Assessments
Screening. HC participants were screened for the absence of mental health symptoms using questions used to predict diagnostic status from a prior population cohort study of child and adolescent’s psychopathology [34]. Children who scored above the screen cut-off were excluded from further participation but were given a small prize.
Determination of Diagnosis and Diagnostic History. We attempted to recruit all individuals with a history of AN who presented at an outpatient medical clinic. These individuals were at various stages of the disorder in terms of degree of weight severity or restoration, and thus the sample had significant heterogeneity. While this had the disadvantage of preventing comparisons between categorical stages of the disorder, it had the advantage of high external validity in that this group reflected individuals presenting for care. We describe individuals dimensionally and categorically. For both, diagnosis and parameters of illness history were determined by systematically combining several sources of data: 1) maternal report of her child’s illness history; 2) adolescent completion of self-report measures of current symptoms; 3) adolescent report of illness history; and 4) medical chart abstraction. This included both BMI and zBMI (i.e., age-adjusted BMI, which accounts for height, weight, and age) and age-adjusted weight percentile. To be classified as a HC, 1) parent report indicated no history of an eating disorder; 2) adolescent self-report of Drive for Thinness values were within 1 standard deviation of normative values; and 3) the medical chart contained no reference to an eating disorder diagnosis.
For the clinical group, we employed the ENIGMA Eating Disorders consortium definitions of weight status to define individuals with AN that were currently ill or partially weight-restored. These definitions were complemented with parent report of disorder history and self-reports of Drive for Thinness as described below. The ENIGMA consortium (http://enigma.ini.usc.edu/about-2/) is an international effort combining data across research sites to accelerate the study of health and disease across development. ENIGMA definitions of weight status are employed in this manuscript for ease of comparison across studies. To these definitions of weight status, we added benchmarks for scores on eating disorder measures as defined below. Acute AN (AN) is a BMI of ≤17.5 kg/m2, < 10th for weight according to age-adjusted weight-percentile, and not in a period of rapid weight gain (<2 kg. over the past 4 weeks). This weight definition was complemented with the following definitions for AN in this study: 1) parent records indicated the child had AN within past 3-6 months; 2) medical chart had a diagnosis of AN; and 3) the adolescent had a Drive for Thinness score >1 std above normative values. For partially weight-restored AN (ANp), according to ENIGMA: participant does not meet criteria for acute AN and either: A) BMI is <18.5 kg/m2 or <10th adjusted percentile; or B) BMI is >18.5 kg/m2 but <19.5 kg/m2, age-adjusted percentile is >10th but <25th, participant must not have regular menses, and still show significant eating disorder symptoms as defined in this study as > 1 standard deviation of Drive for Thinness Normative Values. In this study, weight-restored AN (ANwr), was defined as: 1) BMI ≥ 18.5 or the parent report indicated that the child was without an eating disorder for 3-6 months; 2) the medical chart review did not contain a current diagnosis of AN and 3) there was no evidence of a medical sign that weight was low (e.g., bradycardia, orthostatic hypotension). To determine length of illness, mothers were asked the age at which their child first developed an eating disorder, the type of eating disorder, and whether this diagnosis was verified by a health care professional. This information was compared and combined with the medical record and referenced against the child’s weight history, current weight, and current symptom endorsement. In only one case was there a discrepancy. In this case, the parent indicated that the child no longer had an eating disorder and had been at a healthy weight for 3-6 months. However, the child’s weight and endorsement of clinical symptoms were both above clinical cut-offs. Of interest, this child had a long duration of illness (>7 years) and a lowest BMI of 11. Her current BMI of 18, may have seemed to present as significant progress (as it was), yet an anchor of normality had been lost.
A similar strategy was employed to determine months of weight restoration. Parents were asked the length of time the child had been at a healthy weight and this was verified relative to the child’s weight history and medical record. Again, there was one discrepancy, noted below.
Self-report measures. The Eating Disorder Inventory (3rd Edition) is one of the most widely used measures of eating disorder symptomatology and associated features [38]. This measure was used to characterize the sample relative to other studies and provide a continuous index of current symptoms. Three subscales that measure the core pathology of eating disorders were administered in the current sample: Drive for Thinness, Bulimia, and Body Dissatisfaction. All scales have extensive validity and reliability information as well as normative data from clinical and non-clinical samples. The Drive for Thinness subscale is a 7-item scale that assesses “an extreme desire to be thinner, preoccupation with weight, and an intense fear of weight gain”. Extensive reliability, construct, and predictive validity have been established [39–41]. The internal consistency of this scale was measured via Cronbach’s alpha, which is a measure of internal consistency (between 0 and 1), or how closely related a set of items are as a group. High values indicate high reliability [42]. Cronbach’s alpha for the Drive for Thinness subscale in our sample was α = 0.95. The Bulimia subscale is an 8-item scale used to index the tendency to think about or engage in uncontrollable overeating or eating in response to emotions. The internal consistency in our sample was α = 0.89. The Body Dissatisfaction subscale is a 7-item scale that assesses discontentment with the size and shape of various body parts that are of particular concern to those with eating disorders (e.g., stomach). The internal consistency in our sample was α = 0.94. We also looked at the Perfectionism subscale of the EDI, a 6-item scale that evaluates the personal value that individuals place on personal achievement and meeting their own high standards. The internal consistency in our sample was α = .91
