ABSTRACT
Objective
Negative affect is a primary antecedent for binge eating (BE). Just‐in‐time adaptive interventions (JITAIs) prompt the use of therapy skills when at risk for maladaptive behaviors. JITAIs show promise for improving emotion regulation skill use and decreasing BE. Identifying momentary negative affect has relied on ecological momentary assessment (EMA). EMA surveys are typically delivered 4–6 times/day via smartphone and are self‐report; accordingly, EMA is limited by temporal granularity, participant insight, and adherence. Wearable sensors which continuously, passively measure physiological correlates of affect may be able to detect risk for negative affect‐related BE while overcoming limitations of EMA. This study compared machine learning models using EMA‐ and sensor‐measured negative affect to predict BE.
Method
Thirty adults with recurrent BE wore smartwatches to measure heart rate and electrodermal activity and reported affect and BE on EMA for 4 weeks (preregistration: https://www.researchprotocols.org/2023/1/e47098/). Support vector machines, random forest, and neural network models were trained using the first 3 weeks and evaluated using the last week of EMA data, sensor data, and combined EMA and sensor data.
Results
The best‐performing EMA‐only model for predicting BE had macro‐averaged accuracy of 0.64 (AUROC = 0.69, sensitivity = 0.47, specificity = 0.81), the sensor‐only model had macro‐averaged accuracy of 0.64 (AUROC = 0.78, sensitivity = 0.92, specificity = 0.35), and the combined model had macro‐averaged accuracy of 0.61 (AUROC = 0.68, sensitivity = 0.90, specificity = 0.33); model macro‐averaged accuracies did not significantly differ.
Discussion
Psychophysiological sensor data demonstrate comparable accuracy to EMA for predicting BE, setting the stage for low‐burden JITAIs targeting negative affect‐related BE.
Keywords: affect, affect regulation, binge eating, binge eating disorder, bulimia nervosa, machine learning, OSFED, psychophysiology
1. Introduction
Binge‐spectrum eating disorders (EDs), including bulimia nervosa and binge‐eating disorder, are characterized by recurrent engagement in binge eating (BE; American Psychiatric Association 2013). BE refers to a behavior in which an individual consumes an unusually large quantity of food and experiences a sense of loss of control over eating (American Psychiatric Association 2013). These disorders affect approximately 13 million US adults and are associated with significant physical and psychological consequences and public health costs (Deloitte Access Economics 2020; Hudson et al. 2007; Kessler et al. 2013). Despite these severe consequences, the frontline treatment for BE, Enhanced Cognitive‐Behavioral therapy (CBT‐E), produces remission in only 30%–60% of individuals who undergo a full course of treatment (Linardon 2018; Peat et al. 2017).
Negative affect refers to the experience of aversive emotions. The affect regulation model of BE postulates that negative affect serves as a major momentary risk factor for BE. Individuals with binge‐spectrum eating disorders engage in BE to alleviate negative affect, as these individuals have difficulty tolerating negative emotions and lack adaptive emotion regulation strategies; when negative affect is alleviated by BE, the behavior is negatively reinforced, increasing the likelihood of BE the next time the individual experiences negative affect (Fairburn 2008). Several previous studies using ecological momentary assessment (EMA, smartphone surveys delivered multiple times per day) have found support for the affect regulation model among individuals with anorexia nervosa (Berg et al. 2017), bulimia nervosa (Berg et al. 2017, 2013; Crosby et al. 2009; Lavender et al. 2016; Smyth et al. 2007), binge‐eating disorder (Schaefer et al. 2020), and comorbid BE and obesity (Berg et al. 2015; Ranzenhofer et al. 2014), substantiating its role as a transdiagnostic mechanism underlying BE. There are several effective intervention strategies for improving affect regulation that could be used in treatment for BE to potentially improve outcomes (Blackledge and Hayes 2001; McMain et al. 2001), especially if delivered at times when an individual might be at risk for negative affect‐related binge eating. However, several questions remain about the role of affect in BE, largely due to limitations related to commonly used approaches for measurement of negative affect.
