The effect of body image preoccupation on decision making in eating disorders

Eating disorders

To further describe ED, it is important to mention the three main types: Anorexia Nervosa (AN), Bulimia Nervosa (BN) and Binge-Eating Disorder (BED) as in 5^th edition of The Diagnostic and Statistical Manual of Mental Disorders (American Psychiatric Association, 2013). All share low life satisfaction, greatly impaired quality of life, and increased mortality or suicide risk.

Anorexia Nervosa

AN is characterised by extreme restriction of food intake and the pursuit of weight loss. This is usually associated with an intense fear of gaining weight as well as with body image preoccupation. Individuals with AN tend to perform obsessive body-checks in the mirror and measuring of body parts. Even though it may be very rewarding for people with AN to notice weight-loss, the associated body image disturbance prevents them from recognising how malnourished they are, which furthers the cycle of restriction. Such behaviour usually, but not necessarily, leads to extremely low body weight. AN affects mainly females, but as males also suffer from the disorder.

Bulimia Nervosa

BN differs from AN in that individuals with bulimia undergo recurrent episodes of binge eating followed by compensatory behaviours. Namely, they consume an unusually large amount of food in a relatively short amount of time, often to the point of pain, without much of a sense of control over their eating. To prevent a perceived weight gain that would follow a binge, they engage in behaviours such as self-induced vomiting, use of laxatives and diuretics, fasting, overexercising, and ingestion of weight-loss products or medications. These are usually associated with excessive self-evaluation of their body as a core determinant of self-worth.

Binge-Eating Disorder

BED is very similar to BN in the presence of recurrent episodes of uncontrollable binge-eating, accompanied by the feeling of uncomfortable fullness, disgust and shame with oneself and eating problems. However, BED is not associated with compensatory behaviours as in BN. Moreover, the gender ratio of BED is less biased towards females.

Computational psychiatry

One possible avenue to understand and help devise treatments for eating disorders, alongside the traditional approach of psychotherapy, psychiatry and neuroimaging (Kessler, Hutson, Herman, & Potenza, 2016; Linardon, Kothe, & Fuller-Tyszkiewicz, 2019; Simon et al., 2016) is the emerging field of computational psychiatry (CP). The field allows creating a framework that is able to provide a digital biomarker of different mental disorders. To achieve that, researchers focus on various modes of decision making that are tested with a mode-specific decision-making task and quantified with a computational model that captures individual and group differences (Q. J. M. Huys, Moutoussis, & Williams, 2011; Seriès, 2020). Some mental disorders such as schizophrenia, anxiety, and obsessive-compulsive disorder (OCD) have been linked to various such digital biomarkers of aberrations in a decision-making process (Raymond, Steele, & Seriès, 2017; Rocha, Alvarenga, Malloy-Diniz, & Corrêa, 2011; Valton, Romaniuk, Douglas Steele, Lawrie, & Seriès, 2017). Moreover, employing computational methods to understand mental disorders allows bridging the neurobiological and psychiatric levels of descriptions of an ill mind. Such link can lead to better and more advanced theories of mental disorders, as well as inspires new treatment approaches, and promises improved early prevention (Q. J. Huys, Pizzagalli, Bogdan, & Dayan, 2013).

Computational psychiatry of eating disorders

As mentioned above, depression, OCD, anxiety and schizophrenia have received a lot attention in the field of CP, with very promising results and theories. Eating disorders have received less attention: we could only find eleven papers in the last nine years about CP of eating disorders, either as a main subject or in a study of multiple disorders (Bernardoni et al., 2018; Chan et al., 2014; DeGuzman, Shott, Yang, Riederer, & Frank, 2017; Foerde et al., 2019; Frank, Reynolds, Shott, & O’Reilly, 2011; Giannunzio et al., 2018; Gillan, Kosinski, Whelan, Phelps, & Daw, 2016; Patzelt, Kool, Millner, & Gershman, 2019; Reiter, Heinze, Schlagenhauf, & Deserno, 2017; Verharen et al., 2019; Voon et al., 2015). For comparison, a number of papers on CP of schizophrenia by members of only one single group reaches 27 in the last eight years (TNU - publications, n.d.).

The main focus has been on two main aspects of decision making in eating disorders. One is related to how sensitive to punishments individuals with AN are, that is how well they are able to avoid negative experiences during a decision-making task, and how well they learn from negative feedback from the environment. Unfortunately, the results from five studies prove to be quite contradictory. For example, Bernardoni et al. (2018) show that sensitivity to punishments is increased in AN and they correlate this finding with neural features registered by fMRI using a slightly more nuanced version of the task used by Verharen et al. (2019), whose study suggests decreased sensitivity to punishments. While both studies provide fairly convincing support for their results, it is difficult to say whether AN are ultimately characterised by increased or decreased sensitivity to punishments.

