Body composition in anorexia nervosa: Meta‐analysis and meta‐regression of cross‐sectional and longitudinal studies

Abstract Objective Clinically, anorexia nervosa (AN) presents with altered body composition. We quantified these alterations and evaluated their relationships with metabolites and hormones in patients with AN longitudinally. Method In accordance with PRISMA guidelines, we conducted 94 meta‐analyses on 62 samples published during 1996–2019, comparing up to 2,319 pretreatment, posttreatment, and weight‐recovered female patients with AN with up to 1,879 controls. Primary outcomes were fat mass, fat‐free mass, body fat percentage, and their regional distribution. Secondary outcomes were bone mineral density, metabolites, and hormones. Meta‐regressions examined relationships among those measures and moderators. Results Pretreatment female patients with AN evidenced 50% lower fat mass (mean difference [MD]: −8.80 kg, 95% CI: −9.81, −7.79, Q = 1.01 × 10−63) and 4.98 kg (95% CI: −5.85, −4.12, Q = 1.99 × 10−28) lower fat‐free mass, with fat mass preferentially stored in the trunk region during early weight restoration (4.2%, 95% CI: −2.1, −6.2, Q = 2.30 × 10−4). While the majority of traits returned to levels seen in healthy controls after weight restoration, fat‐free mass (MD: −1.27 kg, 95% CI: −1.79, −0.75, Q = 5.49 × 10−6) and bone mineral density (MD: −0.10 kg, 95% CI: −0.18, −0.03, Q = 0.01) remained significantly altered. Discussion Body composition is markedly altered in AN, warranting research into these phenotypes as clinical risk or relapse predictors. Notably, the long‐term altered levels of fat‐free mass and bone mineral density suggest that these parameters should be investigated as potential AN trait markers. ResumenObjetivo Clínicamente, la anorexia nervosa (AN) se presenta con alteraciones en la composición corporal. Cuantificamos estas alteraciones y evaluamos longitudinalmente su relación con metabolitos y hormonas en pacientes con AN. Método De acuerdo con las pautas PRISMA, realizamos 94 meta‐análisis en 62 muestras publicadas entre 1996–2019, comparando hasta 2,319 pacientes mujeres en pre‐tratamiento, post‐tratamiento, y recuperadas en base al peso con hasta 1,879 controles. Las principales medidas fueron masa grasa, masa libre de grasa, porcentaje de grasa corporal y su distribución regional. Las medidas secundarias fueron densidad mineral ósea, metabolitos y hormonas. Las meta‐regresiones examinaron las relaciones entre esas medidas y moderadores. Resultados Las pacientes femeninas con AN pre‐tratamiento mostraron un 50% menos de masa grasa (MD: −8.80 kg, CI 95%: −9.81, −7.79, Q = 1.01 × 10– 63) y 4.98 kg (CI 95%: −5.85, −4.12, Q = 1.99 × 10– 28) menos de masa libre de grasa, con masa grasa preferentemente almacenada en la región del tronco durante la recuperación temprana del peso (4.2%, CI 95%: −2.1, −6.2, Q = 2.30 × 10– 4). Aunque la mayoría de los rasgos regresaron a los niveles vistos en los controles sanos después de la restauración del peso, la masa libre de grasa (MD: −1.27 kg, CI 95%: −1.79, −0.75, Q = 5.49 × 10– 6) y la densidad mineral ósea (MD: −0.10 kg, CI 95%: −0.18, −0.03, Q = 0.01) permanecieron significativamente alteradas. Discusión La composición corporal es marcadamente alterada en la AN, lo que garantiza la investigación en estos fenotipos como predictores de riesgo clínico o de recaída. Notablemente, la alteración a largo plazo de los niveles de masa libre de grasa y densidad mineral ósea sugieren que estos parámetros debe ser investigados como potenciales rasgos indicadores de AN.

