Main

Early-onset restrictive eating disorders (rEO-ED) in children encompass a heterogeneous group of conditions, including early-onset anorexia nervosa (EO-AN) and avoidant restrictive food intake disorder (ARFID). Both disorders are characterized by severe food restriction and significant weight loss but differ in the primary drivers of restrictive behavior1,2. Pediatrics national surveillance studies have reported an increasing prevalence of rEO-ED, ranging from 1.4 to 3 per 100,000, with an incidence doubling over the past decade3,4,5,6.

The early-onset form of anorexia nervosa (EO-AN) (symptom onset at <13 years old) has been associated with greater detrimental impacts on bone development, statural growth and puberty, as well as faster weight loss compared with patients with typical onset anorexia nervosa (TO-AN)7. Patients with EO-AN show distinct epidemiological and clinical features compared with TO-AN, including a more balanced sex ratio (three boys for seven girls versus one for nine in TO-AN), a higher frequency of the restrictive subtype (rare binge–purge symptoms), a lower incidence rate and more rapid weight loss7. Despite these differences with TO-AN, very little is known about the mechanisms of EO-AN and its impact on brain development. The most extensive study so far of TO-AN was conducted by the Enhancing Neuro Imaging through Meta-Analysis (ENIGMA) Eating Disorders working group, which included 685 anorexia nervosa (AN) and 963 individuals with typical development (TD) (mean age at scan of 21 (15–27) years with a mean illness duration of 5 (1–13) years)8, and revealed a widespread pattern of lower cortical thickness (CT) and smaller thalamic volumes in TO-AN, relative to TD. Brain alterations were less pronounced in weight-restored individuals, suggesting a mediating effect of body mass index (BMI) on brain structure in TO-AN8. Brain charts have been recently developed for cortical morphometric measures in unselected populations9 and revealed how the age of developmental peak varies across regions: from the early developing areas (visual and parietal areas, before age 5 years) to the later-developing areas (frontal, temporal and finally the insula, up to 10 years old). As the brain is in a critical development phase at this age (9–13 years), we hypothesized that the EO-AN group would show greater abnormalities in brain structure than the TO-AN group, especially in frontotemporal areas.

The extent to which brain alterations in AN are driven by poor nutrition and low BMI or represent distinct AN-related pathophysiological processes remains largely unknown. Including BMI as a covariate when comparing rEO-ED and TD is not optimal due to its high colinearity with the diagnosis. One way to tackle this question would be to investigate the effect of low BMI in individuals with restrictive dietary behaviors but without AN. ARFID is a rEO-ED condition characterized by a limitation in the amount or type of food consumed that is not motivated by body image concerns or fear of gaining weight10. Three clinical subgroups have been described: lack of food interest, fear of aversive consequences and sensory sensitivity impairment11. Prevalence estimates are elevated and vary across studies up to 15% in the general population, including both selective and restrictive forms. The prevalence of restrictive forms remains unclear12. As a substantial proportion of patients with ARFID are severely underweight (BMI <3rd percentile), comparing patients with ARFID with EO-AN could provide a way of disentangling brain alterations driven primarily by low BMI and those resulting from EO-AN pathophysiology. So far, no studies have evaluated the impact of ARFID diagnosis on brain morphometry, therefore, our hypotheses regarding ARFID will be exploratory. However, one population-based study investigated the effect of ARFID symptoms on brain structure using a hypothesis-driven approach on six regions and found greater CT in the frontal lobes in participants with ARFID symptoms compared with the general population13. No differences remained significant when including all brain regions.

Clinical observations (overlapping symptoms and comorbidities) and genetic analyses (genetic correlations of risk variants) have highlighted similarities between AN and other serious mental illnesses14,15. Almost one-fifth of patients with AN meet the criteria for obsessive–compulsive disorder (OCD)15,16. Overlapping symptoms and comorbidities with autism spectrum disorder (ASD)17,18, major depression disorder and anxiety disorders are also commonly observed in AN14,16. At the genetic risk level, variants associated with AN derived from genome-wide association studies (GWAS) show significant correlations with variants associated with OCD (rG = 0.45) and major depression disorder (rG = 0.27)15. Similarly, clinical observations in ARFID have revealed an association with anxiety disorders and neurodevelopmental disorders (NDD), including ASD and attention deficit–hyperactivity disorder (ADHD)1,12,19. Recent GWAS have revealed a high heritability of ARFID20,21, but nothing is known so far about similarities with other psychiatric conditions at the genetic level. The extent to which rEO-ED displays overlapping or distinct brain signatures compared with other clinically and genetically associated mental illnesses is currently unknown.

