Association between the frontoparietal network, clinical symptoms and treatment response in individuals with untreated anorexia nervosa

Participants

Seventy-nine females with treatment-naive AN were recruited from January 2015 to March 2022 at the Eating Disorder Treatment Center, Shanghai Mental Health Center (SMHC), of whom 47 were of AN-R subtype and 32 were of AN-BP subtype. Two or more senior psychiatrists evaluated the patients according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-V) criteria. Inclusion criteria for patients with AN were as follows: (1) an initial diagnosis of AN according to the DSM-V criteria, (2) aged 13–25 years, (3) Han Chinese ethnicity, (4) right-handed and (5) a body mass index (BMI) of 13.0–18.5 kg/m². Exclusion criteria for this group were as follows: (1) a past diagnosis of or treatment for eating disorders, (2) evidence of a serious or chronic somatic disease and a history of other psychiatric disorders, (3) any psychiatric or psychological therapy for at least 12 weeks prior to the study, (4) currently pregnant or breast feeding and (5) metal implants in their body that would prevent them from undergoing magnetic resonance imaging (MRI) examination. Forty HCs were recruited among the students and health workers at the SMHC by advertisements and were matched to the patients in terms of age and educational level. The same criteria were applied to HCs, except they had no history of eating disorders and no restrictions on BMI parameters. All participants signed informed consent.

Study design

In this case-control study, all participants were administered a physical examination, a clinical questionnaire and a resting-state functional MRI (fMRI) examination at baseline. Subsequently, patients with AN received routine outpatient treatment, mainly including psychological counselling, nutritional evaluation and pharmacological therapies based on their situations. The pharmacological treatments mainly included selective serotonin reuptake inhibitors (n=32), atypical antipsychotics (n=14) and benzodiazepines (including alprazolam, lorazepam and estazolam; n=6) and zolpidem (n=4). Physical examination and clinical symptoms were evaluated at 4-week and 12-week follow-up.

Outcome assessments

Demographic and clinical characteristics, including age, educational level and BMI (calculated by current weight and height), were collected from all subjects. The severity of eating disorder was measured with the Eating Disorder Examination Questionnaire (EDE-Q 6.0). This widely used scale has 28 items that assess eating disorder symptoms, including thoughts, feelings and behaviours related to eating and body image, over the past 28 days. The scale includes 6 open-ended items—the frequency of overeating, loss of control, binge eating, self-induced vomiting, laxative abuse and compulsive exercise—and 22 attitudinal items, which are combined to create four subscales: restraint, eating, body shape and weight concerns. Purging behaviours are the sum of self-induced vomiting, laxative abuse and compulsive exercise. Higher scores on the four subscales and the total scale indicate more severe symptoms. Our previous study has confirmed good reliability and validity of the scale in mainland China, with Cronbach’s alpha coefficients of 0.95 and test-retest reliability of 0.73.16

Additional questionnaires

The Beck Depression Inventory-II (BDI-II) and the Beck Anxiety Inventory (BAI) were adopted to evaluate the depressive and anxiety symptoms of all subjects. The Chinese version of BDI-II has demonstrated excellent internal consistency, with a Cronbach’s alpha of 0.94,17 and the reliability and validity of the Chinese version of BAI were satisfactory.18

MRI acquisition and preprocessing

All fMRIs were acquired on a 3.0T Siemens Verio MRI scanner (Erlangen, Germany) with a 12-channel head coil at the SMHC. The resting-state scans were acquired using a gradient echo-planar imaging sequence (number of slices =45 170 scans, repetition time (TR) =3000 ms, echo time (TE) =30 ms, flip angle =85, field of view (FOV) =216 mm, matrix =64×64, voxel size =3×3×3 mm³, slice thickness =3.0 mm). High-resolution anatomical scans were acquired with a T1-weighted three-dimension magnetization-prepared rapid gradient echo sequence (192 sagittal slices, TR=2300 ms, TE=2.96 ms, flip angle=85, FOV=256 mm, matrix =256×256, slice thickness =1 mm). Before scanning, participants were instructed to stay still and remain awake with their eyes closed. Foam padding and earplugs were provided to control head motion and reduce scanner noise.