Discussion: Body composition is markedly altered in AN, warranting research into these phenotypes as clinical risk or relapse predictors. Notably, the long-term altered levels of fat-free mass and bone mineral density suggest that these parameters should be investigated as potential AN trait markers.
However, conclusions are limited by small sample sizes and consequent mixed findings.
Molecular genetic studies have revealed that individuals with AN carry genetic variants that increase their liability to AN and concurrently predispose them to lower body fat percentage, lower fasting insulin, and higher high-density lipoprotein cholesterol concentrations, suggesting that metabolic factors may play an etiological role (Duncan et al., 2017;Watson et al., 2019). Additionally, longitudinal investigations of a British birth cohort showed that girls who develop AN later in life are already underweight at the age of 4 years when compared to healthy children (Yilmaz, Gottfredson, Zerwas, Bulik, & Micali, 2019), adding evidence for a developmental component.
A systematic review showed that adolescents and adults differently lose fat tissue when affected by AN, with adolescents losing more central fat tissue and adults more peripheral fat tissue. During weight recovery, individuals with AN show emergent central adiposity which typically attenuates over time .
These clinical and genetic findings encourage the meta-analytic reassessment of the role of body composition traits, such as fat mass and fat-free mass, their regional distribution, and their changes associated with weight restoration and long-term weight recovery in AN.
Meta-analyses have four major advantages compared to systematic reviews. Increasing statistical power through pooling results from independent samples leads to more precise estimates of the underlying effect. Meta-analyses estimate the heterogeneity (i.e., inconsistency) among effect sizes from the individual studies included, which are crucial for the interpretation of the pooled estimates. Meta-regressions are used to investigate potential moderators of the pooled effect sizes and the relationships between the outcomes of interest, while extensions of meta-analytical models can estimate potential publication bias (Nakagawa, Noble, Senior, & Lagisz, 2017).
The goals of these meta-analyses were to (a) replicate findings from the systematic review on fat mass; (b) extend the observations by quantifying them; (c) include fat-free mass; (d) include bone mineral content and density; (e) investigate their associations with each other; and (f) if possible, relate findings to secondary outcomes, such as metabolic and hormonal parameters. This analytical approach is aimed at understanding the potential associations between these factors that are known to be physiologically interrelated. A thorough and rigorous examination of body composition and related laboratory parameters in individuals with AN could elucidate some of the physiological changes associated with this serious disorder, which could lead to more effective medical management, monitoring, and treatment approaches.
2 | METHOD 2.1 | Search strategy, selection criteria, and data extraction Our meta-analysis was conducted according to PRISMA guidelines (Moher, Liberati, Tetzlaff, Altman, & PRISMA Group, 2009)  Psychiatric Association, 2013). We used key search terms including "anorexia nervosa" AND ("body composition" OR "body fat" OR "fat mass" OR "body fat percentage" OR DXA OR BIA OR "fat free mass" OR "lean mass"). The search was repeated by coauthors to avoid selection bias. Furthermore, we screened the references of published articles and reviews. Our search results, including the selection process, are presented in Figure 1 according to PRISMA guidelines. Our selection criteria are presented in Table 1. In case of multiple publications deriving from the same study population, we selected the articles reporting either the largest or the most recent data set. In case of conflict between these two criteria, large sample size was prioritized.
We extracted the information presented in Table 1 from every identified study using a standardized data extraction sheet.
The data extraction sheet was based on two previous metaanalyses (Hussain et al., 2019;Ilyas et al., 2018) and included variables that were hypothesized to be associated with body composition, hormonal, or metabolic measures, including fasting status and period, medications, stage of the menstrual cycle, or treatments for longitudinal studies. If enough studies reported these variables, we performed meta-regressions to investigate their associations with our primary and secondary outcomes.

| Quality of study assessment (Newcastle-Ottawa scale)
We used the Newcastle-Ottawa Scale (NOS) to assess the quality of nonrandomized studies (Wells et al., 2009). Each study is judged on three broad perspectives: (a) the selection of the study groups; (b) the comparability of the groups; and (c) the ascertainment of the outcome of interest for case-control studies. The NOS evaluates these three quality parameters divided across eight specific items. Each item on the scale is scored from one point, except for comparability, which can be adapted to the specific topic of interest to score up to two points. It has been designed to be used in meta-analyses and systematic reviews. For the observational studies, low quality was defined as NOS score ≤8.0 and high quality as score >8.0 (maximum score 9).
We chose a random-effects model, which assumes that the heterogeneity in the differences between clinical and control groups is due to both within-study and between-study variation, as we anticipated differences in procedures and study populations between studies. We quantified the heterogeneity through a restricted maximum-likelihood (REML) approach. For the analysis of subtypes, posttreatment, and weight-recovered patients with AN, the control groups from the acutely-ill/pretreatment analysis were reused because (a) control groups were not measured repeatedly and (b) none of the studies had separate control groups for each subtype analysis. Although some studies included covariates in their statistical analysis (Bratland-Sanda et al., 2010;Bredella et al., 2008;Dellava, Policastro, & Hoffman, 2009;DiVasta et al., 2007;Fernández-Soto, González-Jiménez, Chamorro-Fernández, & Leyva-Martínez, 2013;Haas et al., 2005;Karlsson, Weigall, Duan, & Seeman, 2000;Kosmiski, Schmiege, Mascolo, Gaudiani, & Mehler, 2014;Maïmoun et al., 2018;Nakahara et al., 2007;Schneider et al., 1998), we only used raw values without including study-specific covariates to increase comparability across individual studies. Weight recovery was defined in accordance with DSM-IV and DSM-5 criteria with BMI >18.5 kg/m 2 or >90% ideal body weight. To correct our primary analysis for multiple testing, false discovery rate-adjusted Q values were calculated (Benjamini & Hochberg, 1995).

