Association of Brain Reward Response With Body Mass Index and Ventral Striatal-Hypothalamic Circuitry Among Young Women With Eating Disorders

Key Points Question Is brain reward response associated with specific behaviors across the eating disorder diagnostic spectrum? Findings In this cross-sectional functional brain imaging study of 197 women with anorexia nervosa, other specified feeding and eating disorders, bulimia nervosa, and binge eating disorder and a matched cohort of 120 healthy controls, brain salience response was significantly inversely correlated with body mass index and binge eating severity and positively correlated with ventral striatal-hypothalamic circuitry. Meaning Results of this study suggest that eating disorder behaviors change brain reward processing, which may alter food intake control circuitry and reinforce the individual’s eating disorder behavior.

ating disorders are severe psychiatric disorders with high mortality. 1 Anorexia nervosa (AN) is characterized by severe underweight with intermittent binge eating or purging episodes; individuals with bulimia nervosa (BN) are at normal to high weight and regularly binge and purge. Binge-eating disorder (BED) is associated with binge-eating episodes and frequently elevated body weight. 2 Eating disorders that do not meet full criteria for those diagnoses have been recognized as specific subgroups within the other specified feeding and eating disorders (OSFED) category of the DSM-5. Food restriction, episodic binge eating, or purging vary across the diagnostic groups, whereas body dissatisfaction and drive for thinness are typically elevated across all eating disorders, as are anxious traits and sensitivity to salient stimuli. Identifying how those behaviors are associated with particular biologic mechanisms could help create a better understanding of the underlying eating disorder pathophysiologic factors and development of specific treatments. 3 To adopt a dimensional conceptualization of eating disorder specific behaviors and neurobiologic factors, we recruited individuals across the eating disorder spectrum and applied the prediction error construct from the National Institute of Mental Health Research Domain Criteria (RDoC) project. 4 Brain reward circuits have been repeatedly implicated in eating disorders, and altered reward learning may play a particularly important role. 5 In reward learning, the difference between an expectation and outcome yields a prediction error, a dopamine-assoc iated signal that reinforces new associations. 6,7 The direction of the prediction error is indicated by its sign, which indicates a better (positive) or worse (negative) outcome than expected. The absolute value reflects the degree of deviation of the outcome from the expectation and is related to surprise or conceptualized as a motivational salience signal. 8,9 The dopamine system adapts in opposite directions to extremes of food intake. [10][11][12][13] Food restriction enhances dopamine circuit activity 14,15 and excessive food intake downregulates dopamine circuit activity, 12 which could be relevant for eating disorder pathophysiologic factors. [16][17][18][19] Studies in AN found elevated prediction error response to taste and monetary stimuli compared with healthy controls but a lower response in small studies in individuals with BN and individuals with overweight. [20][21][22][23] Those studies suggested that the prediction error signal is inversely correlated with eating disorder behaviors from restrictive to loss of control food intake (binge eating). 24 Furthermore, prediction error response was positively correlated in adolescent AN with ventral striatum-hypothalamus directed effective connectivity, a circuitry that has been associated with food intake control. 23 For consistency with the RDoC approach, we studied a large group of individuals with eating disorders, varying on a spectrum of restrictive undereating to loss-of-control overeating. To validate previous results, we also recruited a healthy control group. First, we hypothesized that we would find inverse correlations between prediction error response and eating disorder behavior from undereating to overeating, as reflected in body mass index (BMI) and binge eating severity. This hypoth-esis would support basic and translational science research by externally validating a core behavioral dimension via its associations with reward-responsiveness. Second, we hypothesized that effective connectivity would be directed from the ventral striatum to the hypothalamus in the eating disorder sample. This hypothesis would support a potential trait mechanism across eating disorders to attempt to control eating drive. 23 Third, we hypothesized that associations between biological and behavioral data may help develop a model to explain how traits, eating disorder behaviors, and neurobiologic factors interact and reinforce the often chronic nature of eating disorders. 25