The second main aspect of ED that has been of interest to computational psychiatrists is the trade-off between goal-directed and habitual system utilisation in decision-making. As mentioned in the introduction, goal-directed decision-making is related to acting in the environment with a goal in mind, intentionally and deliberately. It is usually characterised by forming an internal model of the environment that describes which states and actions will bring about the best results over prolonged time. On the other hand, habitual decision-making is associated with responding to stimuli in the environment in an automatic manner, without much deliberation, usually repeating those actions that immediately yield the best results. In this case, an agent does not create a model of the environment. As such, goal-directed behaviour utilises model-based learning, which allows building an accurate model of the states and actions and their associated values that takes into account a hidden probabilistic structure of the environment. However, this kind of learning is computationally more demanding, using more resources to support the process. The habitual behaviour employs model-free learning, which updates a running score of possible states and actions, based on the last experience, without registering any hidden structures in the environment. Such a process is computationally efficient as it relies mainly on the memory of the last events (Dolan & Dayan, 2013).

It has been shown that healthy population employs both model-free and model-based learning that trade off against each other (Daw, Gershman, Seymour, Dayan, & Dolan, 2011). Several studies attempted to see how this trade-off is different within eating disorders. The results are converging and suggest that sub-clinical ED group shows reduced model-based learning (Gillan et al., 2016) as well as clinical AN and BED groups (Foerde et al., 2019; Voon et al., 2015). Moreover, a recent study by Foerde et al.(2019) suggests an even more reduced utilisation of model-based learning in AN when the task is strictly food related as compared to a monetary task. This is of particular significance for two reasons. Firstly, AN is characterised by extreme pursuit of weight-loss, which can be understood as extreme goal-directed behaviour. As such, one would expect to see increased model-based learning in AN, yet the results suggest otherwise – such “goal-pursuit” of weight-loss in AN is actually habitual, almost compulsive. Secondly, the study investigates a food-related task (as opposed to monetary) that is relevant to eating disorders, and shows that model-based is further decreased compared to a monetary task.

Problem Statement

Evidently, there is little computational psychiatry research about eating disorders, which is surprising given the high mortality. As such, it is necessary to allocate more resources in that area, especially into the body image disturbance aspects, which being one of the core elements of the disorder has not been investigated.

Furthermore, some studies using model fitting methods during the investigation of the trade-off between model-based and model-free learning do not perform or ensure a good reliability analysis of the model and its fitting procedure (such as model comparison, parameter and model recovery) (Patzelt et al., 2019; Voon et al., 2015). As a result, the lack of reliability may potentially skew the interpretation and conclusions drawn.

Objectives and Hypothesis

To answer the above problems, we implement a decision-making task that captures both model-based and model-free contribution. The task is given to a sub-clinical population who score high on an eating disorder, and body image disturbance questionnaires. For comparison, a healthy control group (HC) is also be recruited.

Our aim is to extend the analysis of the trade-off between model-based and model-free learning (goal-directed and habitual behaviour) and explore how body image preoccupation affects the utilisation of model-based learning during decision-making. To do so, we analyse two task conditions, one of which targets the preoccupation with body image and its manifestation in decision making. The introduction of body image disturbance condition aims to create a context in which ED finds themselves during everyday struggle - excessive worry about body image and how to change it through dieting.

We hypothesise that: 1) In the monetary (neutral) condition and as described by previous studies, the group with ED will report a significantly decreased model-based learning and increased model-free learning as compared to HC. 2) Model-based learning will be further decreased in the body-image condition in the group with ED. HC will not report any significant differences between conditions.

Participants

An online study was performed. Participants for the study were recruited using Prolific, which is an online participant recruitment service used mainly for research and academic purposes (Prolific, n.d.). The platform provides a pool of participants that are reliable and allows to custom screen for different groups before recruitment.

Since for this study two groups were needed – a healthy control and an eating disorder group - we applied two separate sets of pre-screening criteria. For HC we looked for people who:

1. “Have never gone on a diet in the past.”

2. “Do no currently for at least one week restrict food intake to manage weight.”

3. “Have no diagnosed mental health condition that is uncontrolled (by medication or intervention) and which has a significant impact on your daily life/activities.”

4. “Do not have or have not had a diagnosed, on-going mental health illness or condition.”

For ED, participants had to select that they have gone on a diet, as in criterion (a), and that they restrict food intake to either lose or maintain weight, as in criterion (b). Moreover, to ensure data of good quality both groups had additional criteria to meet:

1. age between 18-38

2. have normal or corrected-to-normal vision

3. female as an assigned sex at birth

4. an approval rate of 98%

5. a minimum of 20 previous submissions on Prolific.

