Brain morphology in Anorexia Nervosa and its subtypes: A multi-cohort study of individual participant data

This is an uncorrected proof. Figures Abstract Background In a recent coordinated meta-analysis of neuroimaging data, we reported gray matter (GM) alterations in acutely underweight patients with anorexia nervosa (AN). Here, we extend these findings by examining individual variation in brain structure within AN, individual-level differentiation between AN and healthy controls (HC), and differences between AN subtypes, with potential relevance for understanding clinical heterogeneity. Methods and findings We analyzed individual-level data from 11 international sites in the ENIGMA Eating Disorders Working Group, including 570 female participants with AN and 739 HC. We examined cortical thickness, cortical surface area and subcortical volumes in AN versus HC using three complementary approaches: (i) group-level differences in a mega-analysis correcting for age effects, (ii) frequencies of extreme deviations (infra-/supranormal; z 1.96) based on normative reference models by the CentileBrain Initiative, and (iii) individual-level classification performance using machine learning. The same analytic framework was applied to compare AN restricting versus binge-eating/purging subtype, additionally correcting for BMI effects. Mega-analyses reinforced previous meta-analytic findings of pronounced and widespread GM deficits in AN compared to HC. Normative modelling revealed that the frequency of infranormal z-scores (23/68 cortical thickness, 13/14 subcortical volume metrics) and supranormal z-scores (35/68 cortical thickness, 17/68 cortical surface area metrics) was significantly higher in AN than expected based on reference data. Individuals with AN could be reliably differentiated from HC using machine-learning classifiers (ROC–AUC = 0.75–0.81). In contrast, neither group-level differences nor frequency of extreme z-scores differed between AN subtypes, and individuals with different subtypes could not be reliably differentiated from each other. Importantly, the observational design cannot distinguish neurobiological differences related to AN from the effects of starvation or low BMI in the AN versus HC analyses. The lack of differences between subtypes does not exclude brain structural differences between AN subtypes that might be detectable with other modalities or analytic approaches. Conclusion Using a mega-analytic approach, we confirm widespread GM deficits in AN, show that these alterations are (in some patients) extreme, and demonstrate that they enable robust classification with superior performance compared to most MRI-based psychiatric classification studies. The absence of differences between AN subtypes may reflect shared neurobiology, though other imaging modalities may reveal distinctions beyond brain structure. Author summary Why was this study done? - Previous large-scale studies have shown that people with anorexia nervosa (AN) who are currently underweight often have reduced cerebral gray matter, but it was unclear to what extent these changes vary across individuals. - We also aimed to determine whether brain scans can reliably distinguish individuals with AN from healthy people, and whether brain structure differs between AN subtypes: restricting type versus binge-eating/purging type. What did the researchers do and find? - We combined structural brain MRI data from 11 international sites, analyzing 570 females with AN and 739 healthy controls. - Brain structure was examined at multiple levels: average group differences, individual deviations relative to a large normative reference dataset, and machine-learning classification. - On average, we confirmed that people with AN have reduced gray matter compared with healthy controls (90% of cortical thickness, 84% of cortical surface area, and 100% of subcortical volume metrics affected). However, we also observed a higher-than-expected number of “extreme” brain measurements and overall greater variability, indicating that brain structure changes in AN are highly heterogeneous across individuals. - Despite this heterogeneity, a multivariate pattern across many brain features reliably distinguished AN from healthy controls (ROC–AUC ~0.75–0.81), while AN subtypes could not be reliably differentiated. What do these findings mean? - Brain structure changes in acute AN do not reflect a single uniform pattern; instead, there is marked biological heterogeneity, with some individuals showing pronounced alterations while others are relatively less affected. - Nevertheless, the overall pattern across multiple brain features is sufficiently consistent to distinguish AN from healthy controls, with better performance than is typical for MRI-based classification studies in psychiatry. - The absence of structural differences between AN subtypes suggests shared neurobiology in brain morphology. However, the study is limited by its observational design, focus on currently underweight females, and inability to determine whether brain differences are causes or consequences of illness or how they evolve with recovery. Citation: Bernardoni F, Arold D, Schoppik L, Bahnsen K, Ge R, Moreau C, et al. (2026) Brain morphology in Anorexia Nervosa and its subtypes: A multi-cohort study of individual participant data. PLoS Med 23(5): e1004809. https://doi.org/10.1371/journal.pmed.1004809 Academic Editor: Perminder Singh Sachdev, University of New South Wales, AUSTRALIA Received: October 23, 2025; Accepted: April 30, 2026; Published: May 20, 2026 This is an open access article, free of all copyright, and may be freely reproduced, distributed, transmitted, modified, built upon, or otherwise used by anyone for any lawful purpose. The work is made available under the Creative Commons CC0 public domain dedication. Data Availability: Individual-level data underlying the findings of this study cannot be shared publicly because they are governed by site-specific ethical approvals and national and institutional data protection regulations at the 11 contributing ENIGMA Eating Disorders Working Group sites. The study authors are not the legal custodians of these data. Data access inquiries must be directed to the relevant institutional data custodian or ethics/governance office at each contributing site (Denver: COMIRB@ucdenver.edu, Dresden: ethikkommission@mailbox.tu-dresden.de, Erlangen: ethik-kommission@fau.de, Heidelberg: ethikkommission-I@med.uni-heidelberg.de, London: hampstead.rec@hra.nhs.uk, rec@kcl.ac.uk, Oslo: rek-sorost@medisin.uio.no, Padova: comitato.etico@aopd.veneto.it, San Diego: irb@ucsd.edu, Torino: comitatoetico@pec.cittadellasalute.to.it, Toronto: research.ethics@camh.ca, Utrecht: metc@nedmec.nl, researchoffice@umcutrecht.nl). Any access is subject to local approval procedures and applicable legal and contractual