We investigated three main aims:

  1. (1)

    What structural brain differences distinguish EO-AN from TD, and to what extent does the age of onset of AN influence brain morphometry? Hypothesis: individuals with EO-AN will show a widespread pattern of thinner cortex in later-developing brain regions compared with TD. EO-AN will exhibit greater differences than patients with TO-AN8.

  2. (2)

    Do structural brain profiles differ between EO-AN and ARFID in patients with similar BMI distributions? How does BMI relate to the degree of brain variation in EO-AN? And what structural brain traits differentiate ARFID from TD? Hypothesis: we expect to observe a partial overlap between EO-AN and ARFID. This shared brain pattern may reflect a general consequence of low BMI on brain morphometry.

  3. (3)

    To what extent do brain alterations in rEO-ED overlap with other neurodevelopmental disorders? Hypothesis: brain abnormalities in rEO-ED will significantly overlap with brain profiles previously associated with clinically similar conditions. More specifically, we expect to find similarities between EO-AN and OCD and EO-AN and ASD, as well as between ARFID and ASD, and between ARFID and ADHD.

To this end, we collected and analyzed brain imaging data in children and young adolescents with an early-onset form of AN (<13 years) (n = 124), ARFID (n = 50) as well well as TD (n = 116).

Results

Sample characteristics

Our analysis included 290 participants (124 EO-AN, 50 ARFID and 116 TD) (Table 1). As expected, EO-AN and ARFID had a significantly lower mean z-scored BMI compared with TD (BMI of −1.81 (95% confidence interval (CI) −4.1 to 0.61) for EO-AN, −2.20 (95% CI −4.2 to 0.38) for ARFID and 0.39 (95% CI −2.27 to 2.5) for TD, P = 2 × 10−16. rEO-EDs were slightly older than TD (EO-AN 11.5 ± 1.19 years, ARFID 11.2 ± 1.49 years and TD 10.2 ± 2.04 years, P = 9.8 × 10−8), with no age differences between EO-AN and ARFID (P = 0.6). As expected, there were differences in sex distribution (106 F/18 M for EO-AN, 38 F/12 M for ARFID and 59 F/57 M for TD, P = 2.2 × 10−8). Sensitivity analyses conducted on female individuals only showed no differences compared with the results obtained from the whole dataset (Supplementary Fig. 3).

Table 1 Demographic information of participants included in this study
Full size table

Thinner cortex and smaller SVs in EO-AN > TD

Global effects

A widespread global pattern of thinner cortex (Cohen’s d = −0.86, Padj = 1.1 × 10−9), lower total gray matter volume (GV) (Cohen’s d = −0.58 Padj = 1.9 × 10−0.5) and cerebrospinal fluid (CSF) volume increase (Cohen’s d = 0.75, Padj = 5.9 × 10−8) was observed in patients with EO-AN compared with TD (Fig. 1a). We found no significant effect on the total surface area (SA) (Cohen’s d = −0.18, NS) or the intracranial volume (ICV) (Cohen’s d = 0.11, NS) in patients with EO-AN compared with TD. After accounting for BMI as a mediator, the EO-AN diagnosis was still associated with a lower CT (β = −0.781, P = 0.002) and a larger CSF volume (β = 0.668, P = 0.001), but no association remained significant with GV (Supplementary Table 37 and Supplementary Fig. 4). To assess the robustness of our findings, additional sensitivity analyses were conducted on subgroups of patients: (1) female individuals only (n = 106 EO-AN versus n = 57 TD), (2) participants scanned exclusively on a 3T MRI scanner (50 EO-AN versus 38 TD), (3) controls without a family history of psychiatric disorder (124 EO-AN versus 98 TD) and (4) participants without psychotropic medication (EO-AN (n = 101) versus TD (n = 116). The results remained consistent across all analyses (Supplementary Results and Supplementary Tables 26–31).