2.3. MRI Acquisition and Quality Control.
Whole brain structural and functional (resting state) data was acquired using MRI scans conducted on a 3 Tesla General Electric MR 750 system with 50-mT/m gradients and an 8-channel head coil for parallel imaging (General Electric, Waukesha,WI, USA). Noise reducing headphones were used. To control for the state of acute nourishment on brain activity parameters [43], individuals were asked to fast for 2 hours prior to the scan and then were asked to consume a small, standardized snack just prior to the scan [43]. Twelve of the participants were on medication the day of scanning.
2.3.1. Structural gray-matter.
For registration purposes, a high-resolution structural image was obtained from each subject using a magnetization-prepared rapid acquisition gradient-echo sequence in the axial plane (Ax FSPGR BRAVO, repetition time=7.58ms, structural acquisition time = 3 min 22 sec, echo time = 2.936ms, inversion time = 450 ms, flip angle = 12°, slice thickness = 1mm, 256 slices, 256 x 256 voxel matrix, 1mm voxel size).
2.3.2 Resting State functional connectivity.
Resting state fMRI data was acquired using the following parameters: (34-slice, 150 whole brain volumes, interleaved slices, slice thickness = 3.8 mm, TE: 30 ms, TR 2000 ms, resting-duration (TA) = flip angle = 70°, acquisition matrix = 64 x 128, field of view = 243 mm x 243 mm). Subjects rested with eyes open and instructed to fixate on a cross while functional blood oxygen-level dependent images were acquired.
2.5. MRI Processing.
2.5.1. Resting state functional connectivity pre-processing.
Resting state preprocessing was conducted using SPM12 software (Welcome Department of Cognitive Neurology, London, UK). The first two volumes were discarded to allow for stabilization of the magnetic field. Slice timing correction was performed first, followed by rigid six-degree motion-correction realignment. The motion correction parameters in each degree were examined for excessive motion. If any volume-to-volume motion correction parameter was above 2 mm translation or 2° rotation, it was excluded from the dataset. To robustly take account of the effects of motion, root mean squared (RMS) realignment estimates were calculated as robust measures of motion using publicly available MATLAB code from GitHub [44]. Any subjects with a RMS value greater than 0.25 were not included in the analysis [44]. No participants had a RMS value greater than 0.25. The resting state images were then co-registered to their respective anatomical T1 images. Each T1 image was then segmented and normalized to a smoothed template brain in 2mm Montreal Neurological Institute (MNI) template space. Each subject's T1 normalization parameters were then applied to that subject's resting state image, resulting in an MNI space normalized resting state image. The resulting images were smoothed with a 5 x 5 x 5 mm3 FWHM Gaussian kernel. For each subject, a sample of the volumes was inspected for any artifacts and anomalies. Levels of signal dropout were also visually inspected for excessive dropout in a priori regions of interest.
2.5.1.1 Functional Network Connectivity Construction.
Preprocessed and normalized functional images were entered into the CONN-fMRI functional connectivity toolbox version 17 [45]. Regions from the Destrieux [46] and Harvard-Oxford Subcortical Atlases were entered as ROIs. These atlases were used to accurately capture the ROIs mentioned in the previous research. CompCor, a component-based noise correction method, was applied to remove physiological noise without regressing out the global signal [47]. White-matter, cerebrospinal fluid, six realignment parameters, and first-order temporal derivatives of motion, and RMS were removed using regression. This ensures only signal from gray matter voxels are being considered. Band pass filtering between 0.01 and 0.08 Hz was applied to the residualized time series to reduce the low- and high-frequency noises after regression. Linear measures of ROI-to-ROI functional connectivity were computed using Fisher transformed correlations representing the association between average temporal BOLD time series signals across all voxels in a brain region. The final outputs for each subject consisted of a 165 x 165 matrix consisting of Fisher transformed Z correlation values between each ROI. Overall functional connectivity was computed by taking the mean of all positive values in each individual’s matrix. This was done to determine if proportional thresholding should be used, as because a minimal difference in overall FC can cause a difference in network metrics, which may be due to inherent disease differences [48]. An independent sample t-test was done to determine if there was a significant difference in overall FC between groups.
