Association Between Ghrelin and Body Weight Trajectory in Individuals With Anorexia Nervosa

Key Points Question Is there an association between circulating levels of the orexigenic hormone ghrelin and body weight trajectories in individuals with anorexia nervosa? Findings In this 18-month cohort study of 68 girls and young women, elevated baseline ghrelin levels were associated with prospective weight gain in anorexia nervosa. Meaning This study offers evidence of association between ghrelin and longitudinal weight outcomes in individuals with anorexia nervosa; further studies are warranted to confirm this association and evaluate the potential clinical utility of ghrelin in anorexia nervosa.

Parent Study Overview Findings in this manuscript are derived from data obtained in a larger prospective cohort study aimed at examining multidimensional aspects of abnormal homeostatic and hedonic food motivation pathways including hormones, neural circuits, and psychological symptoms as predictors of eating disorder trajectories over 18 months among girls and young women with low-weight eating disorders, including anorexia nervosa (AN), other specified feeding and eating disorders-atypical AN (OSFED), avoidant restrictive food intake disorder (ARFID), compared to healthy controls with no lifetime history of eating disorders (HC).

1.2.
Parent Study Hypotheses Hypothesis 1: Altered food motivation pathways (ie, hormones, neural circuits, and psychological symptoms) underly patterns of restriction, binge eating, and purging, and differentiates adolescents with low-weight eating disorders from healthy, normal-weight controls. Hypothesis 2*: Within adolescents with low weight eating disorders followed longitudinally, 18-month outcome (ie, dietary restriction, binge/purge frequency, and weight changes) is determined by food motivation pathways (ie, hormones, neural circuits, and psychological symptoms) at baseline and changes over time. * Drs. Eddy, Misra, Lawson, and Kim decided to focus on ghrelin at baseline as independent variable of interest and weight changes in AN vs. HC as the primary outcome of interest, a priori to data analysis, to allow for a focused analysis avoiding overlaps with ongoing projects in our team.

1.3.
Participant Selection Criteria Females with low-weight eating disorders (LWED) Inclusion criteria 1. Female age 10 to 22 years old 2. LWED characterized by (i) low body weight defined as less than 90% of median BMI for age and sex, or less than 90% of ideal body weight for height (IBW, defined as the weight corresponding to the participant's height percentile for age/bone age), with (ii) restrictive eating, binge eating more than once a month, purging more than once a month, excessive exercise more than once a month, and/or in treatment for an eating disorder, to be determined as part of the initial clinical assessment.

Exclusion criteria
1. Use of systemic hormones, pregnancy or breastfeeding within eight weeks 2. Use of Depo-Provera within three months 3. History of psychosis by Schedule for Affective Disorders and Schizophrenia for School Age Children-Present and Lifetime Version (KSADS-PL) 4. Substance or alcohol use disorder active within the past month by KSADS-PL 5. Diabetes mellitus 6. Hematocrit (Hct) < 30.0% 7. Potassium (K) < 3.0 mmol/L 8. Gastrointestinal tract surgery (including gastrectomy, gastric bypass surgery, and small or large bowel resection) 9. Other medical explanation for low weight (e.g., brain tumor) 10. Active suicidal ideation

Reasons for Screen Failures
Reasons for screen failures for individuals screened as low-weight eating disorder (LWED) or healthy controls (HC) included: • Not meeting BMI criteria for 5 LWED and 3 HC.
• Not meeting the eating habits criteria for 7 LWED and 2 HC.
• Related to psychiatric symptoms, with one individual with LWED who was found to have active substance use disorder and 13 screened as HC found to have active or lifetime history of psychiatric illnesses • Other reasons excluded 4 LWED and 2 HC. Individuals screened out may not have gotten a full diagnostic classification with regards to the presence or absence of AN and LWED term is used above.

3.2.
Descriptive Data Data distribution was assessed by the Shapiro-Wilk test and non-parametric data from AN and HC were compared using the Wilcoxon rank sum test. Group differences in medians were computed with the Hodges-Lehmann estimation are shown with non-parametric 95% confidence interval, computed as the median of the set of differences between each value in AN and each value in HC. Contingency tables were analyzed by chi-squared test with Yates continuity correction for goodness of fit. Two-tailed P<0.05 was considered significant.

3.3.
Primary Analysis The main hypothesis was tested using linear mixed-effects regression model (LMM) estimating longitudinal weight change index as a factor of baseline visit ghrelin AUC through the lmer function from lme4 library in R. Repeated-measures data of fold change in BMI percentiles by either 9 or 18 months as response/outcome of interest was fit using LMM with fixed effects for slopes and intercepts and random effects of the individuals as random intercepts, with the term (1 | id), where id represents study id for the participants. Visit interval and outcome of interest were scaled without centering.