Participants first completed a range of questionnaires to be then further selected to complete a decision-making task. The questionnaires were hosted on Qualtrics (an online survey software) (Qualtrics, n.d.), while the decision-making task was hosted partly on Qualtrics and partly on Pavlovia (an online behavioural experiment platform) (Pavlovia, n.d.), designed using PsychoPy software (PsychoPy, n.d.). For each part, participants were paid at a rate of £6.25 per hour. The questionnaire took on average 7 minutes to complete, while the task took on average 27 minutes. Subjects were based all around the world. The study was approved according to the University of Edinburgh’s Informatics Research Ethics Process, with an RT number 2019/48215.

After further selection based on questionnaires and application of exclusion criteria for the decision-making task (see Supplementary material), 38 (ED group) and 34 (HC group) participants were selected for data analysis.

A prior power analysis

A prior power analysis was performed to determine the sample size needed (of each group) that achieves 80% power at the 5% level (for the sought significant difference) and at the 50% level (for the sought lack of significant difference between conditions in HC). We used G*Power software (Faul, Erdfelder, Lang, & Buchner, 2007) in calculating the sample size of t-tests for: 1) the difference in mean model-based contribution between two conditions in both groups (two dependent means; matched pairs); 2) the difference in model-based contribution between two groups in the neutral condition (two independent means; two groups). Using means and twice the standard deviations¹ from a similar study (Foerde et al., 2019), the power analysis revealed that the first set of t-tests requires at least n = 6 ED and n = 9 HC participants, while the second requires n = 38 in each group.

Self-report questionnaires

In the first part of the study, participants answered questions about their age, gender, weight, and height. They also completed three questionnaires - Eating Attitudes Test (EAT-26) (Garner, Olmsted, Bohr, & Garfinkel, 1982), Appearance Anxiety Inventory (AAI) (Veale et al., 2014) and The Obsessive–Compulsive Inventory (OCI-R) (Foa et al., 2002) – to assess the spectrum of: an eating disorder, body image disturbance and obsessive-compulsive behaviours, respectively.

In order to create HC and ED groups, cut-off points were applied for the questionnaires as in the Table 1, chosen based on the literature. For EAT-26, it is suggested to use a cut-off of 11 for subclinical populations (Orbitello et al., 2006) that are characterised as displaying disordered eating behaviours that could warrant further clinical diagnosis. Hence, for HC we set a threshold of ≤ 10 and for ED ≥ 14 to achieve stronger effects. For AAI, a few cut-off points have been suggested: a score above 6 may suggest some issues with body image (Bjureberg et al., 2019), while scores between 15-20 is reported in an appearance-concerned adult population, and a score above 20 indicates a high-risk group for body image disturbance (Mastro, Zimmer-Gembeck, Webb, Farrell, & Waters, 2016). We therefore chose as cut-off for ED a score ≥ 14, whereas for HC a score of ≤ 10 was chosen. Lastly, an additional criterion for HC group was applied so that they do not display worrying obsessive-compulsive behaviours as these have also been associated with reduced model-based learning (Gillan et al., 2016). We selected a cut-off point of ≤ 10 on OCI-R questionnaire for HC, which is twice as low as the recommended optimal cut-off point (Foa et al., 2002). Moreover, two attention checks were implemented in the questionnaires to filter out participants not taking part in the study in good faith.

Two-step decision making task

Selected participants were asked to complete a two-step decision-making task in order to quantify model-based and model-free learning contribution (Daw et al., 2011).

Before proceeding with the task, participants were asked to select a body type that is most similar to their own from a diverse range of body types in the Supplementary Figure S1. Furthermore, they were instructed on how to complete the task by reading through an illustrated tutorial. After reading the instructions, they completed 25 trials of the task as a practice.

The task reproduces that of Daw et al. (2011) but was conceived as a treasure game to make it as engaging and easy to understand as possible (Figure 1). In the first stage (grey stage), participants are presented with a choice between two stimuli - on the left, a ship against a sunset, while on the right, a pirate ship near a lighthouse. They are asked to play a role of a treasure hunter and decide which ship to board. Each ship can sail to two destinations - a blue island and a purple island. The game has a hidden structure, which participants have to learn, such that the normal ship sails to the blue island 70% of the time (common transition) and to the purple island 30% of the time (rare transition), while the pirate ship sails to the blue island 30% of the time and to the purple island 70% of the time.

• Download figure
• Open in new tab

Figure 1.

The structure of the two-step decision-making task. In the first stage (grey), a dashed arrow corresponds to a rare transition with the probability of 30%, while the bold arrow corresponds to a common transition with the probability of 70%. In the second stage, the speckled arrow is associated with the probability of receiving a reward. A duration of the stage is noted on the right side of it (stage one, two or the reward).

Once the participant boards the chosen ship, they are taken to one of the islands (stage two), where two chests await them. Here, they are asked to quickly select the chest to try their chances at finding a treasure (a pirate coin). However, each chest is assigned a probability of containing the coin, which evolves over time according to a Gaussian random walk with a standard deviation σ = 0.0275, as in the Figure 2 (more details in the Supplementary material). The job of the participant is to track, over time, which chest is the most favourable, i.e. has the highest chance of yielding the coin. After the chest is opened and the participant receives the coin (or not), they are taken back to the first stage to repeat the trial.