Image preprocessing and analysis were conducted using SPM V.12 (http://www.fil.ion.ucl.ac.uk/spm) and CONN toolbox V.20.b (http://www.nitrc.org/projects/conn) on MATLAB V.R2021b (Mathworks). The fMRI preprocessing steps were as follows: discarding the first 10 volumes, motion correction, realignment, slice-timing correction and outlier scans for scrubbing. Then, functional and structural images were segmented by grey/white/cerebrospinal fluid and normalised into the Montreal Neurological Institute space. Functional images were spatially smoothed with a Gaussian kernel of 6 mm full width at half maxima, and the bandpass filtering was 0.008–0.09 Hertz. Potential confounders such as white matter, cerebrospinal fluid, realignment, scrubbing and effect of rest were regressed out.

Region of interest selection and quality check

FPCN seeds were identified using the Harvard-Oxford cortical and structural atlas, and brain regions were defined using the Brodmann area (BA) atlas, including four brain regions of the bilateral PPC and dlPFC. The PPC seed was explored in one brain region located at BA 48, and the dlPFC seed was explored in three brain regions located at BA 8, BA 25 and BA 48. Seed-based correlation analysis was used to investigate time series connectivity between the seeds in the FPCN and all other voxels in the whole brain, and the connection strength was extracted from all participants. The quality of brain image values for each participant was checked based on the registration quality and head motion. Head motion was assessed using mean framewise displacement, with the maximal value set to 0.5 mm. Those with poor image quality were excluded from further analysis. Images from all participants passed the quality check, leading to a sample size of 79 patients and 40 HCs who passed quality checks of both clinical questionnaires and brain images.

Statistical analyses

The demographic and clinical characteristics were analysed using SPSS V.25.0. Continuous data are presented as mean (SD), and categorical data are presented as numbers or percentages. Group comparisons were conducted using independent-sample t-tests or Mann-Whitney U tests for continuous variables and the χ² test for categorical variables. Pearson’s correlation was used to assess the correlation between the EDE-Q 6.0 total score and the score on items 13–18, treatment response and the RSFC in seeds of the FPCN. Treatment response was expressed as the rate of decrease or reduction in the EDE-Q 6.0 total score and the score on EDE-Q 6.0 items 13–18 in the sensitive analysis. Multiple linear regression was adopted in the prediction model, with the EDE-Q 6.0 total score, the scores on items 13–18 at baseline and the treatment response as the dependent variables, respectively. The RSFC in the PPC and dlPFC seeds of the FPCN were the independent variables, and age, age of onset and BMI were the covariates. A mixed linear model was used to analyse the clinical response at follow-up time points. All tests were two-tailed, and p<0.05 was considered statistically significant.

A seed-to-voxel analysis was adopted using the CONN toolbox to explore the RSFC between the AN and the HC groups. Fisher-transformed correlation coefficients were generated between blood oxygen level-dependent time series in seeds within the FPCN and all other voxels in the whole brain to create functional connectivity maps. A general linear model was used to explore the differences in functional connectivity between the AN and the HC groups. The BDI-II and BAI total scores were included as covariates to exclude their influence on the results. Connectivity differences were considered significant between the two groups when a voxel height threshold was p<0.001 and a cluster size threshold was corrected at a p<0.05 false discovery rate.

We employed support vector regression (SVR) to construct the prediction model, given the modest size of our data set. SVR is chosen for its documented stability in handling small sample sizes. Five specific brain activities were used as predictive features for each type of variation. Our data set was divided into three subsets: a training set, a validation set and an independent test set comprising 20% of the total data. Data fitting is predominantly assessed through R² values and scatter plots. A higher R² value indicates better alignment between the data and the model, although an excessively higher R² value (~1) indicates overfitting. Scatter plots serve as visual aids to evaluate prediction accuracy, with closer proximity to the diagonal line of equality (where the x-axis is the observed values and the y-axis is the predicted values) indicating the predictive performance is better. Given the limited number of follow-up data points, the risk of overfitting is heightened.

Demographic and clinical characteristics

A total of 79 patients with AN and 40 HC participants took part in the study, of whom 40 patients completed the 4-week follow-up and 35 completed the 12-week follow-up. The other 39 patients did not attend the follow-up, mainly due to difficulty accessing medical care during the COVID-19 outbreak, reluctance to continue outpatient visits after improvement or additional comorbid mental disorders during the study. A flowchart of the study is shown in figure 1.