Fig. 1: Characterization of the effect of rEO-EDs on brain structure.
figure 1

a, The effect of EO-AN (dark red dots, n = 124) and ARFID (yellow dots, n = 50) compared with TD (n = 116) on five global brain metrics (ICV, total SA, mean CT, CSF and total GV). The triangles represent significant effect sizes. Data are presented as mean values ±s.d. b,c, Brain maps (FDR thresholded) showing Cohen’s d values for each of the 68 cortical regions (thickness) and the 14 subcortical regions and ventricles (volumes) for EO-AN versus TD (b) and ARFID versus TD (c).

Full size image

Regional effects

A widespread pattern of lower cortical thickness was identified, with a stronger contribution observed in the parietal and occipital lobes bilaterally. Fifty-nine regions showed significant thickness reduction (Fig. 1b and Supplementary Table 1). The regions with the higher effect sizes indicating thinner cortex included the left superior parietal cortex (Cohen’s d = −1.16, Padj = 8.9 × 10−15) and the precuneus (Cohen’s d = −1.01, Padj = 5.4 × 10−13) along with the right lingual (Cohen’s d = −1.01, Padj = 3.3 × 10−12). No significant regional alterations were found for SA (Supplementary Table 2). The results on CT and SA adjusted for global metrics are reported in the Supplementary Results (Supplementary Fig. 6a and Supplementary Tables 4 and 5). At the subcortical level, analyses comparing EO-AN and TD demonstrated smaller bilateral thalamic volumes (Cohen’s d = −0.71, Padj = 1.2 × 10−6 and Cohen’s d = −0.79, Padj = 1.1 × 10−7 for the left and right side, respectively), bilateral amygdala (Cohen’s d = −0.49, Padj = 7.7 × 10−4 and Cohen’s d = −0.41, Padj = 5.1 × 10−3 for the left and right side, respectively), bilateral pallidum (Cohen’s d = −0.38, Padj = 0.009 and Cohen’s d = −0.31, Padj = 0.03 for the left and right side, respectively) and bilateral putamen (Cohen’s d = −0.37, Padj = 0.009 and Cohen’s d = −0.36, Padj = 0.01 for the left and right side, respectively). We also observed larger bilateral ventricle volumes (Cohen’s d = 0.53, Padj = 3.3 × 10−4 and Cohen’s d = 0.50, Padj = 5.9 × 10−4 for the left and right side, respectively) (Fig. 1b and Supplementary Table 3). After accounting for BMI as a mediator, the EO-AN diagnosis remained significantly associated with lower CT in 32 regions, with the greatest effect observed in the superior parietal (Cohen’s d = −1.07, Padj < 0.0001). However, no subcortical volume (SV) differences remained significant (Supplementary Fig. 5 and Supplementary Tables 38 and 40).

We found an even stronger effect on CT (thinner) and SVs but still no effect on SA in the acutely ill subgroup of EO-AN patients (EO-acAN, BMI <3rd percentile), compared with TD (Fig. 2a and Supplementary Fig. 7).

Fig. 2: Partition of BMI versus rEO-ED effects on brain features.
figure 2

a, The effects on CT and SV of EO-acAN > TD. b, The effects on SA and SV of unARFID > TD. c, The effects of the EO-acAN > unARFID on CT, SA and SV. d, Brain maps summarizing the 35 significant correlations between BMI and CT (68 cortical regions) as well as SV (14 subcortical regions + 2 lateral ventricles) in patients with EO-AN. No correlations were significant in patients with ARFID. e, Correlations between residuals (removing the effect of age, sex and scanner) of the thickness of the right middle temporal gyrus and z-scored BMI in individuals with EO-AN (dark red dots) and ARFID (yellow dots). We selected the right middle temporal gyrus as it showed the highest correlation with BMI in patients with EO-AN. Regression lines are shown with shaded gray areas representing 95% CIs. r represents Pearson correlation coefficient.

Full size image

Brain pattern similarities between the EO versus TO types of AN

We found a correlation between EO-AN and TO-AN8 brain profiles at r = 0.78 (95% CI 0.67 to 0.86, P = 3.9 × 10−15) for the thickness of the 68 cortical regions and r = 0.88 (95% CI 0.68 to 0.95, P = 7.5 × 10−6) for the volume of the 16 subcortical regions. The main differences between EO-AN and TO-AN were observed in effect sizes for specific brain regions. For example, the transverse temporal gyrus showed a larger effect in EO-AN > TD (−0.87) compared with TO-AN > TD (−0.19), while the superior frontal gyrus had a greater effect in TO-AN > TD (−0.85) than EO-AN > TD (−0.63) (Supplementary Fig. 8). A comparison of the distributions of regional effect sizes showed a larger effect on CT in EO-AN > TD (absolute mean top decile of 1.06) compared with TO-AN > TD (absolute mean top decile of 0.89) and a larger effect on SVs in EO-AN > TD (absolute mean top decile of 0.79) compared with TO-AN > TD (absolute mean top decile of 0.66). Overall, these differences between EO-AN and TO-AN remain small and suggest a high degree of replication across independent cohorts, as well as signal stability on CT and SV metrics in AN, independent of the age of onset.