• Download figure
• Open in new tab

Figure 2.

The sample evolution of the reward probability for each of the chest in stage 2 over 150 trials, in the order as in Figure 1. The probabilities evolve according to a Gaussian random walk with σ = 0.0275. One set of two evolutions always starts randomly in a range [0.58, 0.72], while the other set of two starts in [0.31, 0.45]. The values of probabilities are bounded in [0.25, 0.75].

There are two conditions, 150 trials each. Each condition consists of two series of 75 trials, with a short break in between to counteract a possible loss of attention or tiredness. The neutral condition is exactly like in the Figure 1, with a pirate coin as a reward (Figure 3A). The other condition, the body image disturbance (BID) condition, differs from the neutral in that the reward is a pirate coin next to a body type that the participant selected as the most similar to their own (Figure 3C). The aim of the BID condition is to gauge the effects of body image preoccupation on decision-making.

• Download figure
• Open in new tab

Figure 3.

Possible reward outcomes during the task. In the neutral condition, participants could receive either A or B, while in the BID condition they could receive either C or B. An empty box B indicates no reward. The body type was selected by the participant before the task, more body type examples can be found in the Supplementary Figure S1.

For a balanced design half of the participants in each group completed the neutral condition first, followed by the BID condition. The other half completed the task in a reverse order. All results are then based on the average across two subgroups.

Model-based and model-free learning in the task

The reason to use the task to capture the contribution of model-based and model-free learning lies in the stage-like structure and probabilistic nature of the task, thanks to which we can easily distinguish between two strategies a participant can have, usually is a mix of the two.

For the purpose of this example, we focus on two extreme cases (Gillan et al., 2016). On one hand, an agent could utilise only model-free learning strategy. This means that they would make their decision as to which ship and chest to select purely based on the reward they received (or lack thereof), usually repeating the rewarded action, regardless of the fact the transition type. On the other hand, a participant who uses only model-based strategy, takes into account the learnt transition structure of the task, on top of the knowledge about previous rewards. Moreover, they track the probability of receiving the reward to know which chest is the most favourable.

Model-based and model free learning - an RL model

The model has two stages, which are described and quantified below (Foerde et al., 2019; Gillan et al., 2016). For simplicity, we begin with the update equation of the state-action value function, Q^II (s₂, c₂) for the second stage states and actions. There are two possible stage two states s₂: 1 - the blue island; 2 - the purple island, such that s₂ ∈ {1,2}. On each island, an agent can make two separate choices c₂: L - open the left chest; R - open the right chest, such that c₂ ∈ {L, R}. Moreover, after opening the chest, the agent can receive the reward, r ∈ {0,1}, where 0 corresponds to an empty chest, and 1 corresponds to the pirate coin. At the start of the task all value functions are initialised at zero. On any trial, t, we update the value function Q^II(s₂ c₂) of the visited state and action taken as in the Eq. 1. The subscript, t, indicates the current trial values, whereas t + 1 indicates the values at the following trial. α is the learning rate. To calculate the probability of making a choice c ∈ {L, R} at stage two, we use the softmax function as in Eq. 2, with an inverse temperature parameter β₂ quantifying the influence of the value function on making the choice. In stage one, we directly see how model-free and model-based learning play their part. Here, we have two sets of update equations. The first set, Eq. 4-5 is model-based, where the agent uses their knowledge about the environment - which stage two state-action is the best as well as the structure of the task (common vs. rare transition) to update the values of the ships in stage one. In this case, the value functions for both ships are updated simultaneously, such that the value of choosing the normal ship is the maximum over stage two actions on the blue island, while the value of the pirate ship is the maximum over stage two actions on the purple island (common transitions). The second, model-free update in Eq. 6, updates the values of the ships based only on the reward received at the end of the trial. Model-free and model-based contributions are joined together in a weighted value function over two ship choices, Q^I, as in Eq. 7. The contribution of each system is captured with β_MB (model-based) and β_MF (model-free) parameters. An additional indicator I(c = c_1,t−1) tells if the choice made on the current trial is repeated as in the previous one, with a parameter ρ describing how much switching or staying is done regardless of the feedback. Finally, the probability of choosing either ship is calculated as in Eq. 8, analogously to the second stage. Unvisited states and unchosen actions were decayed by (1 − α). The model has a total of five parameters: β_MB, β_MF, β₂, α, ρ.

Model reliability analysis

A model reliability analysis was done before (pre-hoc) and after (post-hoc) data collection. It consists of parameter and model recovery, and model comparison with alternative models. The analysis revealed that the model described above is satisfactorily reliable and superior to other alterative models considered, with average parameter recovery Pearson correlation coefficient (PCC = 0.83). More details can be found in the Supplementary material.

