Phase-Transitions in Disgust to Self-Initiated Eating among Adolescents with Anorexia Nervosa and Related Low Weight Eating Disorders and Matched Controls

Abstract

Background:

Understanding the emotional context of feeding behavior may help identify causal mechanisms of food avoidance among individuals with anorexia nervosa. Although predominant food avoidance models assume fear of fat drives feeding behavior, disgust may be more theoretically and proximally relevant to moment-to-moment experiences of feeding. This study therefore aimed to examine affect and food avoidance using automated affect analysis from facial response by measuring time-specific transitions in disgust during a laboratory eating paradigm. We hypothesized that phase transitions in disgust would distinguish temporally self-initiated eating from food avoidance.

Methods:

Sixty three adolescents with anorexia nervosa or another low weight eating disorder (LWED) and 27 age and sex matched controls were recruited as part of a larger study; forty-five patients and 22 controls provided data on autonomous eating and facial affect during a laboratory meal. Dynamic structural equation models quantified moment-to-moment relationships between disgust and feeding behavior.

Results:

Self-initiated eating was associated with greater increases in disgust, but not fear, intensity among those with LWED relative to control participants and greater disgust intensity predicted lower likelihood of self-initiated eating.

Discussion:

Phasic transitions in disgust provide moment-to-moment evidence of affective influence on self-initiated eating and lend credibility to the hypothesis that disgust contributes to food avoidance and initiation in individuals with LWED.

1.0. Introduction

Pathological feeding is a key element of anorexia nervosa (AN). Studies of objective eating behavior show a pervasive pattern of abnormal behavior, including food avoidance, among individuals with AN.¹ Observing the laboratory feeding behavior of those with AN circumvents inaccuracies frequently noted when using self-report data,^2–4 including consistent over-reporting of calories consumed by patients with AN.⁵ Gianini and colleagues⁶ were able to use laboratory meals to also capture a range of behaviors indicative of delayed eating among individuals with AN that were absent among adolescents without AN. Autonomous-eating paradigms, in which an unfamiliar food is given to an individual with AN in place of a meal (e.g., shake of unknown ingredients in an opaque container), are a generalizable and replicable test of self-initiated eating that have been used to characterize limited consumption in the context treatment² and a positive relationship between pre-meal anxiety and food avoidance in patients with AN.⁷

Given these results, and the emergence of other relevant data, several theories were developed that hypothesize food avoidance emerges as a consequence of emotion, especially fear-based, effects on behavior^8–10 or a process of affective-experiential avoidance¹¹. However, fear-based models of food avoidance may not fully explain persistent food avoidance. Reductions in self-reported anxiety do not appear to yield robust changes in feeding.^12,13 An alternative line of research implicates disgust in the development and maintenance of eating disorder pathology.^14,15 Disgust refers to an aversive and distasteful reflex associated with distinct facial response (i.e., nose/lip curl), physiological reaction (nausea or visceral discomfort), and behavioral response (avoidance).¹⁶ Unlike fear, where rapid appraisal is prioritized, disgust requires a longer and more complex evaluation, yielding prolonged allocation of attention.¹⁷ The aversive—disgust response is highly conserved across mammals and appears to primarily function to facilitate avoidance of exposure to pathogen infection or poison from harmful food.¹⁸ Consequently, disgust may perpetuate food avoidance in those with AN, especially if “feeling fat” is conceptualized as personally or socially toxic. Empirical support for this theoretical extension exists at the trait,¹⁹ neuronal^20,21, psychophysiological^22,23 and behavioral level.²⁴ Despite this cumulating evidence supporting a general role for disgust in differentiating those with AN from individuals without psychiatric diagnosis, limited data are available to understand the influence of disgust on eating behaviors, including objective measures of feeding.

