Resting-state connectivity within and across neural circuits in anorexia nervosa.

INTRODUCTION
Obsessional thoughts and ritualized eating behaviors are characteristic of Anorexia Nervosa (AN), leading to the common suggestion that the illness shares neurobiology with obsessive-compulsive disorder (OCD). Resting-state functional connectivity MRI (rs-fcMRI) is a measure of functional neural architecture. This longitudinal study examined functional connectivity in AN within the limbic cortico-striato-thalamo-cortical (CSTC) loop, as well as in the salience network, the default mode network, and the executive control network (components of the triple network model of psychopathology).


METHODS
Resting-state functional connectivity MRI scans were collected in unmedicated female inpatients with AN (n = 25) and healthy controls (HC; n = 24). Individuals with AN were scanned before and after weight restoration and followed for one month after hospital discharge. HC were scanned twice over the same timeframe.


RESULTS
Using a seed-based correlation approach, individuals with AN had increased connectivity within the limbic CSTC loop when underweight, only. There was no significant association between limbic CSTC connectivity and obsessive-compulsive symptoms or prognosis. Exploratory analyses of functional network connectivity within the triple network model showed reduced connectivity between the salience network and left executive control network among AN relative to HC. These abnormalities persisted following weight restoration.


CONCLUSIONS
The CSTC findings suggest that the neural underpinnings of obsessive-compulsive symptoms may differ from those of OCD. The inter-network abnormalities warrant examination in relation to illness-specific behaviors, namely abnormal eating behavior. This longitudinal study highlights the complexity of the neural underpinnings of AN.