Model-based learning difference - predictor

An additional measure was introduced, Δβ_MB = β_MB,_neutral − β_MB,BID, to capture the difference in model-based learning across conditions (neutral − BID). The measure quantifies how much influence body image preoccupation has on model-based learning. Similar measures were calculated for other parameters.

Δβ_MB was used as a predictor of questionnaire scores and of a group (HC or ED) in a linear (m1 − m3) and logistic regression (m4). Models were implemented in R:

Model reliability

Pre- and post-hoc model reliability analysis was performed for 35 subjects, based on 150 trials of the task (unless otherwise specified). The procedures revealed that model 1, as described in the Methods, provides the best and most reliable fit to the data. It had the lowest BIC score, the best parameter recovery and it was recovered well. Therefore, model 1 was used to analyse the collected data. Detailed results from the model reliability analysis can be found in the Supplementary material.

Post-hoc parameter recovery of the best model

We repeated post-hoc parameter recovery of the best model for sufficient number of EM iterations, this time for 50 runs, to illustrate and provide a more detailed analysis. The average Pearson correlation coefficients (PCCs) from 50 runs of the procedure are presented in Table 2.

The PCCs from a sample parameter recovery run for the best model (model 1) are plotted in Figure 4.

• Download figure
• Open in new tab

Figure 4.

Recovered vs. simulated model parameters from a sample run of parameter recovery procedure for the best model. PCC was marked as a diagonal line with a PCC slope fitted to the recovered parameters. Simulated parameters were drawn from a Normal distribution with mean±SD of the fitted parameter values of ED group in a neutral condition as the mean and covariance matrix of the distribution.

The parameter of most interest, β_MB, has an average PCC of 0.68, which is sufficient to use in the analysis.

Demographic and questionnaires summary

The screening procedures resulted in recruiting two significantly different groups - healthy control (HC) and eating disorder (ED). The summarised information, along with two-sample t-tests, can be found in Table 3. The groups are mainly characterised by average total scores on the three questionnaires. HC scored on average 3.21, 4.76, 5.62, while ED scored: 30.71, 26.97, 26.05 on EAT-26, AAI and OCI-R questionnaires, respectively. This is significantly different across all questionnaires (p<0.001). Since the groups also significantly differ in age and Body Mass Index (BMI), these variables were included as covariates in the regression analysis to see if they explain the variance.

Task performance - rewards and reaction times

Firstly, the performance in the task was analysed independent of the reinforcement learning model. Average characteristics were calculated for each group (detailed results in Table S2).These model-independent measures include: total reward in the neutral and BID condition, and the total reward after completing the full task, as well as mean reaction times (RT) during the neutral, BID, and across both conditions.

The only significant difference (p=0.0463) is in the total reward received during the neutral condition, where, on average, ED received significantly less total reward (77.66) than HC (82.15). All the other measures do not differ between groups.

RL parameter estimates - model-based and model-free learning

RL model parameters were fit as described in the Methods and the Supplementary material. For each subject, a relative model-based fractional reliance was also calculated, treated as a new parameter (Foerde et al., 2019): A comparison of average parameter estimates with standard errors (SE) of β_MB, β_MF,, α, w between groups and conditions can be found in Figure 5.

• Download figure
• Open in new tab

Figure 5.

Mean ± standard error (SE) of the estimated model parameters in HC (blue) and ED (red) during each condition (neutral and BID). A: β_MB, B: β_MF, C: α, D: w.

The initial results suggest that ED is less model-based (β_MB, w) than HC in both conditions. ED is also less model-free β_MF than HC. There is no visible difference in the learning rate, α, between groups in the neutral condition, while ED has a smaller α in the BID condition. A detailed statistical analysis of the results is performed below.

Results from two-sample t-tests

To look for significant differences in parameter estimates between groups and conditions two-sample t-tests were performed. The results of the statistical tests can be found for the: comparison of the parameters between neutral and BID conditions in HC (Table 5); comparison of the parameters between neutral and BID conditions in ED (Table 6); comparison of parameters between groups in a neutral and BID condition (main left and right column in Table 7, respectively). There were no significant results for the ρ parameter (supplementary Tables S12-S14).

For HC there was no significant difference in parameter estimates between the neutral and BID condition. However, there was a strong significant difference in the model-based learning parameter, β_MB, between the two conditions in the eating disorder group. In the neutral condition, ED had an average β_MB = 0.21, which dropped significantly (p<0.001) to β_MB = 0.06 in the BID condition.

Furthermore, comparing the estimated parameters in the neutral condition between groups showed a significant difference in β_MB = 0.43 (HC), β_MB = 0.21 (ED) (p=0.0096) and in β_MF = 1.4 (HC), β_MF = 0.84 (p=0.0049). The difference in the BID condition was also significant between the two groups for β_MB = 0.53 (HC), β_MB = 0.06 (ED) (p=0.0003), but also for β_MF = 1.39 (HC), β_MF = 0.99 (ED) (p=0.0036). There were no differences between groups and conditions for the learning rate, α and the relative model-based fractional reliance, w.