restrictions. Summary-level data underlying the reported findings are provided in Tables A and B in the S1 Appendix. The code used for the analyses is available on OSF (https://osf.io/xrjkf/overview?view_only=b73a380dfaf94c36a69d9354e0c92679) or through the DOI (https://doi.org/10.17605/OSF.IO/XRJKF). Funding: This work was supported by the Else Kröner-Fresenius-Stiftung (https://ekfs.de/, Grant No. 2019_A118 [to FB, DA, and LS]), the European Union and the Saxon State Parliament (https://www.era-learn.eu/network-information/networks/personalised-medicine, EP PerMed project BIOREXIA; 100770101 [to SE]), German Research Foundation (https://www.dfg.de/en, Grant Nos. SI 2087/2-1 and BR 4852/1-1 [to JJS] and Grant Nos. EH 367/5-1 and EH 367/7-1 [to SE]), the Carus Promotionskolleg from the Medizinische Fakultät Carl Gustav Carus, Technische Universität Dresden (https://tu-dresden.de/med/mf/forschung-internationales/nachwuchsfoerderung-dscs/internefoerderprogramme/carus-promotionskolleg-dresden, to KB), National Institutes of Health (https://www.nih.gov/, Grant No. R21MH86017 [to AB-G and CEW], Grant No. R01MH113588 [to AB-G, and CEW], Grant Nos. K23MH080135 and R01MH096777 [to GKWF and MES]), Swiss Anorexia Nervosa Foundation (https://www.fundraiso.com/en/organisations/schweizerische-anorexia-nervosa-stiftung, Project No. 57-16 [to JJS] and to SE), National Institute for Health Research Mental Health Biomedical Research Centre at the South London (https://www.kcl.ac.uk/ioppn/research/nihr-maudsley-brc) and Maudsley NHS Foundation Trust and King’s College London (https://www.kcl.ac.uk/ioppn/our-connections/slam, ICC, OO, and UHS), Research Council of Norway (https://www.forskningsradet.no/en/, Grant Nos. 288083 and 323951 [to CKT]), South-Eastern Norway Regional Health Authority (https://www.helse-sorost.no/, Grant Nos. 2021070, 2023012, and 500189 [to CKT]), National Institute of Health Research Senior Investigator Award (https://www.nihr.ac.uk/, to UHS), CAMH AFP Innovation Fund (https://www.camh.ca/, CAM-14-001 [to AM, AK, and AV]), and Technische Universität Dresden SFB 940 (https://tu-dresden.de/, to SE). The ENIGMA ED Working Group acknowledges the National Institutes of Health Big Data to Knowledge award for foundational support and consortium development (https://commonfund.nih.gov/bd2k, Grant No. U54 EB020403 [to PMT]). The normative CentileBrain Models were developed with support by the National Institutes of Mental Health (R01MH134962 [to SF and PMT]). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: I have read the journal’s policy and the authors of this manuscript have the following competing interests: LAB is a scientific advisor to Juniver, LLC. The other authors report no conflict of interest. Abbreviations: AN, anorexia nervosa; BMI, body mass index; CT, cortical thickness; DSM-IV or DSM-5, Diagnostic and Statistical Manual of Mental Disorders, Fourth or Fifth Edition; ED, Eating Disorders; FDR, false discovery rate; GLMs, General Linear Models; GM, gray matter; HC, healthy controls; ICV, intracranial volume; ICD-10, International Classification of Diseases, Tenth Revision; MRI, magnetic resonance imaging; PCA, principal component analysis; PR-AUC, precision-recall area under the curve; ROC-AUC, area under the receiver operator characteristic curve; SA, surface areas; STROBE, Strengthening the Reporting of Observational Studies in Epidemiology; SV, subcortical volumes; SVM, Support Vector Machine. Introduction Anorexia nervosa (AN) is an eating disorder characterized by severe dietary restriction resulting in low weight and a high mortality rate due to complications of starvation [1]. While the exact aetiology remains unclear, its biological underpinnings are widely acknowledged [2]. No specific pharmacological treatment exists for AN, with therapy typically focusing on psychotherapy and weight restoration [3]. However, clinical presentation and outcomes are highly variable—a significant proportion of patients experience relapses and chronicity, and even suicide or premature death [1]. Understanding why standard therapies work for some patients but not others is critical. While several factors influence clinical trajectories in AN [4], this study specifically focuses on individual-level variation in brain structure, building on prior evidence that such neurobiological differences can help predict treatment outcomes [5]. Previously, a prospective harmonized meta-analysis from the ENIGMA (Enhancing NeuroImaging Genetics through Meta-Analysis) Eating Disorders (ED) Working Group (http://enigma.ini.usc.edu/ongoing/enigma-eating-disorders/) revealed sizeable and widespread gray matter (GM) reductions associated with low weight and body mass index (BMI) in AN [6]. Evidence from related populations suggests that GM reductions may also occur in the context of low BMI outside AN [7], including in early-onset restrictive eating disorders in very young children [8] and in population-based samples of underweight preadolescents (e.g., Generation R cohort [9]). Going beyond these findings, we collected the largest multi-site cohort of individual-level data within the ENIGMA ED Working Group for participants with AN and healthy controls (HC). In contrast to our previous group-level meta-analysis, analyzing individual-level data enables examination of variability related to clinical AN subtypes [10]. Clinically, the restricting subtype (AN-R) is characterized by severe food restriction, which may occur with or without increased energy expenditure (e.g., excessive exercise), while the binge-eating/purging subtype (AN-BP) is marked by purging compensatory behaviors (e.g., self-induced vomiting, laxative misuse, and enemas) alongside food restriction [10]. Importantly, AN-BP is distinguished from bulimia nervosa by the persistence of significantly low body weight, despite the presence of binge-eating and purging behaviors [10]. Although diagnostic crossover between AN-R and AN-BP is common, particularly with increasing illness duration [11], and the validity of this subtype distinction has been debated [12,13], these subtypes are currently retained in diagnostic systems to describe differences in clinical presentation. AN-R is often associated with an earlier age of onset, predominance in adolescent samples, and a more stable course [14]. AN-BP becomes more prevalent with longer illness duration, has a more fluctuating course, higher levels of suicidality [15–17], relapse [11,18], and co-occurring psychiatric symptoms [11,19–22]. While no differences in cerebral blood flow between these subtypes have been reported previously [23], the large sample size of the present study provides a robust opportunity to investigate potential differences in brain morphology. This may advance understanding of the biological underpinnings of structural alterations