Extant data supporting a role for emotion in food avoidance behavior is largely derived from ecological momentary assessment (EMA) studies,²⁵ which generally measure variables in the hours around different target behaviors and require self-reporting of restriction. The time-scale resolution of affect and self-reported feeding prevent fine-grain analysis of feeding behavior. To overcome these obstacles, we used automated affect analysis derived from facial coding of emotion, which generates scaled intensity scores for disgust and fear based on the observed facial movements captured by high resolution camera recording²⁶. The timescale of video recording (30 fps) allows for better resolution of time specific transitions in affect that occur before and after feeding behavior and overcomes the resolution problem in existing EMA research. To capture the dynamic, time-specific effects of these moment-to-moment fluctuations and their relationships in time to eating, we applied dynamic structural equation modeling (DSEM).²⁷ This method allows for quantification of state level fluctuations in a time series (i.e., inertia) and deviations from this state (i.e., innovations) in both lagged and stationary random effects. The movement between steady state and dynamic fluctuations where both affect and behavior occur can be conceptualized as a phase transition. This nomenclature originates from the field of physics where a system experiences an abrupt and discontinuous shift in the properties or dynamics of the system (e.g., when H₂0 transitions from liquid to ice, the molecules go through a phase transition). Consequently, we can conceptualize phase transitions in disgust as time-specific changes in intensity of disgust that serve to change the way an individual interacts with food in the environment; DSEM allows for quantification of that transition through identification of perturbations in disgust in proximity to feeding behavior. Based on the theoretical and observational data supporting a unique role for disgust in food avoidance, we hypothesized that moment-to-moment changes in disgust would correlate with objective behavior (sips of a shake) in the context of laboratory feeding. Specifically, we hypothesized that time-specific elevations in disgust intensity (phase transitions) would follow feeding and to the degree that this occurred, less sips would be initiated.

2. Method

2.1. Participants

Patients with anorexia nervosa (AN) or another low weight eating disorder (LWED) and age-matched age and sex matched controls without an eating disorder (CON) were recruited for the study. Patients were primarily recruited from individuals seeking treatment at the Mount Sinai Eating and Weight Disorders Program, as well as from pediatricians, hospitals, and school counseling centers. Recruitment for the CON group included local street fairs, word of mouth, and posting flyers on community boards. Inclusion criteria for the clinical group included: i) female; ii) aged 12 to 18 years; iii) seeking treatment; iv) failure to maintain greater than minimally low body weight based on body mass index for age percentile and growth trajectory; v) clinically significant food restriction endorsed on the EDA-5; and vi) permission from pediatrician or other medical provider to receive outpatient care. Controls were 12 to 18 year-old females with no history of eating problems. Exclusion criteria for all participants included: i) history of a major psychiatric disorder based on responses on the K-SADS;²⁸ ii) history of a learning disorder or developmental disability; iii) report of active suicidal ideation; iv) current substance abuse; v) medications that impact interoception (e.g., anti-emetics) or performance (e.g., psychostimulants); and vi) medical conditions that influence eating or weight (e.g., Crohn’s disease). In accordance with Institutional Review Board regulations, parental consent was obtained for all participants.

Based on the inclusion and exclusion criteria, 63 adolescent female participants with a LWED and 27 sex and age-matched comparison controls (CON) were recruited for a larger randomized-controlled trial (NCT02795455). All participants were compensated a total of $279.80 for participating in the study. In addition, participants LWED were offered 20 sessions of free open-label family-based treatment through the Mount Sinai Eating and Weight Disorders Clinic and received additional $50 for completing 6 months post-treatment follow-up assessment, making the total compensation as $329.80. The final sample for the current project, based on task completion and video capture with face-on viewing of the camera, included 45 participants with LWED aged 12 to 18 (M = 15.2, SD = 1.92) and 22 CONs aged 12 to 18 (M = 16.1, SD = 1.49). All patients who provided data met criteria for AN (N=43) or Atypical AN (N=2) by EDA-5. The BMI range for participants with LWED was 13.0–21.9 (M = 17.3, SD = 1.94; 12% for age/sex) and 16.9–28.8 (M = 20.7, SD = 3.02; 58% for age/sex) for control participants. In the LWED group, 77.8% of the participants were White, while 40.9% and 27.3% of the CON participants were White and Asian, respectively. Demographic details are provided in Table 1.