These results suggest that ED group is less model-based in the BID condition vs. the neutral condition, but also less model-based than HC across both conditions. Moreover, the ED group appears less model-free in the neutral and BID condition than HC. The unexpected decreased model-free learning result, along with decreased model-based learning in ED may suggest an overall decreased learning capability of ED during the task. Alternatively, model-free parametrisation in model 1 using just β_MF might incorrectly capture model-free contribution, while model 2 parametrisation might be more accurate in splitting model-free learning parameter: β_MF=β_MF0 + β_MF1 (Supplementary material). An additional and more comprehensive study would have to be designed to verify these suggestions.

Results from regression analysis

To further test the effect of the group and condition on the model-based parameter β_MB, a random effects linear regression was performed with group and condition as fixed effects per subject. The regression results can be found in Table 8.

There is a significant effect of ED (p=0.0329) and its interaction with a BID condition (p=0.0226) on the model-based parameter β_MB. The regression coefficient estimates correspond to β_MB values in each group-condition, as in Tables 5-6.

To see if age and BMI can have an effect on model-based learning, a more complex regression model with age and BMI as extra covariates was fitted as in Table 9.

However, in this case, the more complex model does not show any significant effects of age and BMI on β_MB (only the interaction of being in the ED group and the BID condition). The Akaike information criterion (AIC) was computed for that model for comparison with a simpler model. The simpler model provides a better fit since it has a lower AIC=176.18 as compared to the model with age and BMI (AIC=187.38).

Difference between parameters across conditions

To further examine the effect of body image preoccupation on decision-making, new variables, Δβ_MB, Δβ_MF, Δα, Δw, were calculated to capture how much model parameters changed from the neutral condition in the BID condition. The mean values with standard errors of the parameter differences for each group are shown in Figure 6.

• Download figure
• Open in new tab

Figure 6.

Mean ± standard error (SE) of the estimated model parameter differences between conditions in HC (blue) and ED (red). A: Δβ_MB, B: Δβ_MF, C: Δα, D: Δw.

Preliminary inspection suggests that the parameter differences analysis shows a positive difference of model-based contribution in ED as compared to HC (negative difference), which means that model-based learning decreased in ED and increased in HC between conditions. There is also a negative difference in model-free learning for ED, with no change for HC, while the learning rate remains the same in ED and improves in HC. Fractional reliance, w, reflects the behaviour of Δβ_MB.

Results from two-sample t-tests Δβ_MB

Further statistical analysis based on two-sample t-tests revealed a significant difference of Δβ_MB (p=0.0226) and Δα (p=0.0428) between HC and ED, with ED having a larger difference of Δβ_MB = 0.15 than HC (Δβ_MB = −0.1). Also, ED almost do not adjust their learning rate (Δα = 0.01), as compared to an increased learning rate in HC (Δα = −0.11). More detailed results of the statistical tests can be found in Table 10. There were no significant results for the Δρ parameter differences (supplementary Tables S15).

Results suggests that the body image preoccupation significantly reduces model-based learning in ED compared to HC. Moreover, ED does not exhibit any adjustment of learning rate due to body image preoccupation, compared to adjusted α in HC.

Δβ_MB as a predictor - results from regression analysis

The new variable, Δβ_MB, was used a predictor (digital biomarker) in different regression models of independent variables such as: group (logistic regression), as well as disordered eating, body image disturbance, and OCD-like behaviours based on the questionnaire scores (linear regression). The regression results are found in Table 11.

Different regression models show that the new measure Δβ_MB can be used to predict disordered eating, body image disturbance, as well as belonging to the ED group. However, the measure does not significantly predict OCD-like behaviours. The classification performance for the logistic regression model can be found in classification (contingency) Table 12, with HC as a negative class and ED as a positive class.

The logistic regression model appears to be good at predicting that a person is qualified as ED - 33 true positives and 5 false negatives. However, the model performs less accurately in predicting HC membership – 16 true negatives and 18 false positives.

Correlations of Δβ_MB with other covariates

Lastly, we checked if the new variable Δβ_MB correlates with any other covariates such as BMI, age, EAT-26, AAI, and OCI-R scores in the ED group. The correlation plots with PCCs and p-values for the hypothesis of no relationship between Δβ_MB and the covariates can be found in Figure 7.

• Download figure
• Open in new tab

Figure 7.

Correlation plot of BMI, age, EAT-26, AAI and OCI-R scores with Δβ_MB in ED group. Pearson’s correlation coefficient (PCC) and p-values are included.

The correlation analysis revealed that Δβ_MB has a significantly negative correlation with BMI (p=0.013) in ED group, which means that the higher the Δβ_MB the lower the BMI. For the other covariates there is no correlation with Δβ_MB in the ED group.