Distinct brain patterns associated with ARFID and EO-AN

ARFID was associated with smaller ICV (Cohen’s d = −0.51, Padj = 0.02), total GV (Cohen’s d = −0.48, Padj = 0.02) and total SA (Cohen’s d = −0.39, P = 0.04) compared with TD (Fig. 1a,c and Supplementary Table 12). At the regional level, only one region was significantly different between ARFID and TD: the thickness of the right isthmus of the cingulate gyrus (Cohen’s d = −0.62, Padj = 0.04). After including BMI as a mediator, significant differences between ARFID and TD were no longer observed (Supplementary Table 39). No effect on SA at the regional level was observed in ARFID > TD. When investigating specifically the underweight group of patients with ARFID (unARFID, BMI <3rd percentile), we observed a large reduction of the ICV (Cohen’s d = −0.68, Padj = 0.004), total GV (Cohen’s d = −0.69, Padj = 0.004) and total SA (Cohen’s d = −0.58, Padj = 0.01) (Supplementary Fig. 10). At the regional level, we observed lower SA in seven cortical regions (Fig. 2b and Supplementary Table 19). The top impacted regions were the right lingual gyrus (Cohen’s d = −0.72, Padj = 0.02), left postcentral area (Cohen’s d = −0.71, Padj = 0.02) and the right pericalcarine area (Cohen’s d = −0.71, Padj = 0.02). Finally, we observed a smaller left putamen volume (Cohen’s d = −0.65, Padj = 0.04) in unARFID > TD (Fig. 2b). Additional analyses on the ARFID clinical subtype ‘fear of aversive consequences’ (n = 29) were also performed, but no region survived false discovery rate (FDR) correction (Supplementary Fig. 9 and Supplementary Tables 33 and 34).

To disentangle the effects of the low BMI versus the eating disorders mechanisms, we compared patients with EO-acAN > patients with unARFID (both groups with a BMI percentile <3 at scan). We observed a similar profile to the one observed when comparing EO-acAN versus TD for thickness and volumes (r = 0.85, P < 2.2 × 10−16 and r = 0.82, P = 8.7 × 10−5, respectively; Fig. 2c), with a widespread lower CT and smaller SVs in EO-acAN compared with patients with unARFID. These results were similar when BMI was added as a covariate in the model (Supplementary Fig. 11). This suggests that the unARFID group resembles the TD group at the CT and SV levels. This was not the case for SA (r = 0.34, P = 0.004). We found a larger SA in three regions in patients with EO-acAN compared with patients with unARFID. These analyses demonstrate independent brain structural profiles associated with ARFID and EO-AN diagnoses despite similar BMI distributions.

To better understand the impact of weight recovery on brain features, we performed regional correlations between the residuals of each region of interest (n = 82 regions) and the z-scored BMI distribution at the individual level. The thickness of 35 cortical regions and the volume of three subcortical regions were positively correlated with BMI in subjects with EO-AN (Fig. 2d and Supplementary Table 35), showing that a z-scored BMI closest to 0 s.d. (closest to a normal weight) was associated with greater CT and volume. The highest correlation was observed for the middle temporal gyrus (r = 0.37, Padj = 0.001; Fig. 2e). In contrast, despite similar BMI distributions, we found no significant correlation for the CT or volume of any region in subjects with ARFID (Supplementary Table 35). No correlation between SA and BMI survived FDR correction in EO-AN or ARFID.