Table 1:

Demographics by Group

	LWED (N=45)	CON (N=22)	Overall (N=67)
Age
Mean (SD)	15.2 (1.92)	16.1 (1.49)	15.5 (1.84)
Median [Min, Max]	16.0 [12.0, 18.0]	16.0 [12.0, 18.0]	16.0 [12.0, 18.0]
BMI
Mean (SD)	17.3 (1.94)	20.7 (3.02)	18.4 (2.83)
Median [Min, Max]	17.2 [13.0, 21.9]	20.0 [16.9, 28.8]	18.0 [13.0, 28.8]
Race
Native American/Alaska Native	0 (0%)	0 (0%)	0 (0%)
Asian	2 (4.4%)	6 (27.3%)	8 (11.9%)
Black/African American	2 (4.4%)	1 (4.5%)	3 (4.5%)
Native Hawaiian/Pacific Islander	0 (0%)	0 (0%)	0 (0%)
White	35 (77.8%)	9 (40.9%)	44 (65.7%)
Other	4 (8.9%)	4 (18.2%)	8 (11.9%)
Multi-Racial	1 (2.2%)	2 (9.1%)	3 (4.5%)
Not reported	1 (2.2%)	0 (0%)	1 (1.5%)
Ethnicity
Not Hispanic or Latino	38 (84.4%)	17 (77.3%)	55 (82.1%)
Hispanic or Latino	6 (13.3%)	5 (22.7%)	11 (16.4%)
Not reported	1 (2.2%)	0 (0%)	1 (1.5%)

Note. BMI = body mass index. SD = standard deviation.

2.2. Procedures

Following an eligibility screening, participants completed self-reported demographics and questionnaires. Autonomous eating behaviors with emotional responses were assessed using the Single Item Meal (SIM) task with FaceReader monitoring. The SIM task was administered at both baseline and 8-week assessments, and only the baseline data were included in the current project. Video recording occurred through a laptop camera placed approximately 30-inches from the straw that participants used in the SIM task.

2.2.1. Single Item Meal (SIM) Task

The SIM task is a standardized laboratory test meal that has a long history as an objective assessment of feeding behavior in adolescents with low-weight eating disorders.² Participants were presented with a 55 fluid ounce (1626.50 mL) covered opaque container containing approximately 1475 grams (1 kcal per gram, or 1475 kcal) of strawberry yogurt shake. Although informed that the meal consisted of a strawberry yogurt shake, participants were not told the quantity provided in the container. Participants were instructed to consume the shake through a straw while refraining themselves from touching or otherwise manipulating the container. They were told to consume the shake as a replacement for their breakfast/lunch/dinner/snack for the day and to eat as much as they desired. All participants had been instructed to avoid eating for at least 3 hours prior to laboratory assessment. Parents were not informed of the amount consumed to preserve measure of self-initiated (autonomous) eating. The SIM was placed on a modified version of an eating monitor, which measured intake (in grams) every 5 seconds. A closed-circuit video monitor and the FaceReader software were used to ensure participants followed instructions and to observe their eating behaviors and emotional responses.

2.2.2. FaceReader Measure of Moment-to-Moment Affect

FaceReader v 9.0²⁹ is a software program that uses video capture to assess intensity of emotional expression through facial movement and quantifies the intensity of seven core emotions (i.e., happiness, sadness, anger, surprise, fear, disgust, and neutral).³⁰ The software has been extensively tested^31,32. For instance, Lewinski, den Uyl, and Butler³³ found in two independent data sets that FaceReader matched human performance in recognizing basic emotions with an accuracy rate of 88 to 89 percent.