Further analysis of covariate correlation with Δβ_MB in both groups revealed two positive correlations with EAT-26 and AAI scores (Table 13, Figure 8). This shows the value of the predictor for characterising disordered behaviours across a population.

• Download figure
• Open in new tab

Figure 8.

Correlation plot of BMI, age, EAT-26, AAI and OCI-R scores with Δβ_MB across both groups. Pearson’s correlation coefficient (PCC) and p-values are included.

Implications

First of all, as far as we know, this is the first study that looked at the body image preoccupation from a computational perspective. We singled out and verified a potential digital biomarker of eating disorders. Our study complies with a novel RDoC framework (Insel et al., 2010) that aims to investigate and understand mental disorders by integrating many levels of information such as behaviour and self-report. Such a framework helps overcome the downsides of a classical psychiatric approach, namely, comorbidity² of disorders, heterogeneity³, and arbitrary definition of what constitutes a disorder (NIMH, n.d.). The proposed new biomarker is based in a rigorous and quantified behavioural paradigm that addresses an underlying mechanism of the disorder - body image preoccupation effect on model-based learning.

The fact that body image preoccupation is associated with decreased contribution of model-based learning provides some support for the hypothesised mechanism of extreme habitual body preoccupation. “I want to be skinny” once set as a goal, evolves over time into a rigid and habitual body checking. In the neutral condition, there is a fairly good amount of model-based learning contribution (though less so than in the healthy control group). However, when an individual’s body type is displayed onto a screen it may act as a trigger for the habitual behaviour of body comparison and concern. As such, a body-preoccupied state leads to significant reduction in the model-based capabilities that should be allocated towards task completion. Our result adds to the discussion reopened by Foerde et al. (2019) about domain-general/specific deficits in goal-directed learning. It provides support for the view that goal-directed control impairments are not specific to monetary tasks but suggests that it could be aggravated in conditions that trigger body image preoccupations.

Moreover, the new biomarker, Δβ_MB, could be used as a diagnostic tool, since it was shown to be a significant predictor of EAT-26 and AAI scores, as well as of belonging to an eating disorder group. In some cases, where obtaining a reliable and honest set of answer to questionnaires is not available, one could administer the two-step decision making task that includes the BID condition, even in the comfort of an individual’s home, and use the results from the task as a guideline for professional psychotherapists as to whether purse further treatment avenues.

The high heterogeneity (standard deviation) of parameter estimates (model-based/free) in the healthy control may suggest that some healthy controls share some traits with the ED group that explains reduced model-based learning. This might be related to traits that we have not directly explored in our study but that have been associated with deficits in model-based learning, e.g. alcohol addiction, or impulsivity (Gillan et al., 2016). We assume that compulsivity’s negative effect on model-based learning (Gillan et al., 2016) is not a factor here as we have tried to control for this by excluding participants with high OCI-R scores in the healthy control group. Alternatively, it is possible that some participants from a healthy control group do in fact have body image preoccupation issues. A potential factor in group misclassification could be a wide-spread and widely accepted societal preoccupation with dieting, looks as well as the phenomenon of body shaming that could render the recruitment of a ‘healthy’ population difficult (Kar, 2019; Sujoldzić & De Lucia, 2007). This could manifest in participants as selecting the option of “no past diet experience” in the screening stage despite ‘clean eating’ they might engage in as a widely accepted ‘health standard’, which is actually emotionally distressing and linked with functional impairments (Ambwani, Shippe, Gao, & Austin, 2019). As such, the heterogeneity in the parameter estimates in the healthy control may highlight the blurred boundary between health and dieting, and indeed call into question the possibility of a truly healthy control group within this field of research.

Lastly, the results can inspire modern therapy avenues to strengthen model-based learning capabilities and help relax rigidly formed disordered eating and body preoccupation habits. For example, Costin and Grabb (Costin & Grabb, 2012) suggest that within each sufferer of an eating disorder there are two ‘Selves’: The “Healthy Self” - a part of them that wants recovery and fosters healthy habits- and “the Eating Disorder Self”, which overtime gains its identity through repeated binges and/or restriction and depreciative, critical self-talk, becomes a strong and automatic driving force of disordered behaviour. The latter could correspond to the diminished contribution of model-based learning, present as a lack of recognition of all behavioural pathways available, namely the healthy ones that are more beneficial to an individual over time.

Authors suggest that in order to move towards recovery, sufferers should reinforce their Healthy Self. This is done through therapeutic activities that counter negative thoughts (and corresponding behaviours) with healthy thoughts (and corresponding behaviours). As a result, the sufferers begin to identify and choose new pathways. They find that the dire consequences as imagined by the Eating Disorder Self do not come to fruition, while the quality of life and relationship to body and food improves.