Literature examining the affect regulation model of BE in real‐time has relied on EMA to assess BE and contextual risk factors (Berg et al. 2017, 2015, 2013; Crosby et al. 2009; Lavender et al. 2016; Ranzenhofer et al. 2014; Schaefer et al. 2020; Smyth et al. 2007, 2009). Although EMA facilitates examination of the impact of affect on BE within‐day, these surveys are completed only every few hours, preventing the examination of more granular (i.e., minute‐to‐minute) changes in affect (Kockler et al. 2018). Some inconsistencies in previous research examining the relationship between affect and BE using EMA have been attributed to limitations in the capacity of EMA to examine highly proximal changes to affect before and after BE (Berg et al. 2017; Crosby et al. 2009; Engel et al. 2013; Hilbert et al. 2009; Ranzenhofer et al. 2014; Smith et al. 2020; Smyth et al. 2007). These limitations that are inherent to EMA methods could be addressed by continuously measuring affect.
Furthermore, existing research using EMA to study BE has only focused on affective valence, the degree to which an emotional state is experienced as pleasant. Within Bradley et al. (1992) dimensional vector theory of affect, affect states are characterized by both affective valence and affective arousal (i.e., the degree of affect‐related physiological activity). Examination of affective arousal and affective valence may facilitate more accurate momentary prediction of BE, as literature has linked increased affective arousal to increases in eating‐related disinhibition (Cools et al. 1992) and to the aversiveness of negative affect states (Marchewka et al. 2014). Accordingly, high‐arousal negative affect states (e.g., fear and anger) may impact BE risk to a greater degree than lower‐arousal negative affect states (e.g., shame).
Ambulatory, wearable passive sensing could address the limitations of EMA for measuring negative affect. Sensors could augment or replace EMA for collecting momentary data on affect as they facilitate objective, low‐burden measurement of biomarkers of affective arousal on a continuous timescale. The continuous timescale on which sensors measure physiological correlates of affect allow for examination of affect immediately prior to, during, and following EMA‐reported BE episodes. Individuals with eating disorders have difficulties recognizing and labeling their emotions (Lavender et al. 2015; Svaldi et al. 2012; Weinbach et al. 2018), which might inhibit timely recognition of affect‐related risk for BE; sensor data could address this limitation and facilitate earlier detection of BE risk by detecting biomarkers associated with negative affect states without reliance on the individual's subjective awareness of their emotions. Finally, sensors are low‐burden and require minimal user engagement (i.e., wearables simply require the individual to put on the smartwatch) relative to EMA, which may improve compliance and data completeness and reduce the requisite frequency of EMA.
Heart rate, heart rate variability, and electrodermal activity are physiological correlates of negative affect (i.e., physiological features that change during affect states) that can be measured using commercially‐available wearable sensors (Frazier et al. 2004; Kanjo et al. 2015). These biomarkers reflect sympathetic nervous system arousal (e.g., elevated heart rate and increased electrodermal activity) and parasympathetic nervous system inactivation (e.g., decreased heart rate variability) associated with affective arousal. A robust body of literature has identified consistent patterns of change in heart rate, heart rate variability, and electrodermal activity during negative affect states in healthy adult populations (Dissanayake et al. 2019; Golland et al. 2014; Gouizi et al. 2011; Hasnul et al. 2021; Jang et al. 2019, 2015; Kreibig 2010; Valderas et al. 2019). However, only three previous study has used ambulatory sensors to assess heart rate metrics prior to BE episodes (Ranzenhofer et al. 2016, 2021; Shi et al. 2024). Both studies by Ranzenhofer et al. (2016, 2021) found that heart rate was elevated and heart rate variability decreased prior to loss of control eating episodes relative to non‐loss of control eating episodes. Conversely, the study by Shi et al. (2024) identified no significant average group‐level change in heart rate prior to BE episodes compared to control time periods, although significant random effects indicate that individual participants exhibited varying differences in heart rate prior BE episodes compared to control periods. These findings offer preliminary evidence that heart rate and heart rate variability may change in the proximal time period prior to BE episodes, suggesting that they may constitute biomarkers of affect that confer BE risk. Applying these sensors to collect ecological momentary data on affective arousal could therefore improve the field's understanding of the role of affect in maintaining BE and capacity to accurately predict future BE.