| Detection and adjustment for publication bias
The results of meta-analyses can be influenced by publication bias (i.e., small study effects). This describes the phenomenon when certain studies have been selected for publication, while others-mostly due to negative findings-have not been published (Nieminen, Rucker, Miettunen, Carpenter, & Schumacher, 2007). Through graphical diagnosis of asymmetry in funnel plots (Egger, Smith, Schneider, & Minder, 1997) and performing Thompson and Sharp tests (i.e., weighted linear regressions) that take variation between studies into account (Thompson & Sharp, 1999), we investigated potential small study effects or publication bias. If the test resulted in a p value below .05, we adjusted the pooled effect estimates using a Copas selection model calculated with the R package "metasens." The model has two components: the first component estimates the pooled effect, while the second estimates a publication probability for each study. A large correlation between these two components suggests that studies with more extreme effects were more likely to be published (Copas, 1999;Copas & Shi, 2000. The models were iteratively optimized using two tuning parameters γ 0 and γ 1 . We present four diagnostic graphics including (a) a funnel plot, (b) a contour plot, (c) a treatment effect plot, and (d) a p value plot.

| Investigation of potential moderators through meta-regression and stratification
To examine the large between-study heterogeneity per metaanalysis (Table 1), we performed meta-regressions using mixed effects models included in the R package "meta" that take the heterogeneity within and between individual studies into account. The models were optimized via a REML approach. Through meta-regression, we investigated whether relevant participant or study characteristics may be associated with the pooled estimates, such as mean age, the time period of follow-up for longitudinal studies, age at diagnosis, age at menarche, age at amenorrhea, duration of illness, percentage of amenorrhea in patients with AN, percentage of medicated patients with AN, percentage of individuals taking F I G U R E 1 PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram of study selection contraceptives, body composition measurement method, blood sample type, body composition parameters, and their differences between cases and controls.
A second approach to test for potential moderators is stratification of the sample into meaningful subgroups and estimation of statistical differences between the pooled estimates per subgroup. We used this approach and stratified by AN subtype.

| Results of the search and selection of studies
A total of 1,498 papers published between 1996 and 2019 were identified by our search terms, and 1,434 (96%) of them were excluded.

| Characteristics of the included studies
We performed four sets of meta-analyses (a) comparing 2,319 pretreatment/acutely ill AN patients with 1,879 healthy controls; (b) comparing 722 post-treatment AN patients with 809 controls; (c) estimating the change in AN patients (n = 722) from pretreatment to posttreatment; and (d) comparing 398 weight-recovered individuals with AN with 660 healthy controls including samples with a long-term follow-up. The pretreatment AN group comprised 229 individuals suffering from the binge-eating/purging (8% of cases) and 701 from the restricting subtype (26% of cases). The shortest follow-up period was 5.14 weeks, and the longest was 2 years (Table S1). Twenty studies (32%) used bioelectrical impedance analysis (BIA) to assess body composition, 39 (63%) used dual-energy X-ray absorptiometry (DXA), and only 3 (5%) utilized magnetic resonance imaging (MRI)--considered to be the benchmark. Thirty of the 62 studies (48%) investigated body composition as a primary outcome, whereas it was a secondary outcome in the remaining studies. The percentage of AN patients with amenorrhea ranged from 0 to 100%, with 11 studies (18%) not providing information on menstrual status (Agüera et al., 2015;Bachmann et al., 2014;Bredella et al., 2012;de Mateo Silleras et al., 2013;El Ghoch et al., 2012;Gniuli et al., 2001;Iacopino et al., 2003;Kirchengast & Huber, 2004;Schneider et al., 1998;Tagami et al., 2004;Tanaka et al., 2003). Thirty-five of 62 studies (56%) did not provide information on the medication status of AN patients, and 32 (52%) did not indicate whether oral contraceptives were used. In AN cases, the duration of illness was on average 52.2 months (SD = 29.4), the duration of amenorrhea 23.0 months (SD = 18.3), and the age at diagnosis 17.5 years (SD = 3.0). Cases and controls were well matched for age ( Figure S1) and, notably, we did not observe a difference in age at menarche ( Figure S2) or height ( Figure S5) between AN cases and controls.