| INTRODUC TI ON
Anorexia nervosa (AN) is a serious disorder with one of the highest mortality rates of any psychiatric illness (Arcelus, Mitchell, Wales, & Nielsen, 2011). It is characterized by fear of weight gain, preoccupation with body shape and weight, and severe restriction of food intake leading to significantly low body weight (American Psychiatric Association, 2013). AN most commonly emerges during adolescence (Herpertz-Dahlmann, 2015;Swanson, Crow, Grange, Swendsen, & Merikangas, 2011) and approximately half of affected individuals develop chronic illness (Lock et al., 2010). Preoccupations with food, weight and shape, coupled with ritualized eating, and compensatory behaviors are prominent features of AN (Heebink, Sunday, & Halmi, 1995;Sunday, Halmi, & Einhorn, 1995). Among adolescents with AN, obsessive-compulsive symptomatology has been associated with poor treatment outcome (Råstam, Gillberg, & Wentz, 2003) and eating-related obsessions have been shown to be a moderator of treatment outcome (Le Grange et al., 2012). These symptoms, together with high rates of comorbidity with obsessive-compulsive disorder (OCD; Anderluh, Tchanturia, Rabe-Hesketh, & Treasure, 2003;Bastiani et al., 1996;Bulik, Sullivan, Fear, & Joyce, 1997;Kaye, Bulik, Thornton, Barbarich, & Masters, 2004), increased co-occurrence of AN and OCD in family studies (Strober, Freeman, Lampert, & Diamond, 2007), and evidence of a high genetic correlation between the two disorders (Yilmaz et al., 2018) have led many in the field to consider whether AN and OCD share an underlying pathophysiology. This study probes these similarities by examining the neural underpinnings of obsessive-compulsive symptoms in AN.
Resting-state functional connectivity magnetic resonance imaging (rs-fcMRI) is a useful tool to study both specific neural circuits as well as large-scale neural networks implicated in psychiatric disease.
Rs-fcMRI measures spontaneous fluctuations in blood oxygen leveldependent (BOLD) signal. It is used to investigate the functional architecture of the brain at rest (Lee, Smyser, & Shimony, 2013). In AN, a number of rs-fcMRI studies have demonstrated evidence of disturbed functional connectivity in brain regions associated with cortico-striato-thalamo-cortical (CSTC) circuitry which have also been implicated in OCD (Biezonski, Cha, Steinglass, & Posner, 2016;Cha et al., 2016;Ehrlich et al., 2015;Haynos et al., 2018). Using networkbased statistics (NBS), Ehrlich et al identified reduced functional connectivity in a thalamo-insular subnetwork among underweight adolescents and young adults with AN relative to healthy control participants (HC) (Ehrlich et al., 2015). Another study measuring thalamo-frontal functional connectivity in acutely ill individuals with AN relative to HC found evidence of altered connectivity between the thalamus and both the dorsal and anterior prefrontal cortices which were associated with impairments in tasks measuring cognitive control and working memory, respectively (Biezonski et al., 2016). Haynos et al. (2018) used a seed-based approach to investigate ventral and dorsal frontostriatal circuits and found evidence of reduced functional connectivity in both circuits among underweight individuals with restricting type AN relative to HC. Cha et al. (2016) found increased left-sided nucleus accumbens (NAcc)-medial orbitofrontal cortex (mOFC) connectivity among underweight AN relative to HC.
Though both of these studies identified altered connectivity in anatomically similar circuitry, their findings diverged in the direction of the effect (hypo-vs. hyperconnectivity).
In OCD, there is strong evidence that disturbances in corticolimbic circuitry, specifically within the limbic CSTC loop, are involved in the disorder's pathophysiology (Harrison et al., 2013;Maia, Cooney, & Peterson, 2008;Saxena & Rauch, 2000;Vaghi et al., 2017). For example, a relationship between altered limbic CSTC functional connectivity and obsessive-compulsive symptomatology was supported by the finding that reduced left-sided NAcc-mOFC resting-state connectivity was associated with obsessive-compulsive symptom severity among unmedicated adults with OCD (Posner et al., 2014). Taken together, these results suggest that corticolimbic functional connectivity may be important to understanding the neurobiology underlying obsessive-compulsive-like symptoms in AN. It may be that, akin to OCD, limbic CSTC abnormalities underlie the obsessional thoughts and ritualized behavior characteristic of AN (Steinglass & Walsh, 2006).
Brain connectivity can be examined between specific neural regions, or more broadly through functional network connectivity (FNC) analysis, which measures the correlation of BOLD signal between large-scale resting-state networks. FNC is used to investigate network-level hypotheses (i.e., the functional architecture between networks). One such model, the triple network model of psychopathology, proposes that dysfunction within and between three resting-state networks associated with attention and cognitive control results in the development of psychopathology (Menon, 2011): the salience network (SN; nodes include the dorsal anterior cingulate cortex; dACC; and anterior insular cortices), the default mode network (DMN; nodes include the posterior cingulate cortex, medial prefrontal cortex, and lateral parietal lobules), and the executive control network (ECN; nodes include the dorsolateral prefrontal cortex; dlPFC; and lateral posterior parietal cortex). More specifically, the triple network model proposes that the SN, which is involved in detecting and filtering internal and external stimuli (Menon, 2011), is responsible for appropriately engaging the ECN to attend to task-related external stimuli (Seeley et al., 2007) and disengaging the DMN, which is activated during internally oriented, self-referential cogni- McFadden, Tregellas, Shott, & Frank, 2014) and individuals recovered from AN (Cowdrey, Filippini, Park, Smith, & McCabe, 2014).
Altered connectivity has also been identified within the SN in underweight and recovered participants with AN using ICA (McFadden et al., 2014) and a seed-based approach (Lee et al., 2014). Gaudio et al. (2015) identified decreased connectivity in the ECN in adolescents with new-onset illness, while others have found abnormal connectivity within the Frontoparietal Network (FPN; analogous to the ECN) in ill and recovered participants with AN (Boehm et al., 2014(Boehm et al., , 2016. On the other hand, several studies have not found significant differences in the networks implicated in the triple network model in underweight individuals with AN (Boehm et al., 2016;Scaife, Godier, Filippini, Harmer, & Park, 2017) or participants recovered from AN relative to HC (Boehm et al., 2014;Phillipou et al., 2016;Scaife et al., 2017). One study that used FNC to measure cross-network connectivity of these systems found no difference in network interactions between the SN, DMN, and FPN among women recovered from AN and HC (Boehm et al., 2016). Cross-network interactions between the SN, DMN, and ECN have not been studied in acutely ill individuals with AN.
In this longitudinal study, we used rs-fcMRI to measure functional connectivity in corticolimbic neural circuitry in adolescents and young adults with AN relative to age-matched healthy peers.
To better characterize neural mechanisms relevant to persistence of illness, participants were examined before and after weight restoration treatment. Building from existing data showing increased limbic CSTC connectivity among underweight adolescents and adults with AN  and evidence of an association between obsessive-compulsive symptoms and limbic CSTC functional connectivity in individuals with OCD (Posner et al., 2014), we hypothesized that NAcc-mOFC functional connectivity would (a) be increased among underweight AN relative to HC, (b) be associated with greater obsessive-compulsive symptoms in AN, and (c) partially normalize following weight restoration, with a concomitant reduction in obsessive-compulsive symptoms. We examined limbic CSTC functional connectivity bilaterally, though our a priori hypotheses were based on prior studies that found evidence of altered NAcc-OFC rsFC in the left hemisphere only Posner et al., 2014). Secondarily, because studies in OCD have associated obsessive-compulsive symptoms with altered interactions between the SN, DMN, and ECN, we explored broader network connectivity between these networks using FNC before and after weight restoration among AN relative to HC.