rEO-ED versus NDDs similarities at the brain and genetic levels

We extracted and ranked Cohen’s d distributions (Fig. 3a,b) for the brain signatures on CT of three neurodevelopmental conditions (OCD, ASD and ADHD) and compared them with EO-AN. EO-AN had the greatest effect on CT (mean absolute top decile of 1.06), followed by OCD (0.25), ASD (0.21) and ADHD (0.18). Brain-based Pearson correlations (rB) between EO-AN and these three conditions followed by permutation tests showed significant correlations between EO-AN and OCD (rB AN–OCD = 0.59, Pperm = 0.0001). Correlations between EO-AN and ADHD, as well as ASD, were not significant (rB AN–ADHD = −0.14 and rB AN-ASD = −0.03). We also performed similar correlation analyses between brain CT signatures of ARFID and these three conditions and found significant correlations between ARFID and ASD only (rB ARFID–ASD = 0.25, Pperm = 0.025 and rB ARFID–ADHD = −0.07, rB ARFID–OCD = 0.10). No significant correlations were found for SV patterns between EO-AN and ASD, OCD or ADHD. SA measures were not investigated as no significant results were found at the regional level in EO-AN > TD or ARFID > TD.

Fig. 3: Similarities of the EO-AN brain pattern with additional psychiatric conditions.
figure 3

a, Distribution of effect sizes of three psychiatric conditions on regional CT (previously published by the ENIGMA consortium), as well as EO-AN (computed in this study). b, Brain maps representing patterns of abnormalities in CT reported by the ENIGMA consortium for three psychiatric conditions (Methods) and EO-AN. c, The relationship between single-nucleotide polymorphism (SNP)-based correlations (rG, provided by refs. 15,34) and brain-based correlations (rB, computed in this study using cortical thickness). Credit: brain and DNA icons in c, NIAID NIH BIOART.

Full size image

Last, we observed a concordance of 0.94 between brain-based (CT) correlations of EO-AN and NDDs (computed in this paper) and genetic-based correlations of EO-AN and NDDs (previously published) across the three pairs of conditions (Fig. 3c).

Discussion

To the best of our knowledge, we conducted the most extensive neuroimaging study on patients with an rEO-ED. Patients with an early-onset form of AN presented a widespread cortex thinning and smaller subcortical and total GVs (compared with TD), similar to results obtained in patients with a typical onset, suggesting an effect independent of the age of onset. Patients with ARFID presented reduced ICV, total GV and SA. Despite similar BMI distributions, EO-AN and ARFID showed divergent structural patterns, suggesting independent brain mechanisms. Half the regional CT brain measures were correlated with BMI in EO-AN and none in ARFID, indicating a partial mediation of EO-AN signal by BMI. Last, we identified overlapping patterns of CT brain differences between EO-AN and OCD, and between ARFID and ASD, mirroring findings observed at the clinical level. Similar overlap has been previously reported at the genetic level between EO-AN and OCD, pointing toward shared mechanisms underlying both disorders.

Hypothesis 1 predicted that individuals with EO-AN would exhibit a widespread pattern of thinner cortex compared with TD, particularly in later-developing brain regions, such as the anterior cingulate cortex and the insula9. However, our results revealed a general effect of EO-AN on CT without a specific impact on this particular set of regions. The most impacted regions were the left superior parietal cortex and the thalamus, both strongly involved in integrating and processing sensory inputs22. Additionally, the amygdala—a structure involved in emotional response23—was also affected. Such differences could contribute to distorted body image perception and dysregulated emotional responses, which are central symptoms of AN. Furthermore, EO-AN and TO-AN showed highly overlapping brain patterns compared with TD (r = 0.8), indicating similar brain alterations regardless of age of onset. The only notable difference was in the effect’s amplitude: EO-AN showed larger deviations from TD than TO-AN, as well as ADHD, ASD and OCD.

Half of the cortical regions’ thickness correlated with BMI in patients with EO-AN, supporting the existence of a widespread consequence of low BMI on CT, but also a specific effect of EO-AN independent of BMI. This was confirmed by BMI mediation analyses. The underlying mechanisms of brain structural alteration in EO-AN remain unclear. Partial restoration of brain thickness abnormalities following weight partial recovery has been speculated to be associated with hydration level and/or changes in oncotic pressure24. To address these hypotheses, previous studies have investigated gravity of urine samples 1 h before scanning but did not find any abnormality, excluding the likelihood of the dehydration hypothesis25. Serum albumin assessment before scanning was also reported as normal, reducing the likelihood of an increased blood volume after weight partial recovery24. Alternative hypotheses should be investigated, such as differences in metabolic costs of brain regions26 or a direct impact on neuronal membrane thickness, primarily made of lipids. BMI may not be the most appropriate measure for assessing body fat to investigate AN mechanisms. Other metrics, such as leptin levels or body fat derived from MRI data, could provide more accurate measures to disentangle the impact of nutritional status from any independent effect of EO-AN27,28.