Participants
The Colorado Multiple institutional review board approved the study. All participants provided written informed consent. We recruited 197 women with an eating disorder: 69 AN restricting subtype, 22 AN binge-eating/purging subtype, 17 OSFED atypical AN subtype, 17 OSFED purging disorder subtype, 56 BN, 3 OSFED binge-eating subtype, and 13 binge eating disorder (BED). Participants with eating disorders were recruited from eating disorder partial hospitalization specialty care (EDCare Denver or Children's Hospital Colorado) within the first 2 weeks of treatment, to mitigate effects of acute starvation or dehydration. 26 Following RDoC instructions, we recruited any interested patient with eating disorders who was admitted to treatment. In addition, we recruited 120 women as healthy controls (HCs) through local advertisements. The study was conducted from June 2014 to November 2019. Data were analyzed from December 2019 to February 2020. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for crosssectional studies.
Participants were right-handed without history of head trauma, neurological disease, major medical illness, bipolar disorder, psychosis, or current (past 3 months) substance use disorder. The healthy controls were studied during the first 10 days

Key Points
Question Is brain reward response associated with specific behaviors across the eating disorder diagnostic spectrum? Findings In this cross-sectional functional brain imaging study of 197 women with anorexia nervosa, other specified feeding and eating disorders, bulimia nervosa, and binge eating disorder and a matched cohort of 120 healthy controls, brain salience response was significantly inversely correlated with body mass index and binge eating severity and positively correlated with ventral striatal-hypothalamic circuitry.
Meaning Results of this study suggest that eating disorder behaviors change brain reward processing, which may alter food intake control circuitry and reinforce the individual's eating disorder behavior.
of the menstrual cycle to reduce potential hormonal variations. For eating disorders, treatment stage was the primary variable we controlled for, but we recorded days from last menstrual cycle as a proxy to test for hormonal variation.

Assessments
Psychiatric diagnoses were assessed using the Structured Clinical Interview for DSM-5 by a doctoral-level interviewer. 27 Participants completed the Eating Disorder Inventory-3 (EDI-3) for drive for thinness (intense fear of weight gain), bulimia (tendency to engage in binge eating), body dissatisfaction (discontentment with size of body regions), 28 Revised Sensitivity to Punishment and Reward Questionnaire, 29 State-Trait Anxiety Inventory, 30 Temperament and Character Inventory for Novelty Seeking and Harm Avoidance, 31 and Beck Depression Inventory-II 32 ; participants blindly rated sugar solutions for sweetness and pleasantness using a 9-point Likert scale. A subset of participants (eating disorder, n = 128, HC, n = 84) completed the Eating Expectancy Inventory for eating leads to feeling out of control. 33

Brain Imaging Methods
Functional Magnetic Resonance Imaging Between 7:00 AM and 9:00 AM on the study day, participants with eating disorders ate their meal-plan breakfast and HC ate a quality-matched and calorie-matched breakfast ( Table 1). Brain imaging was performed between 8:00 AM and 9:00 AM using the 3-T Signa (General Electric Company) or Skyra 3-T scanner (Siemens) (eMethods 1 in the Supplement). A scanner covariate was included in the multivariate analysis of covariance model for imaging group contrasts.

Taste Reward Task
The design of this study was adapted from O'Doherty et al 34 (eMethods 2 and eFigure 2 in the Supplement). Participants learned to associate 3 unconditioned taste stimuli (1 molar sucrose solution, no solution, or artificial saliva) with paired conditioned visual stimuli. Each conditioned visual stimulus was probabilistically associated with its unconditioned taste stimulus such that 20% of sucrose and no solution conditioned visual stimuli trials were unexpectedly followed by no solution or sucrose unconditioned taste stimuli, respectively.

Functional Magnetic Resonance Imaging Analysis
Image preprocessing and analysis were performed using Statistical Parametric Mapping, version 12 35 (Wellcome Trust Centre for Neuroimaging). Images were realigned to the first volume, normalized to the Montreal Neurological Institute template, smoothed at 6-mm full width at half maximum gaussian kernel. Data were preprocessed with slice time correction and modeled with a hemodynamic response convolved function using the general linear model, including temporal and dispersion derivatives. A 128-second high-pass filter (removing low-frequency blood oxygen level dependent signal fluctuations), motion parameters (as first-level analysis regressors), and the SPM FAST (prewhitening attenuation of autocorrelation effects) were applied. 36

Prediction Error Analysis
Each participant's prediction error signal was modeled based on trial sequence (absolute of positive and negative prediction error) and regressed with brain activation across all trials 20,21,34 (eMethods 3 in the Supplement). We extracted mean parameter estimates across all voxels from 5 predefined anatomical regions of interest (ROIs) bilaterally, based on ROIs that differentiated groups previously 23 : bilateral dorsal anterior insula (automated anatomical labeling Atlas 37 ), ventral anterior insula, 37 caudate head, 37 ventral striatum, 38 and nucleus accumbens. 39 Effective Connectivity Analysis We extracted ROI functional activation for trials of expected receipt of 1 molar sucrose solution (n = 80). 40 The Tetrad-V 41 was used to infer effective connectivity with independent multisample greedy equivalence search and linear non-gaussian orientation, fixed structure search algorithms. We extracted edge coefficients for ventral striatum-hypothalamus (hypothalamus ROI, SPM12 WFU_PickAtlas extension 42 ) connectivity to test for correlations with behavior or PE values based on our previous studies 23 (eMethods 4 in the Supplement).