2.5.1.2 Computing Network Metrics.
The Graph Theory GLM toolbox (GTG) (www.nitrc.org/projects/metalab_gtg) and in-house MATLAB scripts were applied to the subject-specific functional brain networks to compute two local weighted network metrics indexing centrality. We decided to focus on centrality at the microscale level as this allows one to determine characteristic hub roles for specific nodes, or brain regions, which can be easily interpretable by scientists and clinicians alike [49]. Additionally measures of centrality can capture a node’s role in network organization beyond local connections [49]. Measures of centrality quantify the importance of a region’s influence on communication and information flow in large-scale brain networks [25]. These measures include strength and betweenness centrality [23]. Strength represents the weighted sum of the number of connections a given brain region has and reflects a brain region’s total impact in the network [23]. Betweenness centrality describes the degree to which a brain region lies on the shortest path between two other regions [23]. Acting as way stations, regions with high betweenness centrality are topologically primed to control communication between other regions. The magnitude of the Z values represents the weights in the functional network. The Z values in each individual connectivity matrix was thresholded at Z > 0.3, and all other values were set to zero. A threshold of 0.3 was chosen since a correlation of 0.3 represents a medium effect size, and the inclusion of lower correlations could result in the inclusion of less accurate estimates [50]. We did not use a proportional-based thresholding approach because minimal differences in overall functional connectivity may introduce group differences in network metrics in patient vs. control studies [48]. All visualizations were created using in-house visualization schematics along with the BrainNet Viewer [51].
2.6. MRI Data Analysis.
2.6.1. Regions of Interest.
Many of our analyses were based on regions of interest (ROI). For comparison between the combined AN group, consisting of both ANC and ANR individuals, and the HC group, core ROIs of the sensorimotor and basal ganglia networks were examined in relationship to the entire brain parcellated by the Destreiux [46] cortical and Harvard-Oxford subcortical [52–55] atlases, as well as seed-to-voxel whole brain analyses (Supplemental Table 2, Supplemental Figure 1). Core regions of the sensorimotor network included the thalamus [Tha], brain stem [Bstem], hippocampus [Hip], paracentral lobule and sulcus [PaCL/S], primary somatosensory cortex [S1], central sulcus [CS], primary motor cortex [M1], precuneus [PrCun], secondary somatosensory cortex [S2], supplementary motor area [M2], middle insula [part of aINS], and posterior insula [pINS] [10, 11]. Core regions of the basal ganglia network included the basal ganglia [BG] and globus pallidus [Pal] [56]. These core seed ROIs were selected from past research in AN and used to look at differences throughout ROIs across the entire brain parcellated by the Destreiux [46] cortical and Harvard-Oxford subcortical [52–55] atlases, as well as seed-to-voxel whole brain analyses.
2.6.2. Functional seed-to-voxel whole-brain analysis.
In order to determine differences in whole brain connectivity from selected ROIs, we performed a whole brain, seed-to-voxel analysis in CONN utilizing the GLM and controlling for age. This represents the level of functional connectivity between each ROI and every voxel in the brain. The parametric map of t-values were thresholded using an initial height threshold (voxel-level) of p < 0.001 and corrected (using the false discovery rate method); cluster thresholds were set at p(FDR) < 0.05 [57]. In order to perform partial correlations controlling for age with behavioral variables, eigenvalues for each connectivity unit (i.e, the degree of connectivity between the seed and significant cluster of voxels) were extracted from within the CONN toolbox. Significance was set at p < 0.05. Visualizations were done using circus [58] in Linux.
2.6.3 Computing group differences in network metrics.
In order to test for disease-related differences, a GLM was applied and the effect of age was included as a covariate in the model. Significance was determined via Freedman & Lane’s non-parametric permutation testing strategy and specifying 10,000 permutations [59]. This method provides good control over type I error rates and is robust to the presence of outliers [60]. Probability values from the permutation testing strategy were corrected using a false discovery rate (FDR) adjusted p-value, where q <0.05 was considered significant [61, 62]. FDR correction was applied at the whole-brain level. Partial correlations controlling for age were then computed to determine the association between significant network metrics and behavioral measures. Significance was set at p < 0.05.
2.7. Behavioral/Clinical Data.
Group differences in clinical and behavioral measures were evaluated by applying linear contrast analyses in a GLM model using Statistical Package for the Social Sciences (SPSS) software (version 22). To quantify the differences between the various contrasts, we calculated Cohen’s effect size d, reflecting differences on the scale of standard deviation units, where values are interpreted as low (d = .20), moderate (d = .50), and high (d = .80) [63]. Correlations between significant findings for group differences in connectivity and measures of centrality were conducted with behavioral variables, while controlling for age.