3.4.
Secondary Analyses Sensitivity analyses assessed the robustness of primary analysis testing a priori hypothesis. Multivariate LMMs with each of the below as the single change to the main LMM were tested: • Removal of outliers beyond 2 standard deviation units of the mean • Individual ghrelin measurements as alternatives to the main independent variable of interest (ie, ghrelin AUC) • Fasting ghrelin • Postprandial ghrelin 0.5 h after meal initiation • Postprandial ghrelin 1 h after meal initiation • Postprandial ghrelin 2 h after meal initiation • Alternatives to the main outcome variable of interest (ie, weight change index defined as fold change in BMI percentile from baseline to follow-up visit) • BMI z-score delta change, calculated as: (BMI z-score at follow-up -BMI percentile at baseline) • BMI percentile delta fold change, calculated by dividing the BMI percentile delta by baseline value: ((BMI percentile at follow-up -BMI percentile at baseline) / BMI percentile at baseline) • BMI percentile log fold change, calculated by: (log10(BMI percentile at follow-up / BMI percentile at baseline)) • Alternative sets of covariates, replacing the main model's covariate terms with those not included in the main model to account for the non-independence of variables (ie, Tanner staging and age, race and ethnicity, presence or absence of AN diagnosis and BMI percentile) • Tanner stages by breast development instead of age, dichotomized at or below stage 5 • Tanner stages by genital development instead of age, dichotomized at or below stage 5 • Ethnicity instead of race • Ethnicity with race as nonoverlapping categories as the only individuals identifying as Hispanic were white (ie, White Hispanic, White Not Hispanic, Asian Not Hispanic, and Other Not Hispanic) • Baseline BMI percentiles instead of diagnostic group • Alternate model assumptions regarding variance estimators, to account for the possibility that the main LMM solved using maximum likelihood (ML) principle may under or overestimate the true variance, final model was also tested with the restricted or residual maximum likelihood (REML) method.
Alternative sets of random effect terms were assessed with stepwise deletions of single terms, including: • Intercept varying among individuals and in 9-or 18-month follow-up visits • Random intercept and slope by different duration of follow-up and the individual • Random intercept of the individuals and random slope varying by different duration of the follow-up for 9-or 18-month follow-up visits In addition, weight change outcome at either follow-up was modeled individually using simple linear regression models (LM) without repeated-measures outcome as in LMM approach, with adjustment for the same covariate terms of the main model. Given the exploratory nature of LMs, we expanded this to build LMs with each of the ghrelin measurements in addition to the AUC index.
Sensitivity of deriving odds ratios (ORs) directly from the linear model without dichotomizing the continuous outcome variable was assessed by dichotomizing the weight change index into binary outcomes of presence or absence of weight gain at or above specified percentage, to derive ORs of generalized linear mixed effects logistic regression models (GLMMs).
Lastly, subgroup analysis was performed as described in the main text. Given the exploratory nature of LMMs built for each subgroup, we expanded this approach by building LMMs with each of the ghrelin measurements in addition to the AUC index.

3.5.
Handling of Outliers Outliers of weight change outcome were identified by filtering for values beyond 2 standard deviation units of the mean. From the dataset with 124 datapoints, 6 outlier weight change indexes (<5%) from 5 individuals with anorexia nervosa were identified. For one individual, data from both visits were identified as outliers; for 3 individuals, only one of the follow-up visit data was identified as outliers; and for one individual who only had a single follow-up data, this was identified as an outlier, resulting in the final n=33 in this dataset. No healthy control participants had weight change values in the extremes (n=33 for HC). The dataset without outliers, with 118 datapoints from N=66 individuals, was used in LMs for each follow-up visit, LMMs for subgroup analyses, and LMM for sensitivity testing for outlier removal. eFigure 1. Association Between Baseline Ghrelin and Prospective Change in Body Weight Without Adjustment for Covariates.
Abbreviations: AN, participants with anorexia nervosa; HC, healthy control participants; AUC, area under the curve.
Ghrelin AUC data from N=68 participants is plotted with unadjusted Weight Change Index values at respective follow-up visit grouped by diagnosis and with y-axis in exponential scale.

eFigure 2. Sensitivity Analysis for Deriving Odds Ratios from Linear Regression Models
Sensitivity of deriving odds ratios (ORs) directly from linear models without dichotomizing the continuous outcome was assessed by exploratory analysis of ORs derived with dichotomization of the outcome into binary groups to build univariate logistic regression models, or generalized linear mixed effects models (GLMM). Binary outcome groups were the presence or absence of weight gain at or above the specified cutoff % weight gain, the ORs of which were estimated to test the association between ghrelin AUC and future weight gain. For a total of 21 GLMMs with cutoff points ranging from 0 to 20% increases in body weight, univariate GLMM ORs of gaining 7 or higher % weight (ie, cutoff values 7 through 20) were positively associated with ghrelin AUC. After accounting for multiple comparisons with Benjamini-Hochberg method with a false discovery rate of .038, cutoff values 8 through 20 resulted in statistically significant associations. Figure shows the model results with cutoff % weight gain ranging from 0 to 10% and unadjusted and adjusted significance thresholds are indicated with black and grey dashed lines, respectively. eFigure 3. Expanded Subgroup Analyses of Baseline Ghrelin and Prospective Change in Body Weight