The principal idea behind such a therapeutic technique is that eventually the sufferers’ Healthy Self comes to dominate and assimilate into their one true Self. They are left more open to choosing paths that are - according to their (new, truly healthful) experience - more likely to be beneficial for their health, both mentally and physically. Essentially, this corresponds to reinforcing model-based learning by taking a look at one’s environment and following actions that are more thought-through and sufficiently analysed through a non-distorted lens. The results from our study offer additional support that people struggling with eating disorders accompanied by body image disturbance might benefit from such, and similar, therapeutic activities.

Limitations

Kool et al. (Kool, Cushman, & Gershman, 2016) suggest that the structure of the two step decision making task used here does not accurately estimate the trade-off between model-based and model-free learning. The authors propose certain modifications to the task to increase that accuracy such as changing the drift rate in the Gaussian random walk of reward probabilities, reducing the number of stage two choices to one per state, or introducing a deterministic transition structure. However, it has been shown that in some cases of slight variations to the task structure, the reinforcement learning model will not be able to distinguish between model-based and model-free actions as efficiently as in the standard task (Silva & Hare, 2018). More analysis would have to be performed for every variation, while the task in the current form has proven to yield reliable and consistent results (Foerde et al., 2019; Gillan et al., 2016; Voon et al., 2015).

An additional insight that should be taken into consideration is that a plain interaction between just model-based and model-free learning might not be enough to explain the choices and decision-making styles of individuals under the current conception of two step tasks as well as the models associated with it. Possibly, the surprising results about Δβ_MF can be explained by an additional learning system. Moreover, it is important to take extra precautions to ensure a solid understanding of the task by participants, as it has been shown that participants are likely to misconstrue the task. All together, it might be useful to consider additional learning strategies and models aside from the currently used model-based and model-free learning systems (Feher da Silva & Hare, 2020).

Two potential issues with the used modelling procedures are: a relatively low PCC of Δβ_MB (0.68) compared to other parameters (mean: 0.87); negative values of the model-based fractional reliance parameter, w, that is usually in a (0,1) range (Daw et al., 2011; Foerde et al., 2019; Kool et al., 2016). We believe that two issues are connected to each other. The link lies in the fact the estimates of Δβ_MB, Δβ_MF can be negative from the start due to an unsuitable prior distribution (Gaussian) they are drawn from. A better choice for the prior is a Gamma distribution, whose random variables are only positive and hence would constrain Δβ_MB, Δβ_MF to be greater than zero. Further, EM fitting algorithm would have to be slightly tweaked to take into account the new distribution in estimating its hyper-parameters. In turn, this would place the model-based fractional reliance w between 0 and 1, and would hopefully increase the parameter recovery thanks to a correct distribution.

What is more, the variable estimates of Δβ_MF parameter in both groups and of Δβ_MB in HC may suggest (apart from the difficulty in recruiting a truly healthy control) an insufficient number of participants given the online nature of the study. Ideally, more participants would have to be recruited to achieve a lower variability in the parameter estimates. In addition, a different model fitting technique could be used.

An alternative explanation to the reduction of model-based learning on the body image preoccupation condition in the eating disorder group can be attributed to a stress response rather than an effect that is strictly body image related. Otto, Raio, Chiang, Phelps, & Daw (2013) showed that as a result of a stress response (submerging a hand in ice cold water) model-based learning noticed a decreased contribution to decision-making. Given that body image preoccupation can be characterised as a distressing experience, the result from our study could be potentially explained by a generic stress response. However, to verify this alternative explanation, measures of the cortisol level for comparison between conditions would have to be taken. Also, it would be useful to perform a brain imaging study to identify which neural response is associated with reduced model-based learning in the BID condition and compare it to a hypothesised, stress-related response.

Further work

There are a couple of steps that could be taken in order to further explore the mechanisms and effects of body image preoccupation on decision-making in eating disorders. A natural step is to perform a very similar study as here, perhaps on a clinical population (rather than sub-clinical) and use brain imaging such as fMRI to try and correlate the behavioural-computational changes in the ED group with neural signatures to provide a neurobiological basis of body image preoccupation effects on decision-making. Ideally, in the future, a more whole understanding of eating disorders and accompanying body image issues could help devise new and more effective treatment options.

Furthermore, we suggest using a slightly more reliable model fitting technique for the two step task that utilises hierarchical Bayesian estimation methods, for instance hBayes package (Ahn, Haines, & Zhang, 2017). This approach has recently been shown to provide the most reliable estimation of model parameters for the two step decision-making task (Brown, Chen, Gillan, & Price, 2020).

If a more reliable fitting method is used, it might be possible to include an extra parameter that would extend the model used in this study. The extension was not performed here due to very low recovery of this new parameter using the EM fitting method. The new parameter would capture the reward sensitivity to see how much value the eating disorder group places on the silhouette that most resembles their own. It would be interesting to see how the reward sensitivity changes with model-based learning. The extension can be easily implemented by multiplying the reward term in Eq.: 1, 6 by a reward sensitivity parameter Ψ.