Given the sampling frequency of wearable sensors (up to multiple times per second) and the variability in physiological metrics assessed (e.g., resting heart rate can range from 60 to 100 beats/min in healthy adults; American Heart Association 2015), these devices produce datasets that are very large (often containing millions of datapoints) and highly dimensional. Machine learning is a class of data analytic approaches that seek to predict or forecast an outcome of interest (e.g., BE) based on a number of input variables (e.g., EMA ratings of affect, heart rate, electrodermal activity, etc.). Machine learning models are particularly well‐suited to use complex time series data (like those yielded by sensors) to predict a behavioral outcome (Engelhard et al. 2021), as these models learn from data and can improve prediction accuracy as additional data is collected.
If sensor data can improve our capacity to predict BE risk, this innovation will offer transformative potential to deliver ecological momentary interventions (EMIs) targeting negative affect‐related BE risk with minimal user burden. More accurate prediction of BE episodes could allow for delivery of EMIs that promote the real‐time use of therapeutic skills (e.g., emotion regulation skills) to reduce BE when an individual is experiencing negative affect. Such an EMI system may be particularly useful for individuals who have difficulties with self‐regulation (i.e., emotion regulation and impulsivity), based on previous research identifying that self‐regulation moderated treatment response to an EMI system for bulimia nervosa that delivered interventions in response to EMA‐reported risk factors for BE (Presseller et al. 2022). Previous research testing EMIs for BE have demonstrated high feasibility and acceptability of these intervention systems (Juarascio, Presseller, et al. 2022; Juarascio, Srivastava, et al. 2022).
1.1. Aims and Hypotheses
This study examined the comparative efficacy of EMA and wearable psychophysiological sensors measuring aspects of negative affect for the prediction of negative affect‐related BE. The study developed machine learning algorithms using EMA data measuring negative affect alone, sensor data measuring affective arousal alone, and EMA and sensor data together to predict BE episodes to understand the relative accuracy and additive value of sensor data as an augmentation to EMA. We hypothesized that the combined sensor and EMA‐based machine learning algorithm would be significantly more accurate for predicting BE than the algorithms using only EMA or only sensor data. We hypothesized that the sensor‐only and EMA‐only machine learning algorithms would not significantly differ in accuracy for predicting BE. Based on past findings using sensors to predict emotional eating and dietary lapses, we hypothesized that the combined sensor and EMA algorithm would classify the occurrence of BE episodes with > 70% accuracy (Goldstein et al. 2018; Juarascio et al. 2020).
2. Method
The methods and planned statistical analytic approach for the study were preregistered via a registered report (Presseller et al. 2023).
2.1. Recruitment and Participants
Participants were 30 individuals with clinically significant BE recruited from across the United States. Eligibility criteria were at least 12 episodes of BE in the past 3 months, at least 12 instances of nonswimming exercise in the past 3 months, age 18–65 years old, body mass index (BMI) of at least 18.5 kg/m2, ownership of a smartphone, and willingness to wear the Fitbit Sense 2 watch and complete EMA. Exclusion criteria were inability to fluently speak, read, and write English, current eating disorder‐focused treatment, and intellectual disability or current, severe psychopathology (e.g., active suicidality and psychosis) that would inhibit completion of study protocols. One participant (3.3% of the total sample) discontinued participation in the study after 2 weeks of ecological data collection and was excluded from analyses. Table 1 depicts participant demographics and clinical information.
TABLE 1.
Participant demographics and clinical information.