Hypothesis 2 predicted that children and adolescents with EO-AN and ARFID would show partially overlapping brain patterns driven by low weight. Despite equivalent BMI distributions in both conditions, EO-AN and ARFID showed differential brain patterns, which suggests independent underlying mechanisms. One hypothesis for the lack of similarities may relate to the age of onset of both conditions. Children with ARFID typically experience a more early and gradual onset of their eating disorder compared with those with EO-AN. This is particularly the case for two out of three clinical subgroups: a lack of interest in food and sensory sensitivity impairment. This slower progression could allow the brain to compensate for chronic low-calorie intake in children with ARFID compared with children with EO-AN29. ARFID was associated with reduced ICV, SA and GV compared with TD, as well as a thinner right isthmus of the cingulate gyrus. This region is part of the default mode network30 and plays a key role in interoception and sensory integration, which could be associated with ARFID-associated symptoms. Surprisingly, we did not observe any association between BMI and any brain metrics. However, once BMI was included as a mediator, all significant differences between ARFID and TD disappeared. This absence of difference may reflect a lack of statistical power rather than an effect of BMI. Notably, no significant associations were found between BMI and brain metrics in the ARFID population, whereas these associations were significant in the EO-AN population (paths b in Supplementary Fig. 4 and Supplementary Tables 38 and 39). This finding further supports the hypothesis that the brain signature of ARFID is not mediated by BMI, in contrast to EO-AN. Our results do not replicate the unique paper exploring the effect of ARFID symptoms in the general population (Generation R)13. This could be due to differences in participant selections. Our study includes ARFID inpatients (diagnosed based on the DSM-5 criteria) with a BMI under the third percentile at the time of hospitalization (−2.2 s.d.), whereas the Generation R study selected participants from the general population without BMI criteria (−0.44 s.d.), limiting the comparability of results. Additionally, the hypothesis-driven approach used in their study may explain the lack of replication. Similarly to our findings, their results were not significant when analyzing all brain regions (no effect on CT). Future research on ARFID should explore the long-term trajectory of patients and the distinct biological underpinnings of each clinical subtype with larger sample size. A deeper understanding of these factors will help elucidate the mechanisms driving food restriction and avoidance, the progression of the disorder and its potential medical and mental health consequences11.

Hypothesis 3 predicted a significant overlap in brain alterations between EO-AN/OCD and EO-AN/ASD, and between ARFID/ASD and ARFID/ADHD, mirroring previously published clinical observations. Results at the brain thickness level confirmed correlations between EO-AN/OCD and ARFID/ASD, as well as no association of AN/ADHD and ARFID/OCD, in line with clinical observations1,31,32. However, the lack of association of EO-AN/ASD and of ARFID/ADHD at the brain level was notably inconsistent with clinical observations, given that ~20% of patients with AN have ASD as a comorbid condition17,18 and that ~20% of ARFID patients have ADHD as a comorbid condition33,34. The correlation between EO-AN and OCD (rB = 0.59, rG = 0.45) was similar to genetic findings15,35. Overall, this multiscale overlap—at the clinical, brain and genetic levels—suggests shared mechanisms underlying psychiatric disorders that are independent of BMI. These findings highlight the importance of large-scale transdiagnostic approaches to better understand the biological foundations of mental illnesses. They also align with the growing shift toward multidimensional models of psychiatric conditions, such as the Research Domain Criteria36 and the general psychopathology factor. Moreover, they support a personalized treatment approach for AN, grounded in a bio-genetic framework37, in which the disorder emerges from multiple mechanistic pathways and traits rather than being treated solely based on symptoms.