in AN, inform the ongoing debate regarding the validity of the current subtype distinction [13], and ultimately support the development of more targeted treatments [5]. Capitalizing on the richness of individual-level data, we adopted three complementary approaches to more precisely characterize brain alterations associated with AN and AN-subtypes at the group and individual level. First, to examine group-level differences, we conducted a mega-analysis, a multisite data analysis where the individual-level data are shared rather than just the summary statistics from each site. This offers several advantages. For example, research conducted by various ENIGMA working groups has shown that mega-analyses yield lower standard errors and narrower confidence intervals compared to meta-analyses [24–27]. These improvements become even more pronounced when adjusting for different scanning devices and sequences using the ComBat method [28], as opposed to the random-effects approach typically used in meta-analyses [29]. Second, given the clinical heterogeneity observed among individuals with the same psychiatric diagnosis [30,31], which has been linked to variability in the distribution of extreme structural magnetic resonance imaging (MRI) values [32–35], we examined extreme z-scores (infra- or supra-normal) in regional cortical and subcortical brain morphometry in patients with AN relative to a normative reference sample. The normative reference was defined using models from the CentileBrain Initiative [36], which estimate expected values and normative ranges for brain phenotypes (e.g., cortical thickness (CT) in a given brain region), based on age, sex and global GM variables. The location and frequency of infra- and supra-normal z-scores were then compared between AN and the normative reference, and between AN-BP and AN-R. Third, differences between individuals with AN and HC, or between individuals with AN-R and AN-BP might be multivariate, characterized by complex, nonlinear patterns involving multiple metrics of brain structure (e.g., CT in different regions), rather than univariate differences, which focus on single measures such as CT for a specific cortical region. To examine multivariate differences in structural MRI images, we applied machine learning to classify individuals with AN from HCs, as well as individuals with AN-R from those with AN-BP. The classification performance indicates the extent to which measurable and consistent differences in brain structure can predict group membership. By analyzing the feature importances from the model, we can further identify the specific brain regions where these multivariate differences are located. Together, these complementary analyses were designed to address the overarching question of whether AN is associated with consistent alterations in brain morphology at both the group and individual level, and whether these alterations differ between AN-R and AN-BP. Based on previous structural MRI findings in AN, we hypothesized that individuals with AN would show lower brain morphometric values than HCs [6], particularly for CT and SV. We further hypothesized that individual-level normative deviations would be more frequent in AN than in the normative reference sample, reflecting heterogeneity in the spatial distribution of extreme morphometric values. Finally, based on a previous single-site study [5], we expected that multivariate models would classify AN versus HC above chance level, indicating distributed structural differences in brain morphology. Because previous structural MRI studies have rarely examined morphological differences between AN-R and AN-BP directly, and because available evidence does not provide a consistent basis for predicting the direction or regional distribution of such differences, we did not formulate a directional hypothesis for subtype comparisons. Methods Study samples Thirteen cohorts contributed individual-level data to the AN arm of the ENIGMA ED working group, 12 of which included subtype information for each participant with AN. Inclusion and exclusion criteria were as in our previous study [6]. Specifically, patients with AN and HC were selected within each site according to standardized inclusion and exclusion criteria, as in the main analysis by Walton and colleagues [6], but patients were not stratified into two groups according to their weight status. Compared to the sample in Walton and colleagues [6], we excluded cohorts that could not share single participant data, while for some sites, data for more participants became available. To be included in the AN group, participants had to be female and meet the DSM-IV-TR, DSM-5, or ICD-11 criteria for AN, i.e., restriction of energy intake leading to significantly low body weight, intense fear of gaining weight (or persistent behavior interfering with weight gain), and disturbance in body weight/shape experience. Specifically, participants with AN had a BMI 18 years old) or below the 10th percentile for age-adjusted BMI in adolescents. HC participants were also female, with BMI > 17.5 kg/m² for adults or above the 10th percentile for age-adjusted BMI in adolescents, and with no current or lifetime diagnosis of any eating disorder. Although this range includes some underweight individuals, all HC were screened as healthy eaters without evidence of an eating disorder or somatic causes of low weight. Our rationale for including adult HC with BMI 37,000 healthy individuals, aged 5–90 years [36]. For each individual regional morphometric measure, the models provide a z-score which quantifies the deviation from the population mean. In this study, the CentileBrain female-specific models of regional CT and SA and SV were applied to the corresponding harmonized data of the sample of patients with AN. Following Haas and colleagues [43], we defined regional “extreme” z-scores as infranormal if z 1.96, corresponding to the 2.5th and 97.5th percentiles, respectively. Intermediate values (i.e., −1.96 .14; ROC-AUC[ComBat-GAM] = 0.55, PR-AUC[ComBat-GAM] = 0.30, p > .12), with PR-AUC values close to the baseline defined by class prevalence (~0.24), see Fig K in S1 Appendix. Employing a neural network instead of SVM did not change these results (Section A.6, Figs L and M in S1 Appendix). Discussion In this large-scale multisite mega-analysis, we confirmed robust and widespread structural GM reductions (90% of CT, 84% of SA and 100% of SV metrics affected) in individuals with AN compared to HC, consistent with previous meta-analytic findings [6]. The strongest effects were observed for CT (mean Cohen’s d = –0.54), followed by SV (mean d = –0.37), and SA (mean d = –0.23), underscoring the pronounced and spatially extensive nature of brain structure abnormalities in AN. Using a normative modeling approach (CentileBrain), we further showed that