| Participants
Participants were individuals with AN and HC who presented to the Columbia Center for Eating Disorders (Table 1)

| Clinical assessments
Psychiatric diagnoses were established via the Semi-Structured Interview for DSM-IV (SCID) (Spitzer & Williams, 1988) and clinical interview. The diagnosis of AN was made according to DSM-5 crite-
Obsessive-compulsive symptoms were assessed in two ways: (a) the Obsessive-Compulsive Inventory-Revised (OCI-R) (Foa et al., 2002), which measures general obsessive-compulsive symptom severity and (b) the Yale-Brown-Cornell Eating Disorders Scale (YBC-EDS), which measures eating disorder-related rituals and preoccupations . The post-treatment YBC-EDS was administered by phone one week after hospital discharge. As an additional measure of internal preoccupation and repetitive thinking, participants completed the Ruminative Responses Scale, a 22-item self-report instrument (Treynor, Gonzalez, & Nolen-Hoeksema, 2003).
Height and weight were obtained on a beam balance scale (Detecto, Webb, MO) on the day of testing. Patients with AN were contacted during the month after discharge to assess weight trajectory. As weight maintenance in the first four weeks after weight restoration treatment is one of the few variables that predict longer term course (Kaplan et al., 2009), weekly weights for the first four weeks following discharge were obtained and verified by a clinician whenever possible (18 out of 23 participants). Weight slope was defined as the average change in weight per week over the first 28 days following hospital discharge and was calculated using all available weights (Kaplan et al., 2009).
Two 5-min resting-state scans (155 volumes) were obtained for each participant at each time point. The same imaging procedures were used for pre-and post-weight restoration scans.