Request permission

There were no significant differences in age (t=−1.75, p=0.083) and educational level (t=−1.49, p=0.130) between the patients and HCs. The BDI-II scale (t’=5.88, p<0.001), BAI scale (t’=6.78, p<0.001) and EDE-Q 6.0 total score (t’=8.07, p<0.001) were higher for patients with AN than for HCs. As expected, patients with AN had a significantly lower BMI (t=−10.49, p<0.001) than HCs (table 1). There were no significant differences in age, age of onset, BMI, BDI-II and BAI scores, and EDE-Q 6.0 total score between patients who attended the follow-up at the 4th and 12th weeks and those who did not (see online supplemental table 1).

Resting-state functional connectivity

According to seed-to-voxel analyses, RSFC maps in the FPCN seeds showed significant differences between the AN and the HC groups. A significantly increased RSFC (p<0.05, false discovery rate-corrected) was observed in patients with AN compared with HCs between the right PPC seed and the left inferior frontal gyrus (IFG) (x=−42, y=6, z=20, t=5.48, k=117, size p=0.029), the right dlPFC seed and the middle temporal gyrus (MTG) (x=−48, y=−20, z=−6, t=5.16, k=111, size p=0.032), the posterior division left (x=−6, y=38, z=48, t=5.56, k=102, size p=0.032) and the subcallosal cortex (x=−2, y=18, z=−6, t=5.62, k=96, size p=0.032). For the significant clusters in conn, we extracted their specific values and used SPSS to perform an independent samples t-test statistic. The result shows significantly increased RSFC in the FPCN in patients with AN than HCs, and their statistical values are (t=4.04, p<0.001), (t=5.17, p<0.001), (t=4.75, p<0.001), (t=4.82, p<0.001) respectively in the same order as above. was observed in patients with AN compared with HCs (t(degrees of freedom)=, p=, chon'd=). (figure 2 and online supplemental table 2). No significant differences were found in the RSFC maps generated from the seeds of the left PPC and the left dlPFC. Significant results for the dlPFC seed are presented in figure 2. Furthermore, no significant differences in the RSFC maps of the FPCN seeds were found between the AN-R and AN-BP subtypes.

Request permission

Correlations between the RSFC in different brain regions and clinical variables

Pearson’s correlation analysis was conducted to explore the associations between the Z scores of clusters that showed significant differences in the RSFC of the FPCN seeds and relevant self-reported eating disorder symptoms at baseline in the AN and HC groups. Significant correlations were found between the PPC to IFG connectivity strength and the EDE-Q 6.0 total score (r=−0.335, p=0.003), frequency of overeating behaviours (r=−0.238, p=0.036) and frequency of loss of control (r=−0.270, p=0.017) at baseline in patients with AN. No significant correlations were found between the RSFC in the FPCN and self-reported eating disorder symptoms in the HCs (figure 3A–C).

Request permission

Treatment response was expressed as the decrease in the EDE-Q 6.0 total score between baseline and the corresponding treatment node and the decrease rate of the scores on items 13–18 of the EDE-Q 6.0. A mixed linear model was used to analyse the treatment response at the follow-up time points. It showed that the time main effect of the EDE-Q 6.0 total score was not significant over time, and the loss of control and overeating in the model failed to fit due to missing data. The results are shown in online supplemental materials (tables 3–7 and figure 1).

Correlation analysis was performed between treatment response and five different brain regions. It showed that the RSFC of the PPC–IFG had an approximately significant or significant correlation with treatment response both in the 4th week (r=−0.278, p=0.083) and 12th week (r=−0.416, p=0.013) (figure 3D,H), and the RSFC of the right dlPFC–MTG had a significant correlation with improvement in the frequency of loss of control (r=0.451, p=0.031) and binge eating (r=0.556, p=0.013) (figure 3E,F). Also, the RSFC of the right dlPFC–left superior frontal gyrus (SFG) had a significant correlation with improvement in the frequency of overeating behaviours (r=0.502, p=0.029) (figure 3G).