individuals with AN had a significantly higher-than-expected proportion of extreme (infra- or supra-normal) z-scores in regional brain metrics, particularly for CT. Specifically, z-scores for each participant were computed relative to predictions from the normative model for a reference female of the same age and, importantly, matched global structural properties (mean CT, total SA, and total GM volume). While SA measures showed (relative) localized increases in supranormal values, SV were characterized by a predominance of infranormal deviations across nearly all structures. Finally, machine learning reliably distinguished individuals with AN from HC based on structural MRI metrics (ROC-AUC = 0.75–0.81), indicating that brain morphological features are sufficiently altered to allow for accurate individual-level classification. Critically, however, no significant differences emerged between AN-R and AN-BP, neither in group-level comparisons, nor in the frequency of normative deviations. Even machine learning methods, which are sensitive to complex multivariate patterns, failed to classify AN subtypes above chance level. These findings suggest that brain structural differences between AN-R and AN-BP are subtle or nonexistent. The results for the univariate group comparisons closely aligned with those from our previous meta-analysis, which was based on a larger number of sites [6], and other previous studies [40,53]. For CT, the agreement on regions with a significant difference with our previous prospective meta-analysis [6] was 95.59% (in the remaining 4.41% regions, the mega-analysis detected a difference but not the meta-analysis), for SA 72.06% (in 22.06%/5.88% regions the mega-/meta-analysis detected a difference but not the meta-/mega-analysis), and for SV the agreement was 100%. These findings suggest that the current mega-analytic approach provided greater statistical power, enabling the detection of significant effects even for metrics with smaller effect sizes. Specifically, for CT metrics where a significant reduction was also found in Walton and colleagues [6], the effect sizes were comparable (d = −0.55(0.21) versus d = −0.57(0.20) in Walton and colleagues [6]), while in the left rostral anterior cingulate only our analysis detected an effect (d = −0.11). Importantly, in contrast to Walton and colleagues [6] and reviews that reported either cortical thinning or no significant differences—but not increases—in AN [53–56], we observed higher CT in the left temporal pole and the right entorhinal cortex in underweight patients with AN. This finding is compatible with large single-site studies measuring CT vertexwise [40,41], and a machine learning study suggesting that higher CT in these regions helps classify individuals as having AN [5]. Notably, these regions show little to no normative cortical thinning during adolescence [57,58], suggesting that they follow distinct maturational trajectories that may confer resilience to starvation-induced cortical loss. Together with the observation that the effect sizes for CT reductions span a range between d = −0.11 and d = −0.96, this finding suggests that, while cortical thinning and overall reductions in GM volume in AN are widespread and predominantly driven by BMI [40,41], some regions are more susceptible to structural changes, potentially reflecting region-specific cellular composition or metabolic vulnerability [40]. Despite widespread alterations likely dominated by starvation-related metabolic vulnerability, it is informative to consider how the regions showing the largest or most distinct effects relate to AN traits or phenotypes. The strongest CT reductions were observed in posterior parietal regions, implicated in visuospatial processing and mental imagery and repeatedly discussed in relation to altered body representation, body image distortion, and multisensory integration in AN [59,60]. For SA, the most pronounced effects were located in the bilateral transverse temporal gyri, classically associated with early auditory processing [61]; links to core AN phenotypes are necessarily speculative here and may reflect more general, distributed state-related effects rather than a specific auditory phenotype. For SV, the largest reductions involved the bilateral thalamus, a central hub for sensory relay and cortico–subcortical integration [62]. Thalamic volume reductions have been consistently reported in AN and linked to metabolic and endocrine markers, including leptin levels and BMI-related pathways ([63]. Finally, a small number of regions showed opposite-direction effects, with higher CT in the left temporal pole and right entorhinal cortex. Given the modest effect sizes and the multiple-comparisons context, these findings should be interpreted cautiously. Nonetheless, these medial and anterior temporal regions are implicated in socio-emotional [64] and memory-related processes [65] and may reflect inter-individual heterogeneity, compensatory mechanisms, or trait-like differences in subsets of individuals rather than starvation effects per se. The results for the frequencies of extreme z-scores as assessed within the normative model from the CentileBrain group [36] need to be interpreted with care. Since z-scores were adjusted for global structural metrics, a z-score outside the normative range does not imply an absolute increase or decrease in a given region, but rather a deviation relative to normative individuals with similar global brain characteristics. For CT, both supranormal (primarily in temporal and frontal regions) and infranormal (mainly in parietal, occipital, and frontal regions) z-scores were more frequent in AN than expected, indicating significant disruptions of the normative profile beyond the global thinning pattern in a significantly high number of patients. For SA, only the frequency of supranormal deviations was significantly elevated, while for SV, there was a consistent and widespread increase in the frequency of infranormal z-scores. This suggests a more heterogeneous pattern of deviations in the AN group for brain morphology metrics. Indeed, on average, standard deviations of individual z-scores were higher in the AN group compared to the normative reference for both CT, SA, and SV metrics. This finding suggests that AN is not associated with a uniform shift in brain structure, but rather with genuine biological heterogeneity: some patients show severe alterations in a given metric, while others are minimally affected. Normative-modelling studies across multiple psychiatric disorders have reported similar findings: group averages provide limited information, and individuals show heterogeneous deviations from typical brain organization [32,35,66]. Machine learning classification well above chance indicates that brain scans from individual participants with AN can