| Data and analyses results of meta-analyses and meta-regressions
Our results from the 94 meta-analyses show that a wide range of alterations in several key body composition and biochemical measures exist in AN cases compared with healthy controls (Figure 2 and Figure S3). For 95% confidence intervals and Q values, heterogeneity estimates (τ 2 and I 2 ), and adjusted estimates due to estimated publication bias, see binge-eating/purging subtype of AN were detected in our metaanalysis prior to treatment except for total body water (Table S3).
Between-study heterogeneity (I 2 ) was observed in 62 meta-analyses (70%) and ranged from 52 to 99%, confirming our choice of a random-effects model. To investigate moderators implicated in heterogeneity, we performed 411 meta-regressions (Tables S4-S7). Six meta-analyses showed funnel plot asymmetry, indicating small study effects. Therefore, we fitted Copas models to adjust for those effects and estimate the probable number of unpublished studies (Table 1 and Figures S96-S101).  but not through fat-free mass (Table S5). After weight recovery, no statistically significant MD in BMI between female AN cases and controls was detected. The pretreatment BMI difference between male AN cases and controls was −5.48 kg/m 2 (95% CI: −7.87, −3.09, Q = 1.80 × 10 −5 ).
After treatment, female AN patients showed a higher trunk body fat percentage than controls at 12.0% (95% CI: 9.5, 14.4, p adjCopas < 1.00 × 10 −4 ) of total body mass. However, this finding was strongly influenced by publication bias with an estimated 52 unpublished studies. These results on body composition were not influenced by height as female and male cases and controls showed no meaningful difference (i.e., 1 cm pretreatment) or by age as metaregressions were nonsignificant (Tables S4-S7).
Before treatment, we observed a 393.95 kcal/day (95% CI: However, resting energy expenditure was not corrected for fat-free mass or body mass in the original studies, limiting its interpretability. Before treatment, total body water in females was associated with fat mass (β metareg = 0.60, p = .01) and the difference in fat-free mass between AN cases and controls (β metareg = 0.48, p = .003).

|
Before treatment, female AN patients exhibited lower bone mineral density in several regions, including hip, lumbar spine, and femoral neck, with a few being likely to persist after weight recovery. These findings were associated with duration of illness, the age of AN cases, and differences in fat mass between cases and controls (Supporting Information: Secondary Outcomes: Detailed Bone Mineral Measures and Table S4). Cases and controls in our meta-analyses were age-and height-matched ( Figures S1 and S6); therefore, these variables are unlikely to be associated with these results.

| Secondary outcomes: Metabolites and hormones
Exploratory results showed that fasting insulin and glucose concentrations were lower in female AN patients compared with controls but not associated with fat or fat-free mass, while lower leptin was associated with fat mass pretreatment. After treatment, these measures returned to concentrations seen in healthy controls. Before treatment, thyroid hormones, cortisol, and IGF-1 were lower in female AN patients, and all three measures were associated with fat mass, whereas higher cortisol in AN patients was associated with fat-free mass. For detailed results, see Supporting Information: Secondary Outcomes: Metabolites and Hormones.