Limitations

Several limitations should be acknowledged. First, due to the limited sample size, we could not subtype patients with ARFID based on the three clinical subgroups, which would be valuable in the future. Second, power analyses could not be conducted for ARFID, as no prior studies on brain structure in this population have been published. Therefore, it remains uncertain whether our study has sufficient power to detect effects in this group. Similarly, power analyses for EO-AN indicated that additional participants are needed in both groups to achieve sufficient power for SA. As such, the results for SA should be interpreted with caution. Third, we acknowledge that this dataset was cross-sectional rather than longitudinal. However, each patient (diagnosed with AN or ARFID) was hospitalized with a BMI percentile <3. This indicates that those with a BMI above this threshold had partially recovered and recently gained weight. This allowed us to infer about the stages of the brain recovery process. Fourth, we were limited in investigating associations between clinical profile specificities (such as IQ, avoidance of certain food groups or over-exercising) and differences observed at the brain level, due to the lack of such data collection. Fifth, the sample sizes, as well as the protocols used to acquire and compute summary statistics and brain signatures required for genetic-based correlations (rG) and brain-based correlations (rB), were not always consistent, which could have introduced some biases in the results presented in hypothesis 3. Finally, no genetic-based correlations were available for the ARFID condition.

Conclusions

This large-scale neuroimaging study reveals distinct brain alterations in two rEO-EDs, highlighting unique biological mechanisms underlying each condition. Future studies will be required to partition the contribution of low BMI versus rEO-ED mechanisms, as well as to identify shared mechanisms with other mental illnesses toward a dimensional approach of eating disorders.

Methods

Study samples

We analyzed data from 221 patients hospitalized in the eating disorder clinic of the psychiatric department of Robert Debré Hospital (Paris, France). Sample sizes reported in the section below are after MRI image quality control (see the flowchart in Supplementary Fig. 1).

Children with EO-AN

All patients met the DSM-5 criteria for AN and were inpatients at the time of their scan. The age of first inpatient admission for nutrition rehabilitation under 13 years old was used as a criterion to categorize patients as part of the early-onset subgroup of AN7. Details about the inpatient program are provided in the Supplementary Material. The EO-AN sample comprised 124 children (85% female), with a mean age at first inpatient admission of 11.40 ( ± 1.07 s.d.) years and a mean age at scan of 11.5 (±1.19 s.d.) years (Table 1). We split the patients with EO-AN into two subgroups based on the BMI percentile at the time of the MRI scan to account for the disease trajectory: the acutely ill group (EO-acAN, n = 60, percentile <3) and the partially weight-restored group (EO-pwrAN, n = 64, percentile ≥3) (Table 1 and Supplementary Fig. 2). We calculated youth BMI percentile from the Centers for Disease Control and Prevention standards using the PAutilities R package.

Children with ARFID

We enrolled 50 inpatients with an ARFID diagnosis (26%, n = 13 with a lack of interest in food; 58%, n = 29 with fear of aversive consequences of food intake; and 16%, n = 8 with sensory sensitivity impairment). All patients with ARFID included in this study were diagnosed based on the DSM-5 criteria for ARFID10, and a BMI under the third percentile at the time of hospitalization. The mean age and sex ratio was similar to the EO-AN group (mean age at scan 11.2 (±1.49 s.d.) years, 76% female). We subdivided the ARFID group according to each participant’s BMI percentile at the time of the scan. Participants with a BMI percentile <3 were considered as the ‘underweight group’, unARFID (n = 30).

Children with TD

Participants (n = 116) were recruited from the control groups of the Paris Autism Research International Sibpair (PARIS, including siblings of patients with ASD and TD children) and the DEVine cohorts at the Child and Adolescent Psychiatry Department, Robert Debré Hospital, Paris (France) (mean age at scan 10.2 (±2.046 s.d.) years, 54% female). Sensitivity analyses were performed by excluding the subgroup of siblings (n = 18) and did not reveal any differences (Supplementary Results). TD were scanned using the same imaging scanners as the AN and ARFID.

Inferential statistics were performed to compare age at scan, sex and BMI between groups (AN, ARFID and TD) using analysis of variance and chi-squared tests.

Power analyses

Power analyses based on effect sizes reported in ref. 8 (Supplementary Method) indicated that a sample size of N = 51 participants per group would be required to achieve 80% power for CT, N = 397 for SA and N = 120 for SV (Supplementary Fig. 2). These results suggest that while our sample size is sufficient for detecting effects in CT and SV, it may be underpowered for SA. Therefore, we will limit the interpretation of any findings related to SA in AN due to insufficient statistical power. No power analysis was conducted for ARFID, as no prior studies on brain structure in this population have been published so far.