In the current study, video recordings of the Single Item Meal (SIM) (i.e., shake) were processed using Facial Action Coding System (FACS), which categorizes facial muscle movements into specific Action Units.³⁴ It works in three steps: face finding, which detects the position of a face; face modeling, which utilizes the Active Appearance Model (AAM) to create an accurate artificial face model using 500 key points on the faces; and face classification, which classifies the expressions into one of seven core emotional with intensity (range = 0–1).^30,35–37 The FaceReader eating module, which tags facial movements associated with ingestive behavior, also detected the targeted behaviors of “sip”, which indicated participants’ consumption of the shake and categorized the behaviors into either “No Event Marker” or “sip”. Both facial expressions and behaviors in the video recordings were captured for 30 samples/second (i.e., time interval for each sample is approximately 0.033 second) by the FaceReader and all sip ratings were confirmed by independent observation of video to prevent any automation mistakes in coding. No mistakes were found in visual confirmation of the video. Video recordings sampled faces at 30 frames per second (30 fps). To reduce noise and computational time, raw data were downsampled to 3 frames per second and sip data collapsed into single row to avoid modeling of facial movements during objective feeding as affective. For the proposed study, we extracted “disgust” and “fear” data for analysis.

2.2.3. Disgust Scale-Revised (DSR)

Participants completed the Disgust Scale-Revised (DS-R)³⁸, a 27-item self-assessment tool that measures disgust propensity across three domains: core, animal-reminder, and contamination. Responses were given on a 5-point Likert-type scale from 0 (no disgust or repugnance at all) to 4 (extreme disgust or repugnance).

2.3. Statistical Analysis

We conducted a series of multi-level dynamic structural equation models (DSEM)²⁷ to test the moment-to-moment relationship between disgust, fear, and self-initiated eating. We followed the model building steps recommended by Asparahaouv and Muthen³⁹ (all scripts and primary output are in supplementary material). These models allow for the incorporation of autoregressive effects (referred to as inertia) or stability of variables collected on time-series and deviations from the autoregressive effect (referred to as innovations). In our model we included two time series variables (sip) and (disgust or fear intensity derived from Facereader analysis). We built increasingly complex models examining the correlated and cross-lagged effects of disgust and sip, further examined the random (between-subject) variability in autocorrelation and cross-lagged effects between sip and disgust. We then tested whether these between-subject effects were different between group (LWED vs. CON). All models were estimated with Bayesian algorithm and weakly informed priors using Mplus v8.7⁴⁰. Scripts used for the analysis are provided in supplementary material. Our primary hypothesis was that Group (LWED vs. CON) would be associated with greater innovations in disgust following sip (cross-lagged effect). Disgust and Fear intensity scores were log transformed and multiplied by 100 to aid in convergence and account for skew. We also calculated plausible values for each individual on the random effects and correlated these with DS-R total score, hypothesizing that DS-R would be positively associated with disgust innovations. We also conducted the same model building process for fear. Formal power analyses were not conducted as these data were already collected as part of a clinical trial, although we note that our sample size and within-subject observations are well within sample size recommendations for these statistical models.⁴¹ Raw data are not able to be de-identified. Processed data available from the first author upon request.

3.0. Results

All disgust models converged well as evidenced by potential scale reduction factors and visual examination of trace plots of parameter estimation. Fear models were unstable in estimation and results, with no evidence of lagged or stationary effects on eating. The majority of frames were non-missing (91.5%), with 8.4% missing data on Disgust and the remaining missing Sip data. A total of 873 Sips were available for modeling and 177,548 frames where no sips occurred. Mean disgust by frame was 1.27 (SD=3.4; range 0–95.26). Total number of sips were significantly different by group (OR = 0.532, SE = 0.037, 95%CI = 0.465, 0.608) indicating there was approximately 50% lower odds of taking a sip among those diagnosed with LWED.

3.1. Model Building

We estimated models starting with simple two-level model confirming evidence of autocorrelation (φ) of each time series (sip and disgust) and evidence of nonzero effects for slopes (cross lagged and time varying effects). We then examined the extension of this model to include random effects to the autocorrelation, slopes, and residual variance and retained only those random effects that were nonzero. We also compared the time-varying versus cross-lagged model modeling of the relationship between disgust or fear on sip. Inclusion of the time-varying effect led to inconsistent convergence of the disgust model estimation and was of less interest than the cross-lagged effect of disgust on sip. Fear models including either time-varying or lagged effects were unable to converge on non-significant. Inclusion of the lagged effect of disgust on sip led to stable convergence of the model and thus became the focus of further model development (inclusion of both cross-lagged and time varying effects aren’t identified and can’t be estimated). The final model included the group effects as a between level covariate. Figure 1 summarizes the final model and estimated parameters.