be reliably distinguished from those of HC. The superior performance of the ComBat-GAM compared to the LSsO pipeline, together with the absence of overfitting, supports its effectiveness in mitigating site effects. This suggests that the model learned to rely on biologically meaningful structural MRI features rather than technical artefacts. In contrast, the lower performance with LSsO suggests that uncorrected site-specific variability can hinder generalization. Notably, while the ComBat-GAM setup enables testing on unseen data from known sites, the LSsO framework provides a more stringent evaluation by estimating generalization to entirely new sites, where site effects cannot be adjusted for in advance. In the ComBat-GAM pipeline, the most relevant features were CT metrics in the frontal, parietal, and occipital lobes, all showing significant negative importance—indicating that lower values contributed to classification in the AN group. For the LSsO pipeline, in addition to CT metrics, also the thalamus volume had higher negative importance. In contrast, CT in the temporal pole and entorhinal cortex had positive importance for both pipelines (albeit 4–5 times smaller in magnitude than the negative importances found for CT metrics in the parietal cortex), suggesting that higher values supported classification in the AN group. These results were consistent with the univariate results, which found larger effect sizes for CT reductions in parietal regions and higher CT in the AN group in the entorhinal cortex and temporal pole. Similarly, the frequency of infranormal z-scores for CT in parietal regions was significantly elevated in AN compared to the normative reference. SA metrics were not among the most important features for classification, and compatibly, the reductions in univariate analyses had smaller effect sizes. SV had somewhat in-between negative effect sizes in univariate analyses, and the frequence of infranormal z-scores was elevated for nearly all metrics, but SV features contributed less to classification compared to CT metrics. While the aim of brain morphology–based AN versus HC classification was not to propose a substitute for BMI in routine diagnosis, using brain measures might become relevant in situations where differential diagnosis is challenging, such as atypical presentations [67], apparent absence of other AN criteria due to poor insight or treatment ambivalence [68], or comorbid medical conditions [69,70]. For this reason, we believe that future machine-learning studies would benefit from explicitly including underweight HC or individuals with low BMI due to nonpsychiatric causes, as this would allow a more direct assessment of the incremental value of brain morphology beyond BMI. However, assembling such samples requires careful clinical characterization to exclude eating disorders and to account for somatic causes of low weight [71]. Compared to other multicentric studies in psychiatric populations, the classification performance achieved here (ROC-AUC[LSsO] = 0.75, ROC-AUC[ComBat-GAM] = 0.81) was notably higher than that reported by the ENIGMA Bipolar Disorder group [48] using similar cross-validation methods (ROC-AUC[LSsO] = 0.61), and also exceeded performance in major depressive disorder, where accuracies below 60% were reported using brain morphology data [72], but was comparable to results reported for schizophrenia [73]. Except for Arold and colleagues [5], previous machine learning studies in AN were based on single-site data with fewer than 50 participants per group, yet also reported high classification accuracy [74–78]. Overall, these findings suggest that brain alterations in underweight individuals with AN can be detected at the individual level, despite substantial variability—both across imaging sites (e.g., scanner differences and protocol variations) and across individuals (e.g., genetic and environmental factors). However, whether such alterations differ meaningfully between clinical subtypes of AN remains unclear. In this respect, this study investigated structural brain differences between the restricting (AN-R) and binge-eating/purging (AN-BP) subtypes of AN using three complementary analysis approaches. After controlling for BMI, no significant differences emerged between subtypes in metrics of CT, SA, or SV—neither in conventional univariate analyses nor in the frequency of supranormal and infranormal z-scores derived from the CentileBrain normative model [36]. Furthermore, it was not possible to classify individuals from these subgroups using a machine learning algorithm. These findings suggest that AN-R and AN-BP may share core neurobiological features that are not distinguishable using structural MRI. While the clinical distinction between these subtypes remains relevant—given documented differences in emotion regulation [79], treatment outcome [14], and medical complications [80–84]—frequent transitions between subtypes, particularly from AN-R to AN-BP and often accompanied by clinical worsening [11] have led some researchers to propose that AN-R may represent a transitional phase rather than a stable subtype [13,19,51,52]. From this perspective, the absence of marked structural brain differences observed in this study may reflect the temporal fluidity of subtype expression, such that individuals classified as AN-R and AN-BP do not constitute biologically distinct groups but rather represent different points along a shared disease continuum. To better capture the biological heterogeneity within AN, future studies should consider data-driven approaches—such as unsupervised machine learning—to identify subgroups that transcend traditional symptom-based classifications and may better reflect treatment outcomes or long-term illness trajectories [85]. These results should be interpreted in light of several limitations. First, we did not assess or control for psychiatric comorbidities (e.g., obsessive-compulsive disorder, depression). However, previous large-scale studies have reported relatively small effects of these conditions on brain structure [24,86,87], making it unlikely that our findings in the AN versus HC comparison are primarily driven by comorbidities. Second, the present design does not allow a definitive disambiguation between neurobiological differences related to AN and effects related to starvation or low BMI, as BMI was not included as a covariate in the primary AN versus HC analyses due to its strong confounding with diagnostic group membership. In this respect, a recent study comparing underweight HC to individuals with AN found both alterations related to underweight and alterations specific to AN [7]. Further insights may come from studies including individuals recovered from anorexia nervosa, in whom the effects of acute starvation are expected to be minimal [40,41,88]. Third, because