Image acquisition and processing

Scanning protocols

Data were collected between 2010 and 2024 on three Philips scanners at two magnetic field strengths (1.5 and 3 Tesla). Acquisition parameters and specific sequence information are available in the Supplementary Methods. Sensitivity analyses were performed on data collected on the 3T machine only (between 2019 and 2024) (Supplementary Results).

Brain segmentation and metrics extraction: all T1-weighted structural brain MRI scans were processed using FreeSurfer (version 7.2.0 (ref. 38)) through the computing cluster at the Human Genetics and Cognitive Functions lab, Institut Pasteur (Paris, France). Quality control of the T1-weighted brain MR images and the segmentation of the cortical and subcortical regions was performed by the same rater (A.A.). Images were segmented using the Desikan–Killiany atlas39 into 68 cortical and 14 subcortical regions + 2 lateral ventricles. CT and SA measures were extracted for the 68 cortical regions (34 per hemisphere). Volumes were calculated for the subcortical regions and lateral ventricles. We also extracted the following global metrics from FreeSurfer: mean CT, total SA, ICV, GV and CSF volume.

Statistical analyses

Case–control analysis

Group difference analyses (EO-AN > TD, ARFID > TD, EO-acAN > unARFID, EO-acAN > TD and unARFID > TD) were performed to assess the impact of rEO-ED diagnosis on brain metrics. We used the ComBat batch adjustment method40, which has been extensively used in neuroimaging analyses to reduce scanner-related heterogeneity and increase statistical power for multiscanners analysis. We treated each sequence–machine pairing as an individual site. Batch-adjusted analyses were conducted using the lm function to investigate the effect of the rEO-ED diagnostic on brain metrics, using sex and age at scan as covariates. We applied the FDR procedure41 to correct for multiple testing. Analyses of the SVs were adjusted for ICV. CT and SA analyses reported in the main text were not adjusted for global metrics. CT and SA results adjusted for mean thickness and total SA, as well as percentages of differences between groups for global metrics, are reported in the Supplementary Results.

BMI mediation analyses

Mediation analyses were conducted to investigate whether BMI mediates the relationship between diagnosis (EO-AN and ARFID) and brain metrics. To account for site-related variability, ComBat harmonization was applied to correct for scanner effects. Brain metrics were normalized based on the variance observed in the control group. Mediation models were implemented using the Lavaan R package42, with age and sex included as covariates. Standardized path coefficients were extracted (Supplementary Tables 37–40).

Brain profiles correlation with TO-AN

Analyses comparing EO- and TO-AN were performed using previously published TO-AN data from the ENIGMA Eating Disorders working group8. Additional information on this sample is provided in Supplementary Material.

Concordance analyses

Brain structural patterns for three neurodevelopmental conditions

We extracted 84 Cohen’s d values (for 68 cortical and 16 subcortical regions) for each of the three following psychiatric conditions (published data from the ENIGMA consortium): ADHD (n = 1,513 cases versus 1,395 controls)43,44, ASD (n = 1,571 cases versus 1,651 controls)45 and OCD (n = 407 cases versus 324 controls)46,47. The results from each of these studies were derived from the largest, multisite cohorts ever analyzed at the time, and were analyzed using the same, publicly available ENIGMA processing pipelines.

Brain-based correlations across disorders

Recent analytic methods allow resemblance among brain maps to be tested formally, allowing similarities among different conditions to be evaluated48. We computed Pearson correlations between beta value vectors for each pair of conditions. To avoid inflation of spatial correlations, we performed spin permutation tests (5,000 null correlations per pair of brain maps) proposed by ref. 48 and implemented in the ENIGMA toolbox49.

Transdiagnostic genetic risk correlations

We extracted genetic correlations across the same three pairs of conditions computed by the Psychiatric Genomics Consortium15,35 using the most extensive GWAS per condition.

Ethics

The study was conducted in accordance with the principles of the Declaration of Helsinki and followed the Good Clinical Practice (ICH GCP) standards. Data from patients with eating disorders (AN and ARFID) were studied following the French regulation (no. MR-004) and approved by the local Ethics Committee (ref. CEER 2022-610ter). Data from the TD group were used under authorizations of the PARIS (ref. Inserm C07-33) and DEVine cohorts (protocol CEA 100 054), approved by the local Ethics Committee (CEER 2008-A00019-46 and CCP 3772-RM, respectively). All participants and their guardians provided written informed consent.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.