Figure 1.

The final model including within-subject and between-subject decomposition of Sip and Disgust intensity into autoregressive ($\Phi$) and slopes (β) and variances $\psi$. The random effects are denoted by dark circles in the top panel (within) and the shaded circles in the bottom panel (between). Transparent lines reflect those dropped from the model. The between level effects are regressed on group (anorexia nervosa vs. age and sex-matched controls without eating disorder).

3.2. Phase Transitions in Affect and Feeding Behavior

Table 2 summarizes the parameters for the cross-lagged autoregressive model with random effects. The inertias (φ_sip, φ_dis) of both sip and disgust indicated that feeding behavior and disgust intensity had stable correlations over time, but that the inertia for Sip had evidence of credible between-subject variability. The credibility intervals for the random autocorrelation of disgust included zero indicating the majority of variability in inertia occurred within-subjects. In contrast, there was evidence of both significant within-subject and between-subject variability in inertia of Sip. The cross-lagged effects (β_{sip->disgust+1}, β_{disgust->sip+1}) indicated higher disgust was associated with reduced probability in taking a sip and taking a sip led to elevated disgust. The latter phase transition supports the hypothesized model, but the former suggests that disgust intensity also has a prospective (not just reactionary) association with feeding behavior.

Table 2.

Standardized Model Results of Cross-lagged Dynamic Structural Equation Model of Disgust

Parameter	Estimate	95% Credibility Interval
Within-Level Standardized Estimates Averaged Over Individuals
φsip (inertia)	0.184	[0.148, 0.219] *
β sip->dis (cross-lagged effect/innovation)	0.147	[0.139, 0.156] *
φdis (inertia)	0.997	[0.995, 1.000] *
βdis->sip (cross-lagged effect/innovation)	−0.144	[−0.179, −0.104] *
ψdis	0.031	[0.029, 0.034] *
Between-Level Standardized Estimates for Covariate (Group)
γφsip Group	−0.270	[−0.430, −0.083] *
γsip->dis Group	0.359	[0.191, 0.497] *
γφdis Group	−0.125	[−0.328, 0.085]
γdis->sip Group	−0.201	[−0.378, −0.006] *
βsip Group	−0.214	[−0.371, −0.047]*
Intercepts
γdis	1.817	[1.195, 2.477] *
γφsip	0.910	[0.467, 1.304] *
γsip->dis	−0.224	[−0.682, 0.196]
γdis->sip	0.016	[−0.355, 0.414]
Residual Variances
ψsip	0.954	[0.863, 0.998] *
ψdis	0.983	[0.893, 1.000] *
ψφsip	0.928	[0.816, 0.993] *
ψsip->dis	0.960	[0.857, 0.999] *
ψdis->sip	0.871	[0.753, 0.963] *
Within-Level R-Square Averaged Across Individuals
Sip	0.325	[0.300, 0.355]
Disgust Intensity	0.969	[0.966, 0.971]
Between Level R-Square
Sip	0.046	[0.002, 0.137]
Disgust Intensity	0.017	[0.000, 0.107]
φsip	0.072	[0.007, 0.184]
β sip->dis	0.129	[0.037, 0.247]
β dis->sip	0.040	[0.001, 0.142]

Note. All results reported in standardized effects. Dis = disgust intensity, which was transformed into log(1+disgust) to aid in model convergence.

^*significant one-tailed .05 based on posterior standard deviation.

3.3. Covariate Effects on Phase Transition

The effect of group on inertia and innovations indicated that those with LWED had greater phase transition innovations in disgust intensity than controls (γ =0.36, PSD 0.08, 95%CI = 0.19, 0.50). This effect provided support for our hypothesized effect of disgust on feeding behavior. In addition, there was evidence of a more robust negative effect of disgust on likelihood to take a sip (γ =−0.27, PSD = 0.09, 95%CI = −0.43, −0.08) among the members of the LWED group. The group effects indicate that disgust phase transitions are more closely linked to feeding behavior in those with LWED and supported by the evidence of lower between-subject variability in disgust intensity inertia (γ =−0.20, PSD 0.09, 95%CI = −0.38, −0.01). Figure 2 illustrates these differences between groups in both fear (which was not significantly related to sip) and disgust in a single LWED participant and matched control. The γ parameters of the model capture increased variability in affect intensity surrounding sips for the LWED participant but not the control participant.