| N | % | |
|---|---|---|
| Sex | ||
| Male | 13 | 43.3% |
| Female | 17 | 56.7% |
| Gender | ||
| Cis man | 13 | 43.3% |
| Cis woman | 16 | 53.3% |
| Trans man | 1 | 3.3% |
| Trans woman | 0 | 0.0% |
| Genderqueer/Gender nonconforming | 0 | 0.0% |
| Race | ||
| American Indian/Alaska Native | 0 | 0.0% |
| Asian | 5 | 16.7% |
| Hawaiian/Pacific Islander | 2 | 6.7% |
| Black/African American | 5 | 16.7% |
| White | 19 | 63.3% |
| More than one race | 2 | 6.7% |
| Unknown or prefer not to say | 1 | 3.3% |
| Ethnicity | ||
| Latinx/Hispanic | 5 | 16.7% |
| Not Latinx/Hispanic | 25 | 83.3% |
| Employment Status | ||
| Full time | 14 | 46.7% |
| Part time | 6 | 20.0% |
| Occasional/per diem | 3 | 10.0% |
| Disability/SSI | 0 | 0.0% |
| Unemployed/no income | 7 | 23.3% |
| DSM‐5 Eating Disorder Diagnosis | ||
| Bulimia nervosa | 16 | 53.3% |
| OSFED, bulimia nervosa (SBEs only) | 9 | 30.0% |
| Binge‐eating disorder | 3 | 10.0% |
| OSFED, binge‐eating disorder (SBEs only) | 2 | 6.7% |
| Mean (SD) | Range | |
|---|---|---|
| Age | 37.13 (10.23) | 20–62 |
| BMI | 28.77 (8.40) | 18.88–48.23 |
| Past Month Total Binge Episodes | 21.23 (19.73) | 3–84 |
| Past Month Total Compensatory Behaviors | 17.13 (11.84) | 0–46 |
| EDE Global Score | 2.82 (0.98) | 0.29–5.06 |
| EDE Restraint | 2.37 (1.24) | 0.00–5.00 |
| EDE Eating Concern | 1.92 (1.28) | 0.00–5.00 |
| EDE Shape Concern | 3.72 (1.16) | 0.75–5.75 |
| EDE Weight Concern | 3.27 (1.41) | 0.00–6.00 |
Note: Race categories are not mutually exclusive.
2.2. Procedures
Participants completed a phone screen and baseline assessment (including the Eatng Disorder Examination; Fairburn et al. 2014) to establish eligibility. Eligible participants wore Fitbit Sense 2 watches during all hours of the day and reported negative and positive affect and eating episodes on EMA surveys for 4 weeks. Fitbit accounts were created by the study team for each participant for the purpose of the study and study data were downloaded from these accounts. Participation was fully remote, and assessments were conducted using the institution's HIPAA‐compliant Zoom platform. Informed consent was obtained from all participants and study procedures were approved and overseen by the institution's Institutional Review Board. Participants received $20 for completion of the baseline assessment, up to $15/week for completion of EMA and wearing the smartwatches (with $1 subtracted for each day with EMA or sensor data compliance below 70%), and $20 for completion of the end of study visit.
2.3. Measures
2.3.1. Eating Disorder Examination
The Eating Disorder Examination (EDE) is a very widely used, semi‐structured diagnostic interview for the assessment of eating disorders (Fairburn et al. 2014) and demonstrates test–retest reliability, inter‐rater reliability, and internal consistency across eating disorder populations (Berg et al. 2012). The EDE was used to assess the presence and frequency of BE episodes at baseline.
2.3.2. EMA Surveys
Participants reported affect and BE and typical eating episodes on signal‐contingent and event‐contingent EMA surveys. Signal‐contingent surveys were delivered via Avicenna EMA software (formerly Ethica; Avicenna Research 2025) at 5 semi‐random times per day (one survey prompt delivered at a random time between 9 am and 11 am, 11 am and 1 pm, 1 pm and 3 pm, 3 pm and 5 pm, and 5 pm and 7 pm in the participant's time zone). Participants also completed an end‐of‐day survey at 8 pm. Participants were instructed to complete event‐contingent surveys whenever they ate a meal or snack or engaged in BE. Ratings of positive and negative affect were collected at all EMA surveys using the Short‐Form Positive and Negative Affect Schedule (PANAS) modified to also include guilt, joy, excitement, and satisfaction (Thompson 2007). These items were rated on a five‐point scale from 1 = Very slightly/Not at all to 5 = Extremely and instances where a participant rated any item more than one standard deviation above their mean level (for that item) were considered instances of high negative affect or positive affect.