the ethnoracial/ancestry composition of the contributing cohorts could not be harmonized and is likely skewed toward participants of European ancestry, the generalizability of our findings to other ethnoracial groups remains uncertain. Fourth, we compared two relatively simple classification algorithms—support vector machines and shallow neural networks—for distinguishing individuals with AN-R from AN-BP and observed no better-than-chance performance. It remains possible that more advanced models, larger datasets, or the use of raw neuroimaging data as input could yield different results. Finally, distinctions between AN subtypes might emerge more clearly by investigating recovery dynamics longitudinally or when employing other imaging modalities, such as diffusion tensor imaging or functional MRI, which may capture dimensions of brain function or connectivity not reflected in structural MRI. We confirmed large and widespread reductions in GM in AN, most pronounced in CT and SV, while SA was less affected. Analysis of extreme deviations from the normative reference indicated a higher frequency of extreme z-scores, and a higher degree of heterogeneity in AN. In line with this, machine learning models successfully distinguished individuals with AN from HCs, with CT metrics contributing more strongly to classification than SV metrics—even though univariate analyses also revealed significant SV reductions. No significant differences between AN subtypes were found in any structural metric or in the frequency of extreme deviations, and machine learning classifiers could not distinguish between AN-R and AN-BP above chance level. To better understand heterogeneity in AN, future studies should apply data-driven techniques—such as unsupervised machine learning—to identify biologically meaningful subgroups or psychopathological dimensions linked to treatment outcomes. Supporting information S1 Appendix. Contains Supplementary Methods, Supplementary Results, Supplementary References, Supplementary Tables A–G, and Supplementary Figs A–M. https://doi.org/10.1371/journal.pmed.1004809.s001 (DOCX) S1. STROBE Checklist. STROBE checklist for case-control studies. Checklist reproduced from the STROBE Statement (https://www.strobe-statement.org/; von Elm and colleagues, PLoS Med. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296) under the Creative Commons Attribution 4.0 International License (CC BY 4.0). https://doi.org/10.1371/journal.pmed.1004809.s002 (DOC) References - 1. Steinhausen H-C. Outcome of eating disorders. Child Adolesc Psychiatr Clin N Am. 2009;18(1):225–42. pmid:19014869 - 2. Kaye W. Neurobiology of anorexia and bulimia nervosa. Physiol Behav. 2008;94(1):121–35. pmid:18164737 - 3. Crow SJ. Pharmacologic treatment of eating disorders. Psychiatr Clin North Am. 2019;42:253–62. - 4. Vall E, Wade TD. Predictors of treatment outcome in individuals with eating disorders: A systematic review and meta-analysis. Int J Eat Disord. 2015;48(7):946–71. pmid:26171853 - 5. Arold D, Bernardoni F, Geisler D, Doose A, Uen V, Boehm I, et al. Predicting long-term outcome in anorexia nervosa: a machine learning analysis of brain structure at different stages of weight recovery. Psychol Med. 2023;53(16):7827–36. pmid:37554008 - 6. Walton E, Bernardoni F, Batury V-L, Bahnsen K, Larivière S, Abbate-Daga G, et al. Brain structure in acutely underweight and partially weight-restored individuals with Anorexia Nervosa: a coordinated analysis by the ENIGMA eating disorders working group. Biol Psychiatry. 2022;92(9):730–8. pmid:36031441 - 7. Wang Y, Wang Q, Chan RCK, Wang Y, Lei X, Qiu Q. Distinct structural alterations in cortical and subcortical regions in females with acute anorexia nervosa: a cross‐sectional MRI study with BMI ‐matched healthy controls. Int J Eat Disord. 2025. - 8. Moreau CA, Ayrolles A, Ching CRK, Bonicel R, Mathieu A, Stordeur C, et al. Neuroimaging insights into brain mechanisms of early-onset restrictive eating disorders. Nat Ment Health. 2025;3(7):780–8. pmid:40655158 - 9. Steegers C, Blok E, Lamballais S, Jaddoe V, Bernardoni F, Vernooij M, et al. The association between body mass index and brain morphology in children: a population-based study. Brain Struct Funct. 2021;226(3):787–800. pmid:33484342 - 10. Falkai P, Wittchen H-U, Döpfner M, Gaebel W, Maier W, Rief W. Diagnostisches und statistisches manual psychischer störungen dsm-5. 2nd ed. Hogrefe. 2018. - 11. Serra R, Di Nicolantonio C, Di Febo R, De Crescenzo F, Vanderlinden J, Vrieze E, et al. The transition from restrictive anorexia nervosa to binging and purging: a systematic review and meta-analysis. Eat Weight Disord. 2022;27(3):857–65. pmid:34091875 - 12. Wonderlich SA, Joiner TE Jr, Keel PK, Williamson DA, Crosby RD. Eating disorder diagnoses: empirical approaches to classification. Am Psychol. 2007;62(3):167–80. pmid:17469896 - 13. Robert-McComb JJ, Albracht KD, Gary A. The physiology of Anorexia Nervosa and Bulimia Nervosa. In: The Active Female. Springer New York; 2013. p. 149–76. - 14. Miskovic-Wheatley J, Bryant E, Ong SH, Vatter S, Le A, National Eating Disorder Research Consortium, et al. Eating disorder outcomes: findings from a rapid review of over a decade of research. J Eat Disord. 2023;11(1):85. pmid:37254202 - 15. Bulik CM, Thornton L, Pinheiro AP, Plotnicov K, Klump KL, Brandt H. Suicide attempts in Anorexia Nervosa. Psychosom Med. 2008;70:378–83. - 16. Valenciano-Mendoza E, Fernández-Aranda F, Granero R, Vintró-Alcaraz C, Mora-Maltas B, Valero-Solís S, et al. Common and differential risk factors behind suicidal behavior in patients with impulsivity-related disorders: the case of bulimic spectrum eating disorders and gambling disorder. J Behav Addict. 2022;11(4):963–78. pmid:36287739 - 17. Joiner TE, Robison M, McClanahan S, Riddle M, Manwaring J, Rienecke RD, et al. Eating disorder behaviors as predictors of suicidal ideation among people with an eating disorder. Int J Eat Disord. 2022;55(10):1352–60. pmid:35792367 - 18. Zerwas S, Lund BC, Von Holle A, Thornton LM, Berrettini WH, Brandt H, et al. Factors associated with recovery from anorexia nervosa. J Psychiatr Res. 2013;47(7):972–9. pmid:23535032 - 19. Peat C, Mitchell JE, Hoek HW, Wonderlich SA. Validity and utility of subtyping anorexia nervosa. Int J Eat Disord. 2009;42(7):590–4. pmid:19598270 - 20. Zipfel S, Sammet I, Rapps N, Herzog W, Herpertz S, Martens U. Gastrointestinal disturbances in eating disorders: clinical and neurobiological aspects. Auton Neurosci. 2006;129(1–2):99–106. pmid:16962383 - 21. Caldiroli A, Affaticati LM, Coloccini S, Manzo F, Scalia A, Capuzzi E, et al. Clinical factors associated with binge-eating episodes or purging behaviors in patients affected by eating disorders: a cross-sectional study. J Pers Med. 2024;14(6):609. pmid:38929830 - 22. Dechartres A, Huas C, Godart N, Pousset M, Pham A, Divac SM, et al. Outcomes of empirical eating disorder phenotypes in a clinical female sample: results from a latent class analysis. Psychopathology. 