Figure 2.

Intensity of disgust and fear in two matched participants throughout laboratory assessment of single item meal. Intensity of emotion is scaled from 0–1 and time is measured in 1/30 second. Sips were all initiated by participant and reflect the consumption of food by straw of an opaque shake of unknown caloric or macronutrient content. Phase transitions in affect occur proximal to the sip and are modelled by the innovation parameter of dynamic structural equation models.

Table 3 includes parameters of the best model for Fear where convergence criteria were met. In contrast to our disgust model, no significant lagged effects were found, although significant group differences in inertia of fear intensity (γ =−0.33, PSD = 0.12, 95%CI = −0.55, −0.09) suggesting that variability in correlations from moment-to-moment fear in the LWED group. In contrast, there was no significant difference between group difference in the lagged-effect of sip to fear (γ =−0.13, PSD = 0.30, 95%CI = −0.63, 0.52). The magnitude of these effects are also evident in the Figure 2 tracings of fear intensity.

Table 3.

Standardized Model Results of Cross-lagged Dynamic Structural Equation Model of Fear

Parameter	Estimate	95% Credibility Interval
Within-Level Standardized Estimates Averaged Over Individuals
φsip (inertia)	−0.155	[−0.221,− 0.063] *
φFear (inertia)	0.947	[0.945, 0.948] *
βFear->sip (cross-lagged effect/innovation)	0.006	[−0.032, 0.045]
ψFear (residual variance)	0.104	[0.100, 0.106] *
Between-Level Standardized Estimates for Covariate (Group)
γφFear Group	−0.334	[−0.546, −0.090]
γFear->sip Group	−0.129	[−0.633, 0.519]
βsip Group	−0.458	[−0.650, −0.222]*
Intercepts/Means
γFear	2.509	[1.195, 2.477] *
μφsip	−2.390	[−5.389, −0.619] *
μgroup	1.394	[1.065, 1.731] *
γFear->sip	0.301	[−0.585, 1.126]
Residual Variances
ψsip	0.791	[0.863, 0.998] *
ψFear	0.889	[0.702, 0.992] *
ψFear->sip	0.946	[0.558, 1.00] *
Within-Level R-Square Averaged Across Individuals
Sip	0.031	[0.010, 0.054]
Fear Intensity	0.896	[0.894, 0.900]
Between Level R-Square
Sip	0.209	[0.049, 0.422]
Fear Intensity	0.111	[0.008, 0.298]
β Fear->sip	0.054	[0.000, 0.442]

Note. All results reported in standardized effects. Fear = fear intensity, which was transformed into log(1+fear) to aid in model convergence.

^*significant one-tailed .05 based on posterior standard deviation.

3.4. Generalization of Phase Transition Findings

We estimated Pearson correlations between plausible individual values for model estimated parameters and disgust propensity ratings. Disgust sensitivity was not significantly correlated with φ_sip (r = .24, t = 0.19, p = .85) or the pre sip (r = .24, t = 0.19, p = .85) and post sip (r = .07, t = 0.54, p = .59) variability in in disgust intensity. These findings were inconsistent with our hypothesized model.

4.0. Discussion

We found initial support for the hypothesis that phase-transitions in disgust follows feeding behavior among adolescents with LWED. These transitions appear to deviate from stable fluctuations over time and occur immediately before and following eating. The finding that this transition occurs to a greater extent among those with a LWED suggests it may be a unique response to feeding that limits food intake. Although we did not hypothesize pre-sip influences of disgust intensity on probability of feeding, our model demonstrates initial evidence of pre-meal disgust increases reducing the probability of feeding. We could extend this initial evidence to theorize that disgust acts as an aversive governor on feeding behavior, reducing its likelihood by reducing motivation to initiate feeding and coding the aversive nature of food consumed. This ‘bookending’ of feeding may provide an obstacle to regular or sustained feeding among those with LWEDs by reducing the value of seeking and consuming food during a meal.