2.3.3. Psychophysiological Sensors to Measure Affect
Participants wore fitbit sense 2 smartwatches on their wrist for the duration of the data collection period. Fitbit Sense 2 s contain several psychophysiological sensors, including PPG sensor, accelerometer, galvanic skin response sensor, and infrared thermopile. Sensor features extracted from the fitbit sense 2 data include average heart rate and standard deviation of heart rate, and tonic skin conductance level (average skin conductance level). The fitbit sense 2 devices could not be blinded while ensuring the data were collected as intended, so participants were able to view their data in the fitbit application on their smartphone.
2.4. Statistical Analyses
Figure 1 summarizes the data splitting, preprocessing, and model development procedures for the study. Three sets of supervised binary classification machine learning models were developed to predict BE, the first using only EMA‐reported negative affect, the second using only sensor‐measured affective arousal, and the third integrating both EMA‐ and sensor‐measured negative affect. Missing EMA outcome data were not replaced, as these surveys were used to provide the “ground truth” outcomes (i.e., BE vs. typical meals and snacks) for the developed machine learning models and we did not want to introduce simulated data into our “ground truth” sample. Missing EMA‐and sensor‐measured predictors included in our machine learning models (i.e., EMA‐measured affect, heart rate) were replaced using bagged trees imputation; this imputation approach was selected as it balances robustness with computational efficiency (Kuhn and Johnson 2019). Sensor features across the 60‐min prior to BE episodes were included in the models, based on previous research examining timeframes between 30‐ and 90‐min prior to eating episodes (Crochiere et al. 2021; Juarascio et al. 2020; Ranzenhofer et al. 2016). These time periods were divided into 5‐min moving windows, during which heart rate and electrodermal activity indices were computed. Figure 2 depicts the correlations between all EMA and sensor features in the training data. Elastic net regression was used on the full training dataset to select sensor features that predict BE episodes (Zou and Hastie 2005). Three types of machine learning models (i.e., support vector machines (Boser et al. 1992), random forest (Breiman 2001), and neural networks (Walczak and Cerpa 2003)) were developed using the first 3 weeks of data (training data) and evaluated on the last week of data (test data). Both group‐level (trained on all participants' data in aggregate) and individual‐level (trained for each individual participant on that participant's data alone) algorithms were generated (James et al. 2017). Individual and group‐level models were trained using a subset of the training data (75%, the subtraining dataset) and model parameters were tuned using 10‐fold cross validation repeated 10 times. To integrate the individual‐ and group‐level models to make predictions for an individual participant, weights from 0 to 1 by 0.05 increments (e.g., 0.00, 0.05, 0.10, etc.) were tuned using cross‐validation. For each participant, accuracy was characterized (on the held‐out 25% of training data, the validation set) using each weight for the individual model and the inverse weight (i.e., 1—weight) for the group‐level model. The weight that maximized model performance for the individual participant was retained and the weighted models for each participant were used to characterize the overall algorithm performance in the test data. A hybrid resampling approach was conducted using the ROSE (Random oversampling examples) package in R (Lunardon et al. 2021) to create balanced classes in the training and subtraining datasets. The test data set was not resampled for class balance to avoid incorporating synthetic data into our ground truth data and ensure that model performance metrics reflect the reality of class imbalance between binge eating and typical meal sand snacks. The best performing model was selected by maximizing macro‐averaged accuracy, area under the receiver operating characteristic curve (AUROC), and F1 scores. The accuracy of the best performing combined model was compared to the a priori threshold of 0.70 to evaluate model performance.
FIGURE 1.

Flow diagram of data splitting, preprocessing, and model development.
FIGURE 2.

Heat map of correlations between EMA and sensor features. All EMA features are single lagged from eating episodes (i.e., binge episodes, typical meals and snacks). All sensor features are computed for the period of time (in minutes) prior to eating episodes as designated by variable labels. Darker fill indicates higher correlation strength.
