10.1371/journal.pone.0177847.g002 The relationships between candidate brain structural features and the group membership at each time point.The graph presents point-biserial correlation coefficients (r) between candidate features and the group membership at (A) time 1, (B) time 2, and (C) time 3 assessments. Error bars represent standard errors, which were calculated using 5,000 bootstraps. Asterisks in each graph indicate the first 10 brain structural features based on the rank of the absolute r values. Amy, amygdala; OMPFC, orbitofrontal and ventromedial prefrontal cortex; Hippo, hippocampus; Thal, thalamus.
For each classification model, performance for correctly assigning an individual to the trauma-exposed group was assessed in terms of the area under a receiver operating characteristic curve (AUC). We used bootstrap tests (n = 1,000) to assess whether an AUC is significantly greater than the chance level (AUC = 0.5) [27] and to compare AUCs between any two classification models [28]. The classification model with the greatest AUC was then determined as the best model for classifying the trauma-exposed group from the trauma-unexposed group at each time point. Having identified the subset of brain structural features corresponding to the best classification model, the accuracy of each individual feature on its own was measured using the AUC. All statistical analysis was performed using the Stata/SE software (release 13.1) (Stata Corp LP, College Station, TX, USA), and the statistical significance level was set at P = 0.05 in all statistical inferences.
|
![[This page as RDF]](http://data.gesis.org/somesci/static/rdf-icon.gif){kind=icon}