2011;44(1):12–20. pmid:20980783 - 23. Yonezawa H, Otagaki Y, Miyake Y, Okamoto Y, Yamawaki S. No differences are seen in the regional cerebral blood flow in the restricting type of anorexia nervosa compared with the binge eating/purging type. Psychiatry Clin Neurosci. 2008;62(1):26–33. pmid:18289138 - 24. Boedhoe PSW, Heymans MW, Schmaal L, Abe Y, Alonso P, Ameis SH, et al. An empirical comparison of meta- and mega-analysis with data from the ENIGMA obsessive-compulsive disorder working group. Front Neuroinformatics. 2019;12:102. - 25. Koshiyama D, Fukunaga M, Okada N, Morita K, Nemoto K, Usui K, et al. White matter microstructural alterations across four major psychiatric disorders: mega-analysis study in 2937 individuals. Mol Psychiatry. 2020;25(4):883–95. pmid:31780770 - 26. Kochunov P, Jahanshad N, Sprooten E, Nichols TE, Mandl RC, Almasy L, et al. Multi-site study of additive genetic effects on fractional anisotropy of cerebral white matter: comparing meta and megaanalytical approaches for data pooling. Neuroimage. 2014;95:136–50. pmid:24657781 - 27. Zugman A, Harrewijn A, Cardinale EM, Zwiebel H, Freitag GF, Werwath KE, et al. Mega-analysis methods in ENIGMA: the experience of the generalized anxiety disorder working group. Hum Brain Mapp. 2022;43(1):255–77. pmid:32596977 - 28. Johnson WE, Li C, Rabinovic A. Adjusting batch effects in microarray expression data using empirical Bayes methods. Biostatistics. 2007;8(1):118–27. pmid:16632515 - 29. Radua J, Vieta E, Shinohara R, Kochunov P, Quidé Y, Green MJ, et al. Increased power by harmonizing structural MRI site differences with the ComBat batch adjustment method in ENIGMA. Neuroimage. 2020;218:116956. pmid:32470572 - 30. Fried EI, Nesse RM. Depression is not a consistent syndrome: an investigation of unique symptom patterns in the STAR*D study. J Affect Disord. 2015;172:96–102. pmid:25451401 - 31. Young G, Lareau C, Pierre B. One quintillion ways to have PTSD comorbidity: recommendations for the disordered DSM-5. Psychol Inj Law. 2014;7(1):61–74. - 32. Wolfers T, Doan NT, Kaufmann T, Alnæs D, Moberget T, Agartz I, et al. Mapping the heterogeneous phenotype of schizophrenia and bipolar disorder using normative models. JAMA Psychiatry. 2018;75(11):1146–55. pmid:30304337 - 33. Lv J, Di Biase M, Cash RFH, Cocchi L, Cropley VL, Klauser P, et al. Individual deviations from normative models of brain structure in a large cross-sectional schizophrenia cohort. Mol Psychiatry. 2021;26(7):3512–23. pmid:32963336 - 34. Floris DL, Wolfers T, Zabihi M, Holz NE, Zwiers MP, Charman T, et al. Atypical brain asymmetry in autism-A candidate for clinically meaningful stratification. Biol Psychiatry Cogn Neurosci Neuroimaging. 2021;6(8):802–12. pmid:33097470 - 35. Wolfers T, Beckmann CF, Hoogman M, Buitelaar JK, Franke B, Marquand AF. Individual differences v. the average patient: mapping the heterogeneity in ADHD using normative models. Psychol Med. 2020;50(2):314–23. pmid:30782224 - 36. Ge R, Yu Y, Qi YX, Fan Y-N, Chen S, Gao C, et al. Normative modelling of brain morphometry across the lifespan with CentileBrain: algorithm benchmarking and model optimisation. Lancet Digit Health. 2024;6(3):e211–21. pmid:38395541 - 37. van Eeden AE, van Hoeken D, Hoek HW. Incidence, prevalence and mortality of anorexia nervosa and bulimia nervosa. Curr Opin Psychiatry. 2021;34(6):515–24. pmid:34419970 - 38. Desikan RS, Ségonne F, Fischl B, Quinn BT, Dickerson BC, Blacker D, et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage. 2006;31(3):968–80. pmid:16530430 - 39. Pomponio R, Erus G, Habes M, Doshi J, Srinivasan D, Mamourian E, et al. Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan. Neuroimage. 2020;208:116450. pmid:31821869 - 40. Bahnsen K, Bernardoni F, King JA, Geisler D, Weidner K, Roessner V, et al. Dynamic Structural brain changes in Anorexia Nervosa: a replication study, mega-analysis, and virtual histology approach. J Am Acad Child Adolesc Psychiatry. 2022;61(9):1168–81. pmid:35390458 - 41. Bernardoni F, King JA, Geisler D, Stein E, Jaite C, Nätsch D, et al. Weight restoration therapy rapidly reverses cortical thinning in anorexia nervosa: a longitudinal study. Neuroimage. 2016;130:214–22. pmid:26876474 - 42. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Ser B Methodol. 1995;57(1):289–300. - 43. ENIGMA Clinical High Risk for Psychosis Working Group, Allen P, Baldwin H, Bartholomeusz CF, Chee MW, Chen X, et al. Normative modeling of brain morphometry in clinical high risk for psychosis. JAMA Psychiatry. 2024;81:77. - 44. Belov V, Erwin-Grabner T, Aghajani M, Aleman A, Amod AR, Basgoze Z, et al. Multi-site benchmark classification of major depressive disorder using machine learning on cortical and subcortical measures. Sci Rep. 2024;14(1):1084. pmid:38212349 - 45. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O. Scikit-learn: machine learning in Python; 2012. - 46. Snoek L, Miletić S, Scholte HS. How to control for confounds in decoding analyses of neuroimaging data. Neuroimage. 2019;184:741–60. pmid:30268846 - 47. Saito T, Rehmsmeier M. The precision-recall plot is more informative than the ROC plot when evaluating binary classifiers on imbalanced datasets. PLoS One. 2015;10(3):e0118432. pmid:25738806 - 48. Nunes A, Schnack HG, Ching CRK, Agartz I, Akudjedu TN, Alda M, et al. Using structural MRI to identify bipolar disorders — 13 site machine learning study in 3020 individuals from the ENIGMA Bipolar Disorders Working Group. Mol Psychiatry. 2020;25(9):2130–43. pmid:30171211 - 49. Dinga R, Schmaal L, Penninx BWJH, Veltman DJ, Marquand AF. Controlling for effects of confounding variables on machine learning predictions. Bioinformatics. 2020. - 50. Haufe S, Meinecke F, Görgen K, Dähne S, Haynes J-D, Blankertz B, et al. On the interpretation of weight vectors of linear models in multivariate neuroimaging. Neuroimage. 2014;87:96–110. pmid:24239590 - 51. Eddy KT, Dorer DJ, Franko DL, Tahilani K, Thompson-Brenner H, Herzog DB. Diagnostic crossover in anorexia nervosa and bulimia nervosa: implications for DSM-V. Am J Psychiatry. 2008;165(2):245–50. pmid:18198267 - 52. Eddy KT, Keel PK, Dorer DJ, Delinsky SS, Franko DL, Herzog DB. Longitudinal comparison of anorexia nervosa subtypes. Int J Eat Disord. 2002;31(2):191–201. pmid:11920980 - 53. King JA, Frank GKW, Thompson PM, Ehrlich S. Structural neuroimaging of anorexia nervosa: future directions in the quest for mechanisms underlying dynamic alterations. Biol Psychiatry. 2018;83(3):224–34. pmid:28967386 - 54. Seitz J, Bühren K, Von Polier GG, Heussen N, Herpertz-Dahlmann B, Konrad K. Morphological changes in the brain of acutely ill and weight-recovered patients with anorexia nervosa: a meta-analysis and qualitative review. Z Für Kinder- Jugendpsychiatrie Psychother. 