Furthermore, multiple linear regression was conducted with the EDE-Q 6.0 total score (F=5.16, p=0.008, R²=0.12), frequency of overeating (F=2.21, p=0.076, R²=0.11) and loss of control (F=2.91, p=0.027, R²=0.14) at baseline as dependent variables, respectively, with the RSFC in seeds of the FPCN as independent variables, and age, age of onset and BMI as covariates. Multiple linear regression was conducted with the decrease in the EDE-Q 6.0 total score in the 4th week (F=3.74, p=0.061, R²=0.38) and 12th week (F=3.098, p=0.039, R²=0.38), and with the improvement in loss of control (F=4.08, p=0.032, R²=0.29) and binge eating (F=3.85, p=0.050, R²=0.66) as the dependent variables, respectively. The RSFC in seeds of the FPCN were the independent variables, and age, age of onset and BMI were the covariates. As shown in table 2, the RSFC in the PPC–IFG was a significant factor for both the clinical symptoms at baseline and the treatment response in the 4th and 12th weeks for patients with AN.

We also explored the machine learning prediction model using SVR. Five brain activities were used as predictive features for each type of variation. These results are shown in the support vector machine regression parts of online supplemental tables 8–39 and figures 2–17.

Main findings

To our knowledge, the present study is the first to investigate the role of significant seeds in the FPCN in self-reported eating disorder symptoms and treatment response in patients with AN. First, after controlling for depression and anxiety symptoms, the RSFC from the PPC and dlPFC of the FPCN increased in patients with AN versus HCs. Second, the RSFC of the PPC to IFG was a significant neural marker of self-reported eating disorder symptoms after controlling for age, and it was a significant neural marker of treatment response to cognitive preoccupations about eating/body image after controlling for age, age of onset and BMI. Likewise, the RSFC of the dlPFC to MTG/SFG may be a significant neural marker of the treatment response to binge eating and loss of control/overeating behaviours in patients with AN. Because no significant differences in the RSFC of the FPCN were found between patients with AN-R or AN-BP, the RSFC of the FPCN may not be a neural marker to differentiate the two subtypes of AN. Thus, these results provide evidence for the important role of the FPCN in patients with AN.

Moreover, the RSFC in the PPC to IFG and the dlPFC to MTG/SMG increased in AN versus HCs in the study, further supporting that AN is a disorder of excessive cognitive control, a finding that agrees with previous studies that found the function in the FPCN increased in patients with AN.7 19 The FPCN engages in various executive functions by allocating top-down attentional resources to arrange cognitive control processes.20 The impaired cognitive control ability has been associated with decreased cognitive flexibility and hyperdetailed information processing in patients with AN.10 One study confirmed that the FPCN functional connectivity contributed to the loss of control in patients with binge drinking, possibly by impairing cognitive function and response inhibition.21 In addition, the right IFG is involved in inhibiting control and stopping the upcoming impulsive responses, and the impairment of the right IFG is closely related to impaired inhibitory control.22 One study pointed out that age was an important moderator of overall cognitive performance in AN, including executive function, with younger participants had better performance than the older participants.19 Considering the increased function in the FPCN at a young onset age, we put forward a viewpoint that the neural circuitry of the FPCN in patients with AN is premature.

Furthermore, our study showed that the aberrant RSFC in the FPCN was related to the abnormal eating disorder symptoms in patients with AN but not in the HCs, supporting the hypothesis that the RSFC in the FPCN may contribute to the regulation of pathological symptoms in patients with AN. The alteration of the FPCN neural function may be an important factor in causing pathological symptoms and psychological characteristics in patients with eating disorders, which is consistent with our findings.23 The FPCN system is responsible for various cognitive functions and regulates eating behaviours. A lower function connectivity in the FPCN was associated with worse self-control in eating, further contributing to disturbed eating disorders.24 Additionally, the PPC and dlPFC of the FPCN are involved in regulating eating behaviours through cognitive control when confronted with tempting food.25 It has been suggested that the food consumption of individuals with AN is closely related to the dlPFC and dorsal striatal connectivity.23 The above studies have provided the theoretical basis for our results.