2.5. Power
Previous studies using EMA among individuals with BE have recorded an average of 2–6 BE episodes per week per participant, with most episodes occurring on separate days (Juarascio, Srivastava, et al. 2021; Juarascio, Parker, et al. 2021; Schaefer et al. 2020; Smith et al. 2020), and overall rates of missing data of 9.7%–32.6% (Schaefer et al. 2020; Smith et al. 2020; Williams‐Kerver et al. 2021). Conservatively expecting 3 binge episodes per week and 70% compliance, we anticipated approximately 189 BE episodes (~3 episodes/week * 3 weeks * 30 participants * 0.7 compliance) and 1134 typical eating episodes (~18 episodes/week * 3 weeks * 30 participants * 0.7 compliance) in the training data. Simulation studies have indicated that supervised machine learning models trained on high‐dimensional data require at least 150 observations in each class to be powered at 0.80 to detect a small‐medium effect, which 189 and 1134 exceed, respectively (Guo et al. 2010). 95% confidence intervals were constructed for each set of models (EMA‐only, sensor‐only, and combined EMA plus sensor) to evaluate their accuracy; if intervals do not overlap, this indicates a significant difference in model accuracy. For the confidence interval comparison, an a priori power analysis conducted in G*Power 3.1 (power = 0.80; alpha = 0.05) indicated that ≥ 159 total test observations were required to detect a medium effect, indicating power > 0.80 (given 63 BE episodes (~3 episodes/week * 1 week * 30 participants * 0.7 compliance) and 378 anticipated typical eating episodes (~18 episodes/week * 1 week * 30 participants * 0.7compliance) in the test data; Faul et al. 2007). The test sample, after removal of participants with inadequate data availability, included 319 total observations in the test dataset, indicating power > 0.80 for detection of a medium effect.
2.6. Deviations From Preregistration
The methods described above include a few deviations from the preregistration (Presseller et al. 2023). First, we developed a set of machine learning models using sensor data alone to contextualize the stand‐alone performance of wearable sensors for predicting naturalistic BE. Second, we initially intended to include heart rate variability features including root mean square of successive differences between normal heart beats, high frequency heart rate variability, and low frequency heart rate variability. However, Fitbit Sense 2 devices do not collect beat‐by‐beat heart rate data, inhibiting the extraction of these heart rate variability metrics. Third, prior to selecting sensor features with elastic net regression on the full dataset, we centered these features within‐person based on preliminary evidence suggesting that doing so improves individual algorithm performance (Langener and Jacobson 2025). Finally, we initially planned to select the optimal models by maximizing sensitivity while maintaining > 50% specificity. However, doing so produced models whose performance was primarily driven by their capacity to classify typical meals and snacks correctly. Given the greater importance of correctly classifying binge eating episodes for the anticipated future applications of these models, we instead selected optimal performing models based on maximizing macro‐averaged accuracy, AUROC, and F1 scores, which provide an assessment of model performance unbiased by unbalanced classes.
3. Results
Table 2 summarizes the results of the developed machine learning models for predicting BE episodes using EMA data alone, sensor data alone, and the combined EMA and sensor data model. The selected EMA‐only model was the random forest model, which demonstrated accuracy = 0.73, macro‐averaged accuracy = 0.64, AUROC = 0.69, F1 score = 0.46, sensitivity = 0.47, and specificity = 0.81 (see Figure 3). The best‐performing sensor data‐only model was the random forest model, which demonstrated accuracy = 0.49, macro‐averaged accuracy = 0.64, AUROC = 0.78, F1 score = 0.47, sensitivity = 0.92, and specificity = 0.35 (see Figure 3). The selected combined model was also the random forest model, which exhibited accuracy = 0.47, macro‐averaged accuracy = 0.61, AUROC = 0.68, F1 score = 0.45, sensitivity = 0.90, and specificity = 0.33 (see Figure 3); the combined model did not achieve the a priori accuracy threshold of 0.70. None of the developed machine learning models met both a priori thresholds of ≥ 0.70 sensitivity and ≥ 0.50 specificity simultaneously.
TABLE 2.
Confusion matrices, accuracies, sensitivities, and specificities for EMA‐only models predicting BE episodes.