| MRI analysis: seed-based connectivity
To test NAcc-OFC resting-state functional connectivity, we used a seed-based methodology similar to the approach by Cha et al. (2016). Specifically, standard image preprocessing was performed using SPM 12 and the CONN toolbox version 17.b for functional connectivity analysis. Slice timing and spatial realignment were applied.
Images were then coregistered with a high-resolution anatomical scan, normalized into the Montreal Neurological Institute space, and resampled at 2 mm. Images were smoothed with a Gaussian kernel of 8 mm FWHM (full width at half maximum). To minimize the influence of non-neural contributors to the fMRI signal, the BOLD time series was regressed against (a) five orthogonal time series extracted from white matter and cerebrospinal fluid separately using CompCor (component-based noise correction) methods (Behzadi, Restom, Liau, & Liu, 2007) and (b) 12 motion-related regressors (six estimated motion parameters plus their 1st-order derivatives from the rigid realignment preprocessing step in SPM). BOLD signal was also "scrubbed" by identifying and removing volumes (and the adjacent, neighboring volumes) with framewise displacement (FD) > 0.5 mm or global signal intensity changes > 3 SD estimated using the artifact detection tool (ART) (www.nitrc.org/projects/artifact_detect).
Although no runs met this criterion, a threshold of >20% scrubbed volumes was used to exclude runs from subsequent analyses. Other signal processing in the pipeline included temporal band-pass filtering (0.008-0.09 Hz) and linear detrending. Head motion was comparable across groups (Supporting information Figure S1).
Seed-based correlation analysis was used to investigate connectivity between the NAcc and OFC bilaterally. The NAcc seed was predefined from FSL Harvard-Oxford Atlas maximum-likelihood subcortical atlas (HarvardOxford-sub-maxprob-thr25-2 mm.nii).
Connectivity maps were generated by calculating the Pearson correlation between the time series within the NAcc seed and the time series for all other voxels in the brain; Pearson correlations were then Fisher r-to-z transformed to allow for group-wise comparisons.
For hypothesis testing, mean connection strength between the NAcc and the mOFC, as defined by a published atlas (http://www. gin.cnrs.fr/en/tools/aal-aal2/), was extracted. Whole-brain connectivity maps were used in exploratory analyses.

Independent components analysis + FIX was conducted by
randomly selecting 20 study participants for group ICA, and then manually identifying components as either "good" (i.e., indexing resting networks) or "bad/noise" (i.e., indexing noise/artifacts) following published procedures (Griffanti et al., 2017). A classifier was then trained using FIX to automatically identify and regress out noise components for the remaining resting fMRI datasets, yielding a denoised resting fMRI dataset for subsequent analyses. Lastly, to determine the effectiveness of ICA-FIX in denoising this rs-fcMRI dataset, head motion parameters derived from this dataset after ICA-FIX denoising were compared to another denoising technique (Supporting information Figure S2). Group ICA was applied to the denoised rs-fcMRI data.

| Statistical analyses
Demographic and clinical characteristics were compared between diagnostic groups (AN vs. HC) for continuous variables (age, BMI, estimated IQ, STAI-T, and BDI) using independent samples t tests.
Eating disorder symptoms, mood, anxiety, and obsessive-compulsive symptoms before and after weight restoration were examined

| Confounding variables
In each statistical analysis, we tested a full model with nuisance covariates, including demeaned age and estimated IQ. As standardized norms for the WTAR are not available for the 6 individuals under the age of 16, demeaned WTAR raw scores were used as the covariate. In a sensitivity analysis, the six participants under the age of 16 years were removed from the neuroimaging analyses, which did not change the pattern of results. We performed an additional GLM (Bonferroni corrected for multiple comparisons) to measure the effect of AN subtype (AN-R vs. AN-BP) on connectivity differences, with connectivity strength as the dependent variable, Group (AN-R, AN-BP, HC) as the between-subject variable, Time (pre-and post-weight restoration) as the withinsubjects variable, and demeaned age and WTAR raw score as covariates.

| Clinical correlates
Partial correlations were performed to measure associations between fMRI results and measures of obsessive-compulsive symptomatology and weight slope, controlling for age and estimated IQ (demeaned). The significance and confidence level was set at α = 0.05 (95% confidence interval).    Table S1). Whole-brain exploratory analyses yielded nonsignificant group effects.

| MRI results: functional network connectivity (FNC)
There was reduced cross-network functional connectivity between the SN-left ECN in the AN group relative to HC (Group:  Table S2).