The current study also showed that the aberrant RSFC in the PPC to IFG was negatively associated with self-reported eating disorder symptoms, loss of control and overeating, and that the treatment response was negatively associated with cognitive preoccupations about eating/body image. Interestingly, the aberrant RSFC in the dlPFC to MTG/SMG was positively associated with the treatment response to binge eating and loss of control/overeating behaviours of AN. On the other hand, we found a stronger RSFC in the PPC predicted poorer clinical outcomes regarding cognitive preoccupations about eating and body concerns. It may be that a stronger RSFC in the PPC implies more attention to eating, body shape and body weight, which leads to potential excessive attention to eating, body shape and body weight, as well as excessive control over diet. However, the increased RSFC in the dlPFC predicted an improvement in loss of control and binge eating or overeating behaviours, analogous to previous results that showed increased dlPFC activity of the cognitive control circuit predicted dietary improvements.26

A recent study indicated that the FPCN, as a crucial cognitive control network, and the functions of its different parts were associated with the selection and maintenance of various stimuli and features. Its functional connectivity can also combine external information with internal representations to guide decision-making.27 It has been argued that treatment response in patients with depressive disorder was associated with altered connectivity within and between networks, including the FPCN.14 In addition, increased functional connectivity between the frontal cortex and the FPCN may be one of the neurological mechanisms involved in successfully resolving response conflicts.27 Patients with AN may develop an adaptive neural mechanism to maintain their extreme eating behaviours due to prolonged malnutrition. The above theory may provide the rationale for the seemingly opposing results derived from the study. Further validation of the effect of the PPC and the dlPFC on patients with AN and the clinical outcomes is still needed. It is currently difficult to predict treatment response in patients with AN; however, the resting-state brain activity of the FPCN may serve as a biomarker to predict treatment response in AN. As noted in this study, the dlPFC is an important target for understanding the pathogenesis of AN,28 and the symptoms of binge eating and loss of control/overeating behaviours. The study also found that the PPC may be a significant target for cognitive preoccupations about eating/body image in patients with AN, suggesting that it may serve as an important target for interventions addressing cognitive preoccupations that could be considered in future treatment protocols. This study did not track the FPCN function after treatment to determine long-term effectiveness, which needs further investigation.

No significant differences in the RSFC of the FPCN between the AN-R and AN-BP subtypes were found in this study. A long-term follow-up study on patients with different subtypes of AN showed that most patients with AN-R have an early age of onset, and most patients with AN started with the AN-R subtype. Approximately 88% of those with AN-R develop bulimic behaviours, and 62% of patients with AN-R eventually develop into the AN-BP subtype.29 Based on our results, we conjecture that the RSFC of the FPCN is altered in the early onset of AN. Moreover, most patients are adolescents and not yet adults, a stage when the neural activity of the brain is still at a premature stage of development, and the brain is more plastic in structure and function compared to adults,30 which may explain the short-term improvement of clinical symptoms.

Limitations

There are several limitations to the present study. First, the study was based on a cross-sectional analysis, which could not determine the causation, and it was not clear how the RSFC of the FPCN changed after treatment. Second, the study did not consider the effect of undernutrition on brain networks. Third, more participants than expected were lost to follow-up due to COVID-19 outbreak; thus, their BMI was not tracked. Fourth, treatment response in AN was expressed as a reduction rather than a reduction rate in the EDE-Q 6.0 total score, which may increase the power to predict the outcome. As the study was exploratory, we did not perform multiple tests for corrections to increase the likelihood of obtaining significant results. Since this study is the first to explore the neurobiological mechanisms of treatment response in AN, future research with larger samples and a longitudinal design is needed to validate our findings.

Implications

The current study provides evidence that the RSFC in the FPCN is increased in patients with AN versus HCs after controlling for emotional symptoms, and it yields an initial indication that the RSFC in the PPC to IFG may be a significant neural marker of self-reported eating disorder symptoms and treatment response to cognitive preoccupations about eating/body image. Also, the RSFC of the dlPFC to MTG/SMG may be a significant neural marker for treatment response to binge eating behaviours and loss of control/overeating behaviours of AN. These findings are helpful in further understanding the pathogenesis of AN, as well as opening up potential avenues for relevant target strategies for the prevention and treatment of this pervasive mental disorder.