| Confusion matrix | ||||||||
|---|---|---|---|---|---|---|---|---|
| Observed typical meal/snack | Observed binge episode | Accuracy (95% CI) | Macro‐averaged accuracy | AUROC | F1 score | Sensitivity | Specificity | |
| EMA Data Only | ||||||||
| Support Vector Machine | ||||||||
| Predicted typical meal/snack | 190 | 47 | 0.69 [0.64, 0.74] | 0.59 | 0.69 | 0.39 | 0.40 | 0.79 |
| Predicted binge episode | 51 | 31 | ||||||
| Random Forest Machine | ||||||||
| Predicted typical meal/snack | 195 | 41 | 0.73 [0.68, 0.78] | 0.64 | 0.69 | 0.46 | 0.47 | 0.81 |
| Predicted binge episode | 46 | 37 | ||||||
| Neural Network | ||||||||
| Predicted typical meal/snack | 187 | 50 | 0.67 [0.62, 0.73] | 0.57 | 0.65 | 0.35 | 0.36 | 0.78 |
| Predicted binge episode | 54 | 28 | ||||||
| Sensor Data Only | ||||||||
| Support Vector Machine | ||||||||
| Predicted typical meal/snack | 190 | 54 | 0.67 [0.62, 0.72] | 0.55 | 0.68 | 0.31 | 0.31 | 0.79 |
| Predicted binge episode | 51 | 24 | ||||||
| Random Forest Machine | ||||||||
| Predicted typical meal/snack | 85 | 6 | 0.49 [0.44, 0.55] | 0.64 | 0.78 | 0.47 | 0.92 | 0.35 |
| Predicted binge episode | 156 | 72 | ||||||
| Neural Network | ||||||||
| Predicted typical meal/snack | 212 | 77 | 0.67 [0.62, 0.72] | 0.45 | 0.65 | 0.06 | 0.01 | 0.88 |
| Predicted binge episode | 29 | 1 | ||||||
| EMA and Sensor Data | ||||||||
| Support Vector Machine | ||||||||
| Predicted typical meal/snack | 187 | 47 | 0.68 [0.63, 0.73] | 0.59 | 0.70 | 0.38 | 0.40 | 0.78 |
| Predicted binge episode | 54 | 31 | ||||||
| Random Forest Machine | ||||||||
| Predicted typical meal/snack | 80 | 8 | 0.47 [0.42, 0.52] | 0.61 | 0.68 | 0.45 | 0.90 | 0.33 |
| Predicted binge episode | 161 | 70 | ||||||
| Neural Network | ||||||||
| Predicted typical meal/snack | 210 | 60 | 0.71 [0.62, 0.72] | 0.55 | 0.66 | 0.02 | 0.23 | 0.87 |
| Predicted Binge episode | 31 | 18 | ||||||
FIGURE 3.

Heat maps depicting confusion matrices for best‐performing EMA‐only, sensor‐only, and combined EMA and sensor models. Heat maps depict the distribution of observations in the test data as classified by developed models. Darker fill indicates a higher proportion of observations, while lighter fill indicates a lower proportion of observations.
The EMA data only model demonstrated significantly higher overall accuracy compared to the selected sensor data only and the combined EMA and sensor data models, due to higher specificity of the EMA only model and the class imbalance in the test data that favored typical meals and snacks. When accounting for class imbalance using macro‐averaged accuracy, all three models demonstrated comparable performance. The selected sensor and combined models exhibited superior specificity (0.92 and 0.90, respectively) compared to the EMA only model (0.47), indicating that these models were better able to classify binge eating episodes correctly than the EMA only model.
4. Discussion
This study sought to understand the relative performance of EMA data, sensor data, and EMA and sensor data together for the naturalistic prediction of BE episodes in response to negative affect, with the ultimate goal of informing the development of an ecological momentary intervention targeting affect‐related BE. Findings indicate that EMA data and sensor data demonstrate comparable macro‐averaged accuracy and that sensor data do not significantly augment EMA data for the prediction of negative affect‐related BE.