| D ISCUSS I ON
This study used rs-fcMRI to examine neural substrates of obsessive-compulsive symptomatology longitudinally in individuals with trait (Smith & Alloy, 2009). Rumination has rarely been examined in individuals with AN (Fürtjes et al., 2018;Rawal, Park, & Williams, 2010;Smith, Mason, & Lavender, 2018;Startup et al., 2013), and only one other study that we are aware of has examined rumination longitudinally in AN (Fürtjes et al., 2018). In depressive disorders, rumination has been associated with both the onset and maintenance of illness (Nolen-Hoeksema, 2000;Nolen-Hoeksema, Larson, & Grayson, 1999;Nolen-Hoeksema, Morrow, & Fredrickson, 1993;Watkins, 2008) and has been considered as a target in the treatment of chronic depression (Watkins et al., 2011). Further research is warranted to investigate whether a ruminative thinking style is related to the persistence of AN.   (Biezonski et al., 2016;Favaro et al., 2013;Phillipou et al., 2016) and other studies only included AN-R (Haynos et al., 2018). One study had unbalanced numbers in each subtype group which limited between-group comparisons (Boehm et al., 2016). Cha et al included both AN-R and AN-BP and found no effect of subtype on imaging results . Our patient sample was slightly larger, which may have allowed us to detect a group difference that was not evident with a smaller group.
This is the third study that we are aware of to find NAcc-prefrontal cortex rs-fcMRI differences between individuals with AN and HC.
The direction of effect differs across studies. Our patient sample was unmedicated and included individuals with both AN subtypes, whereas the study by Haynos et al. (2018) permitted psychotropic medications and studied individuals with AN-R only. As frontal activation has been shown to change with psychotropics (Posner et al., 2011), medication status may influence frontostriatal circuit connectivity results. However, while the discrepancy in our results is likely due in part to these methodological differences, the divergence in our findings also mirrors results from the broader fMRI literature in AN in which some studies have found evidence of increased neural activity within the frontostriatal system (Cowdrey, Park, Harmer, & McCabe, 2011;Frank et al., 2012) while others have found the opposite (Brooks et al., 2011;Scaife et al., 2017;Wagner et al., 2007).
Moving forward, it will be important to discern whether the heterogeneity in results stems from variability in methodology or reflects an underlying complexity of frontostriatal circuitry in AN. This study has notable strengths. First, the patients with AN were not taking psychotropic medication. The full effects of psychotropic medications on the brain, and especially on functional connectivity, are unknown. However, as functional connectivity within large neural circuits has been shown to change with psychotropics (Posner et al., 2011), medication status may have a substantial influence upon frontostriatal circuit connectivity results.
Second, fMRI testing occurred at two time points, before and after weight restoration in the same participants. As the effect of stage of illness on brain architecture and connectivity is not fully established, this longitudinal analysis provides information about the stability of neural network changes in AN. As there may be important differences in the neurobiology of individuals who recover, a longitudinal approach can provide different information than cross-sectional studies. A potential limitation of this study is that estimated IQ scores were not available for individuals under 16 years old. However, a sensitivity analysis of our imaging results in which we removed the six participants under the age of 16 did not meaningfully alter results. Studying adolescents also raises unique challenges due to the influence of hormones on brain development. This study did not include assessments of puberty, which may be useful covariates in future studies of adolescents with AN. Imaging-related limitations include the possibility that participants fell asleep during resting-state scans, despite being instructed before beginning fMRI procedures to rest with their eyes closed but to remain awake, and the difficulty in scanning the OFC given this region's vulnerability to OFC signal dropout.
Finally, the FNC analysis was exploratory and should be considered as useful in generating new hypotheses about the relationship of altered triple network FNC to the underlying neurobiology of AN.
This study found evidence of increased left NAcc-left mOFC connectivity among underweight AN that was not observed following weight restoration. Unlike in OCD, we did not find an association between altered NAcc-OFC connectivity and obsessive-compulsive symptoms in AN. Our exploratory investigation of the triple network model revealed evidence of disturbed connectivity in AN across large-scale neural networks which persisted following weight restoration. Taken together, these results suggest that while there is clear value in probing the neurobiology of illness-specific phenomena to understand the neurobiology of AN, advances in psychiatry more broadly may come through probing complex inter-network relationships.

ACK N OWLED G M ENTS
The authors thank Xingtao Zhou and Jiook Cha for their thought-

ful contributions and the patients and staff of the Eating Disorders
Research Unit at the New York State Psychiatric Institute. Joanna Steinglass and Jonathan Posner were coprincipal investigators of the grant that supported this work (R21-MH099388) and contributed equally to this project.

CO N FLI C T O F I NTE R E S T
Dr. Steinglass receives royalties from UpToDate.