We hypothesized that the dynamic pattern of behavior would also correlate with trait measure of disgust propensity, but there was no evidence of this effect. The dynamic relationship between disgust and feeding behavior was predicted to occur for the LWED patients, but not our control participants. This prediction did not presume that the control participants are resistant to the bookending effect observed among the LWED group, but that the relatively pleasant milkshake would not elicit an anticipatory or reactionary disgust response in those with otherwise functional feeding behavior. Manipulating disgusting vs. palatable foods might help tease apart the state-trait like differences in propensity for disgust regulation of food intake as well as relationships to self-reported disgust traits.

The relationships between fear and feeding were not clear from the moment-to-moment analysis of behavior. The absence of findings should be cautiously interpreted. A null finding cannot be used to support the absence of an effect and there remains an open question about whether fear plays an additive or independent role in the sequence of behaviors that control ingestion. It is possible, or even likely, that fear operates at a different stage of the approach—avoidance continuum to influence food avoidance (e.g., well before feeding opportunities exist, such as planning to attend a birthday party with cake). Disgust, in contrast, appears to have a strong proximal relationship with feeding behavior, so may influence larger patterns of food avoidance by coding the aversive experience and reducing value of feeding.

Disgust has recently become a potential treatment target for those with LWED⁴² and has been formally integrated into a form of family therapy.⁴³ Disrupting this governing process is likely to be difficult in the context of starvation, particularly because the goals of refeeding require consumption of large volume of food consistently for weeks to months in order to restore physical health. Disgust is also likely to be an important part of broader issues with self-identity and body image. Glashouwer and de Jong⁴⁴ have extended the disgust model to include bodily perception and the ‘fear of fatness’ characteristic of the disorder. Recent evidence also indicates that disgust accounts for correlations between negative affect on eating disorder symptoms⁴⁵. Furthermore, disgust has a unique physiology that allows it to be directly connected to the frequently observed disturbances in gastrointestinal discomfort and function among patients with AN.⁴⁶

The role of affect in expression and dynamics of food avoidance are important to understanding how feeding disturbances function among patients with AN and related LWED. Much of the moment-to-moment changes in affect that occur among patients are lost to timescale of measurement methods or subject to self-report limitations. The availability of laboratory feedings studies offers a solution to these problems when integrated with passive high-resolution data collection. Because DSEM methods allow for the study of random effects as they vary over time, higher-resolution time-series modeling is possible. This approach is also computationally expensive, limiting the number of random effects available for incorporation into the model. Further investigation of this current model should examine alternative emotion or even psychophysiological measures that can scale computationally to examine the interplay of different emotions on behavior. There is some interest in incorporating electrophysiological signals into fMRI so that specific emotional responses can be linked to neurocircuit functioning.²³

Despite the evidence to support our hypothesis, there were several limitations that warrant consideration in interpreting the statistical model. There were very few sips relative to the near continuous sampling of affect. The sparsity of the behavior may be better examined in a series of hurdle or hazard models where the time between events can be incorporated directly into the model with phase-transition in disgust predicting longer duration between sips. An alternative approach would be to standardize sip time and frequency (although not amount consumed), so that sparsity could be controlled across participants. In addition, the sampling frequency of the affect may be optimized as higher resolution cameras and time-scales are possible using existing software and hardware. An alternative set-up might yield more salient results or alternatively reveal other associations. Because of these limitations, we believe that these findings should be cautiously interpreted and followed-up with more thorough testing of the model.

The future applications of DSEM methods and integrated passive affect coding and eating behavior may help build more complex and integrated theoretical models of affect and food avoidance. For instance, extending the disgust governance model of feeding to disgust-food cue learning²² or the valuation of safety learning with pleasure.⁴⁷ It would be possible to additionally use this method to examine other physiological markers such as stress reactivity, eye-blink startle, or interoceptive signals (heart rate, gastrointestinal changes) in efforts to integrate individual and general sources of disruption in feeding behavior.