| DISCUSSION
Our primary meta-analyses showed marked alterations in body composition traits in patients with AN before and after treatment. Before treatment, all three major body compartments-fat, fat-free, and bone F I G U R E 4 Cross-sectional metaanalysis of studies reporting fat-free mass content in acutely-ill/pretreatment female AN patients compared with healthy controls. Thirty-seven samples had the appropriate data for the metaanalysis with 2,319 AN cases and 1,879 controls. A random-effects meta-analysis revealed a pooled estimate of the mean difference (MD: −4.98 kg; 95% CI: −5.8, −4.1; Q = 1.99 × 10 −28 ) with the mean differences ranging from −12.16 to 0.20 kg. Heterogeneity between studies was statistically significant (τ 2 = 5.92; p = 1.22 × 10 −58 ; I 2 = 90.5%). C, subtypecombined sample mass-showed significant reductions that were only partially restored after treatment. Our meta-analysis estimated~50% lower body fat in AN patients which was mirrored by leptin concentrations (Perry & Shulman, 2018), both of which recovered with treatment. In females, significant differences were observed in body fat distribution after treatment as body fat is primarily stored in the trunk. This distribution pattern may be due to increased insulin sensitivity observed in AN patients (Ilyas et al., 2018) potentially similar to observations in healthy individuals after short-term overfeeding (McLaughlin et al., 2016). We did not detect meaningful or statistically significant differences in body fat distribution in weight-restored patients, indicating potential redistribution occurring over longer term follow-up.
A new finding from our meta-analysis is that lower fat mass in females with AN was correlated with significantly low bone mineral content and density across the whole body. We speculate that the hormonal cross-talk between fat and bone tissue may be influencing this association (El Ghoch et al., 2016;Hawkes & Mostoufi-Moab, 2018), potentially mediated through greater bone marrow adipose tissue observed in AN (Fazeli & Klibanski, 2018;Suchacki & Cawthorn, 2018). Whole-body bone mineral content remained low in weightrecovered individuals with AN. However, as only two studies followed patients for longer than 6 months (Dellava et al., 2009;Karlsson et al., 2000), the duration of follow-up was insufficient to draw firm conclusions because bone mineral mass is slow to normalize. Future studies should be designed to capture long-term changes. In men with AN, fat mass and fat-free mass were lower before treatment than in controls.
However, long-term follow-up studies are missing. It has been reported that short-term weight restoration may normalize body composition patterns but could also lead to central adiposity , but sample sizes of reports of males are very small. Additionally, in our analysis alterations in bone mineral mass did not affect the height of individuals with AN.
Another new finding in our meta-analysis is that we observed a 5 kg lower fat-free mass in female AN patients, which remained lower even after treatment and in weight-recovered AN patients, indicating that current treatment regimens may insufficiently target fat-free mass.
Future studies should also assess muscle mass to identify the components of fat-free mass that are most associated with this reduction.
Our secondary outcomes-associations between detailed body composition and laboratory parameters in AN-were difficult to assess as only a few published studies reported both outcomes consistently.  Prioletta et al., 2011;Zuniga-Guajardo, Garfinkel, & Zinman, 1986).
This approach is supported by epidemiological associations of AN with type 1 diabetes (Hedman et al., 2018) and its genetic overlap with fasting insulin (Duncan et al., 2017), type 2 diabetes (Watson et al., 2019), and insulin sensitivity .
AN was associated with body fat percentage-associated low T 3and T 4 -syndrome pretreatment, whereas thyroid-stimulating hormone concentrations were normal. Associations between fat mass and thyroid hormones have been described before (Kwon et al., 2018); however, sufficiently powered long-term follow-up studies in AN are absent.
Steroid hormone concentrations were altered showing high cortisol, low estradiol, and normal testosterone. Estradiol was negatively associated with fat-free mass, whereas cortisol was positively associated with fat mass. These findings suggest that fat-free mass may be a potential moderator for the return of menses in AN patients and should be further investigated as most research in recovery of menses primarily focused on BMI-or weight-related cutoffs (Misra et al., 2006;Swenne, 2004). Potential reverse causation should also be taken into account where altered estradiol concentrations may precede changes in fat-free mass.
Overall, the meta-analyzed study sample was highly selected and biased as it comprised mostly European females aged between 14 and 31 years, emphasizing the urgent need for studies including diverse ancestries, such as Asia, South and Central America, and Africa.  Table 3 and should be carefully assessed in future studies (Hernandes et al., 2017).
Most importantly, blood comprises approximately 3,500 highly correlated and interacting proteins (hupo.org; Schwenk et al., 2017) and 4,600 metabolites (serummetabolome.ca; Psychogios et al., 2011); therefore, the measurement of single proteins, hormones, or metabolites is ill-advised. Metabolomics, proteomics, and lipidomics can capture large amounts of information at adequate statistical power when used in large samples (Hernandes et al., 2017

| CONCLUSIONS
Detailed measurement of body composition with simple methods, such as BIA or DXA, which offers additional information on bone tissue, may help refine our understanding of the nature of AN and its diagnosis. Our meta-analyses showed that all body compartments were markedly altered in AN. Individuals with AN presented with 50% lower fat mass and prolonged recovery periods for fat-free mass and bone mineral content. The core implication of body composition differences are alterations in metabolism, growth, and development.
Although results must be interpreted with caution given small samples, we found evidence indicating alterations in fasting insulin, thyroid, sex, and stress hormones in AN, which appeared to partially normalize with weight gain and recovery. Large birth cohorts that collected information on eating disorders along with metabolomic information offer a rich and exciting opportunity for prospective investigations that add to our understanding of body composition and metabolic mechanisms in risk and maintenance of eating disorders. T A B L E 3 Minimum requirement of variables that should be assessed, reported, and included in statistical analyses of case-control studies examining anorexia nervosa or other eating disorders to facilitate reproducibility, meta-analysis, and meta-regression