Statistical Analysis
Statistical analysis was performed with SPSS 27 software (IBM). Data were tested for normality with Shapiro-Wilk test and ranked and normalized using Rankit procedure if nonnormally distributed. 43 Demographic and behavior data were analyzed using analysis of variance, and post hoc analyses were Bonferroni corrected. Multivariate analysis of covariance and correlation analyses were used to test effect sizes of potential confounding categorical or continuous variables such as comorbidity, medication use, or age. Variables associated with the primary outcome variable brain response were included in a group-comparison multivariate analysis of covariance and estimated marginal means post hoc comparisons Bonferroni corrected. Partial η2 was calculated for effect size in addition to power calculations. Pearson correlation analysis was used to test associations between behavior and brain activation, CIs were calculated using bootstrap (1000 samples) and results were multiple comparisons controlled using false discovery rate. 44 All P values were 2-tailed, and a P value less than .05 was considered statistically significant.
Forty-five HC and 40 participants with eating disorders took oral contraceptives (χ 2 = 16.329; P < .001). Use of antidepressant or antipsychotic medication, or comorbidity with major depression, obsessive-compulsive disorder, or posttraumatic stress disorder were not differentially distributed between eating disorder groups, but comorbid anxiety disorder was (χ 2 = 7.935; P = .047).

Brain Response-Behavior Correlations
In the HC group, correlations between age, BMI, or behavior and brain imaging values were not significant or were not found after multiple comparison correction.
The  Table 2; eFigure 3 in the Supplement). In a partial correlation analysis, significant correlations between regional prediction error response and BMI were found when controlling for binge-eating frequency (

Effective Connectivity
Effective connectivity was directed in HC from hypothalamus to ventral striatum. In the eating disorder sample, effective connectivity was directed from ventral striatum to hypothalamus ( Figure 1). eFigure 4 in the Supplement shows individual graphs for AN and BN groups. Extracted effective connectivity edge coefficients from right ventral striatum to hypothalamus in eating disorder correlated significantly with right-sided ventral striatum prediction error response (r = 0.189; 95% CI, 0.045-0.324; P = .01); left-sided correlation was also positive but nonsignificant (r = 0.104; 95% CI -0.030 to 95% CI 0.231; P = .15). Edge coefficients correlated in eating disorders significantly in 3 ways: first, bilaterally negatively with eating leads to feeling out of

Prediction Error Group Contrasts
Prediction error response significantly differentiated groups (Wilks λ, 0.843; P = .001). After Bonferroni correction, prediction error remained elevated in AN compared with HC, OSFEDr and BN in the left caudate head, compared with HC and BN and BED in left nucleus accumbens, compared with HC and BN in the right nucleus accumbens, compared with HC and BED in the left ventral striatum, and compared with BN in the left dorsal insula ( Table 3; eFigure 5 in the Supplement).