2014;42:7–18. - 55. Dreier MJ, Van De Water AL, Kahn DL, Becker KR, Eddy KT, Thomas JJ, et al. Meta-analysis of structural MRI studies in anorexia nervosa and the role of recovery: a systematic review protocol. Syst Rev. 2021;10(1):247. pmid:34517926 - 56. Su T, Gong J, Tang G, Qiu S, Chen P, Chen G, et al. Structural and functional brain alterations in anorexia nervosa: a multimodal meta-analysis of neuroimaging studies. Hum Brain Mapp. 2021;42(15):5154–69. pmid:34296492 - 57. Wong AP-Y, French L, Leonard G, Perron M, Pike GB, Richer L. Inter-regional variations in gene expression and age-related cortical thinning in the adolescent brain. Cereb Cortex. 2018;28:1272–81. - 58. Tamnes CK, Herting MM, Goddings A-L, Meuwese R, Blakemore S-J, Dahl RE, et al. Development of the cerebral cortex across adolescence: a multisample study of inter-related longitudinal changes in cortical volume, surface area, and thickness. J Neurosci. 2017;37(12):3402–12. pmid:28242797 - 59. Gaudio S, Wiemerslage L, Brooks SJ, Schiöth HB. A systematic review of resting-state functional-MRI studies in anorexia nervosa: evidence for functional connectivity impairment in cognitive control and visuospatial and body-signal integration. Neurosci Biobehav Rev. 2016;71:578–89. pmid:27725172 - 60. Brizzi G, Sansoni M, Di Lernia D, Frisone F, Tuena C, Riva G. The multisensory mind: a systematic review of multisensory integration processing in Anorexia and Bulimia Nervosa. J Eat Disord. 2023;11(1):204. pmid:37974266 - 61. Khalighinejad B, Patel P, Herrero JL, Bickel S, Mehta AD, Mesgarani N. Functional characterization of human Heschl’s gyrus in response to natural speech. Neuroimage. 2021;235:118003. pmid:33789135 - 62. Guillery RW, Sherman SM. Thalamic relay functions and their role in corticocortical communication. Neuron. 2002;33(2):163–75. - 63. Wronski M-L, Gronow F, Schlömer J, Bernardoni F, Geisler D, Doose A, et al. Structural alterations of thalamic nuclei and their associations with leptin levels in patients with anorexia nervosa. Prog Neuropsychopharmacol Biol Psychiatry. 2025;136:111248. pmid:39761816 - 64. Herlin B, Navarro V, Dupont S. The temporal pole: from anatomy to function-A literature appraisal. J Chem Neuroanat. 2021;113:101925. pmid:33582250 - 65. Ekstrom AD, Hill PF. Spatial navigation and memory: a review of the similarities and differences relevant to brain models and age. Neuron. 2023;111(7):1037–49. pmid:37023709 - 66. Haukvik UK, Wolfers T, Tesli N, Bell C, Hjell G, Fischer-Vieler T, et al. Individual-level deviations from normative brain morphology in violence, psychosis, and psychopathy. Transl Psychiatry. 2025;15(1):118. pmid:40175325 - 67. Vo M, Golden N. Medical complications and management of atypical anorexia nervosa. J Eat Disord. 2022;10(1):196. pmid:36522787 - 68. Gregertsen EC, Mandy W, Serpell L. The egosyntonic nature of anorexia: an impediment to recovery in anorexia nervosa treatment. Front Psychol. 2017;8:2273. pmid:29312100 - 69. Perera LAM, Chopra A, Shaw AL. Approach to patients with unintentional weight loss. Med Clin North Am. 2021;105(1):175–86. pmid:33246517 - 70. Treasure J, Duarte TA, Schmidt U. Eating disorders. Lancet. 2020;395(10227):899–911. pmid:32171414 - 71. Bailly M, Thivel D, Isacco L, Verney J. Unique energy profile associated with persistent thin phenotype. Annu Rev Nutr. 2025;45(1):65–91. pmid:40327537 - 72. Winter NR, Blanke J, Leenings R, Ernsting J, Fisch L, Sarink K, et al. A systematic evaluation of machine learning-based biomarkers for major depressive disorder. JAMA Psychiatry. 2024;81(4):386–95. pmid:38198165 - 73. Rozycki M, Satterthwaite TD, Koutsouleris N, Erus G, Doshi J, Wolf DH, et al. Multisite machine learning analysis provides a robust structural imaging signature of schizophrenia detectable across diverse patient populations and within individuals. Schizophr Bull. 2018;44(5):1035–44. pmid:29186619 - 74. Lavagnino L, Amianto F, Mwangi B, D’Agata F, Spalatro A, Zunta-Soares GB, et al. Identifying neuroanatomical signatures of anorexia nervosa: a multivariate machine learning approach. Psychol Med. 2015;45(13):2805–12. pmid:25990697 - 75. Lavagnino L, Mwangi B, Cao B, Shott ME, Soares JC, Frank GKW. Cortical thickness patterns as state biomarker of anorexia nervosa. Int J Eat Disord. 2018;51(3):241–9. pmid:29412456 - 76. Cerasa A, Castiglioni I, Salvatore C, Funaro A, Martino I, Alfano S, et al. Biomarkers of eating disorders using support vector machine analysis of structural neuroimaging data: preliminary results. Behav Neurol. 2015;2015:924814. pmid:26648660 - 77. McClure Z, Fuller-Tyszkiewicz M, Messer M, Linardon J. Machine-learning applications in eating-disorder-outcome prediction: a systematic scoping review. Clin Psychol Sci. 2025. - 78. Monaco F, Vignapiano A, Di Gruttola B, Landi S, Panarello E, Malvone R, et al. Neuroimaging and machine learning in eating disorders: a systematic review. Eat Weight Disord. 2025;30(1):46. pmid:40450619 - 79. Weinbach N, Sher H, Bohon C. Differences in emotion regulation difficulties across types of eating disorders during adolescence. J Abnorm Child Psychol. 2018;46(6):1351–8. pmid:29101588 - 80. Gibson D, Workman C, Mehler PS. Medical complications of anorexia nervosa and bulimia nervosa. Psychiatr Clin North Am. 2019;42:263–74. - 81. Mehler PS, Brown C. Anorexia nervosa − medical complications. J Eat Disord. 2015;3:11. pmid:25834735 - 82. Mehler PS, Anderson K, Bauschka M, Cost J, Farooq A. Emergency room presentations of people with anorexia nervosa. J Eat Disord. 2023;11(1):16. pmid:36759897 - 83. Douglas L. Caring for dental patients with eating disorders. BDJ Team. 2015;1(1). - 84. Forney KJ, Buchman-Schmitt JM, Keel PK, Frank GKW. The medical complications associated with purging. Int J Eat Disord. 2016;49(3):249–59. pmid:26876429 - 85. Sun J, Lu T, Shao X, Han Y, Xia Y, Zheng Y, et al. Practical AI application in psychiatry: historical review and future directions. Mol Psychiatry. 2025;30(9):4399–408. pmid:40456953 - 86. Thompson PM, Jahanshad N, Ching CRK, Salminen LE, Thomopoulos SI, Bright J, et al. ENIGMA and global neuroscience: a decade of large-scale studies of the brain in health and disease across more than 40 countries. Transl Psychiatry. 2020;10(1):100. pmid:32198361 - 87. Schmaal L, Hibar DP, Sämann PG, Hall GB, Baune BT, Jahanshad N, et al. Cortical abnormalities in adults and adolescents with major depression based on brain scans from 20 cohorts worldwide in the ENIGMA Major Depressive Disorder Working Group. Mol Psychiatry. 2017;22(6):900–9. pmid:27137745 - 88. King JA, Geisler D, Ritschel F, Boehm I, Seidel M, Roschinski B, et al. Global cortical thinning in acute anorexia nervosa normalizes following long-term weight restoration. Biol Psychiatry. 2015;77(7):624–32. pmid:25433902