Discussion
This cross-sectional study in a large sample of women across the eating disorder diagnostic spectrum indicates elevated prediction error response in AN compared with HC, BN, and BED, which is consistent with previous studies. In eating disorders, prediction error response was inversely correlated with BMI and binge eating behaviors. Furthermore, ventral striatal prediction error response correlated with effective connectivity from the ventral striatum to the hypothalamus in eating disorders, indicating an association between prediction error responsiveness and strength of a circuitry that has been associated with food intake control. 45 The results support basic science studies showing that prediction error response adapts to patterns of food intake. [10][11][12][13] Regional prediction error response was higher the more restrictive a person's food intake was, reflected by BMI. Binge eating frequency and the EDI-3 bulimia score were also inversely correlated but not after controlling for BMI.
Previously, prediction error was associated with harm avoidance in adolescent AN, which we did not find in adults across the eating disorder spectrum. 23 It is possible that at an earlier age prediction error response affects anxiety and triggers eating disorder behaviors. 46 However, during longer illness, such associations may be attenuated, after eating disorder behaviors have transformed to become a way of maintaining a sense of control. 47 Ventral striatal-hypothalamic effective connectivity during sugar tasting in opposite directions between HC and eating disorder groups, together with positive correlation with ventral striatum prediction error response, extends previous results in smaller eating disorder samples. 23,40 Dopamine and prediction error signaling have been associated with energy homeostasis regulation, 48 and a fear mediated dopamine circuit from the ventral striatum to the hypothalamus has been identified that inhibits food intake. 45,49,50 A fear-driven and dopamine-mediated circuitry to suppress eating drive thus could be a trait common to all eating disorders, which is most effective in the context of sensitized dopamine circuits, reflected in high prediction error response.
The results suggest that a data-driven model of biological and behavioral interactions that promote restrictive or excessive eating behaviors is warranted ( Figure 2). Consistent with previous research, trait anxiety correlated positively with the EDI-3 bulimia score, supporting that negative emotions trigger binge eating. 51 Novelty seeking and bulimia are known risk traits for elevated BMI. 52,53 The negative correlation between BMI and prediction error response supports basic science, indicating a modulatory effect of amount of food intake on dopaminergic circuitry. 11,12 The positive correlation between prediction error and effective connectivity further implicates dopamine circuitry in modulating brain connectivity 54 and suggests that higher dopaminergic activity strengthens the ventral striatal-hypothalamic food-control circuitry enabling individuals with AN to override normal hunger cues. BN or BED, however, have lower dopaminergic activity than AN, cannot maintain consistent food intake control, which facilitates intermittent binge-eating episodes. 55,56 The negative correlation between effective connectivity and eating leads to feeling out of control could indicate that stronger food intake control circuitry leads to less of an out-of-control sensation, but this is speculative and requires further study. Behaviorally, higher BMI was associated with higher body dissatisfaction, which correlated positively with intolerance of uncertainty, harm avoidance, drive for thinness, depression and sensitivity to punishment, which also correlated positively with Trait Anxiety, consistent with previous research. 57-59 Body dissatisfaction triggers drive for thinness, which reinforces and is reinforced by anxiety, depression, and punishment sensitivity, increasing poor self-esteem and further promoting eating disorder behaviors.

Limitations
This study has limitations. The study was well powered for group comparisons, but effect sizes were small to moderate. Correlation analyses in the eating disorder sample showed moderate to large or very large effect sizes, but correlation analyses cannot prove mechanism. The prediction error model is based on dopamine function, but other neurotransmitter systems, such as serotonin, noradrenaline, or adenosine, are likely factors in reward processing and behavior control in eating disorder behaviors. 60-62 Furthermore, dopamine neuronal function was not directly measured in this study and functional magnetic resonance imaging prediction error response is only an indirect approximation. 63 Whether altered prediction error response affects food intake acutely will require further study, and inverse relationships between this brain response and BMI may also exist in other conditions. The hypothalamus ROI did not separate subnuclei. While HC were studied during the first 10 days of the menstrual cycle to keep hormonal variation low, the eating disorder population was either amenorrheic or had more days from the last menstrual cycle. Not having hormonal measures is a limitation, but days from last menstrual cycle did not correlate with brain response in either group. Because eating disorder results were either higher or lower compared with HC, we do not believe that there was a systemic confound. For the prediction error analysis, we used the unsigned (absolute) prediction error. Pleasantness ratings for the 1 molar sucrose solution varied from very high to very low. Unexpectedly receiving sucrose solution could therefore be associated with positive (better than expected) or negative (worse than expected) prediction error. Studying the absolute prediction error accounts for interindividual variation and measures degree of deviation from expectation, reducing effects of subjective pleasantness. 64,65 Our theoretical framework was primarily based on sensitivity to salient stimuli and adaptation of the related circuitry to food intake. We believe that using the unsigned prediction error yields more reliable results, independent from individual value computation.

Conclusions
Results of this study suggest that behavioral traits are factors in eating disorder initiation and extremes of eating and then alter prediction error-related reward response. This process reinforces in opposite ways the ventral striatal-hypothalamic food control circuitry, which is activated in response to sugar taste as a trait in eating disorders. Clinically it therefore may be important to implement weight gain in eating disorders in people with underweight and weight loss in eating disorders associated with overweight to normalize brain function and behavior. This topic is controversial, though, and the critical question remains what the best BMI for a person is in this context. Furthermore, temperamental traits are biologically oriented behaviors that affect eating disorder behaviors. Treatment modules that specifically target those behaviors may be a key element to promote behavior change and lasting recovery.