Both the sensor data only and combined sensor and EMA best‐performing models exceeded the a priori sensitivity threshold of 70%, while the EMA data only model did not. Although the accuracy of the combined EMA and sensor data model fell short of our hypothesized accuracy threshold, all three selected models demonstrated macro‐averaged accuracy fairly comparable to previous research using machine learning to predict eating behaviors in the naturalistic environment (i.e., 61%–80% accuracy; Arend et al. 2023; Crochiere et al. 2021; Goldstein et al. 2018; Juarascio et al. 2020). Accuracy of the best‐performing models is likely biased by the overrepresentation of typical meals and snacks in the test data (241 out of 319 total test episodes, 75.5%), which likely explains the lower accuracy exhibited by the sensor and combined models. Further, negative affect constitutes one of the two primary mechanisms of risk for BE, alongside dietary restraint (Bray et al. 2022; Fairburn 2008). The independent, prospective contributions of dietary restraint and negative affect to BE have been substantiated by numerous longitudinal studies identifying significant associations of both dietary restraint and negative affect with BE (Allen et al. 2012; Dakanalis et al. 2014; Holmes et al. 2014; Kukk and Akkermann 2022; Sehm and Warschburger 2017; Stice 2001). Accordingly, the lower‐than‐expected overall accuracy may reflect the misclassification of instances of BE driven primarily (or exclusively) by dietary restraint, rather than negative affect, as data about dietary restraint was not included in the developed models.
The developed machine learning model for predicting negative affect‐related BE episodes using EMA and sensor data did not demonstrate significantly greater accuracy than the machine learning models using EMA data alone or sensor data alone, as hypothesized. However, the sensor data only and combined models exhibited AUROCs of 0.78 and 0.68, respectively, which were comparable to the AUROC of the EMA only model (0.69). Further, the AUROC of the sensor‐only model falls into the acceptable or moderate discrimination range (0.7–0.8), whereas the AUROC of the EMA only and combined models fall in the poor discrimination range (< 0.7; Hosmer et al. 2013). This finding offers some preliminary evidence that psychophysiological data collected by wearable sensors may outperform EMA data for the differentiation of BE and typical meals and snacks using negative affect data in the naturalistic setting. Existing research has identified reliable changes in heart rate and electrodermal activity during times of self‐reported elevated negative affect among individuals with binge eating (French and Chen 2021; Messerli‐Burgy et al. 2010; Yiu et al. 2018), indicating that these psychophysiological data are likely valid indicators of negative affectivity. These data may augment self‐report of negative affectivity for the prediction of BE as individuals with BE consistently demonstrate deficits in emotional awareness (Brockmeyer et al. 2014; Kittel et al. 2015; Ruscitti et al. 2016); psychophysiological data may overcome limitations to self‐report of negative affect resulting from deficits in emotional awareness, as these data do not rely on patient insight.
None of the EMA data only models exhibited sensitivity greater than 0.50, indicating that self‐reported affect data demonstrates poor capacity to predict negative affect‐related BE episodes accurately. Conversely, the random forest machine learning model using sensor data alone and the random forest model integrating both EMA data and sensor data demonstrated significantly higher (p < 0.05) sensitivity (i.e., the capacity to correctly classify BE episodes as BE; sensor data random forest model = 0.92, EMA plus sensor data random forest model = 0.90) than chance. Importantly, the improved sensitivity came at the cost of specificity, as these models exhibited much poorer specificity (0.35 and 0.33, respectively) and accordingly, lower overall accuracy (0.49 and 0.47, respectively). Nonetheless, this pattern of findings may indicate that psychophysiological features including mean heart rate, standard deviation of heart rate, and mean electrodermal activity may facilitate correct prediction of negative affect‐related BE episodes specifically.
The goal of this research is to inform the development of a momentary intervention system that can predict BE in response to negative affect. Use of sensor data alone could significantly reduce user burden, as smartwatches are generally well tolerated and require no active engagement from the user once the smartwatch has been put on. Conversely, EMA requires repeated engagement (typically 4–6 times per day) and can be experienced as burdensome by research participants (van Genugten et al. 2020; Williams et al. 2021). Self‐reported burden has been associated prospectively with decreased EMA adherence (Tate et al. 2024) and EMA adherence typically declines across extended data collection protocols (such as those collected across a course of treatment; Kim et al. 2025; Ono et al. 2019). Delivery of momentary interventions that are prompted by low‐burden sensor data, rather than reliant on EMA, could improve user experience and, subsequently, engagement and response. However, for this momentary intervention system, correct prediction of BE risk in real‐time is more important to achieve than correct prediction of typical meals or snacks, as it is better for patients to practice therapeutic skills more often than needed than to miss opportunities to use therapeutic skills when at risk for BE. Therefore, the sensor data only model, which exhibited high sensitivity, is likely well‐suited for the future development of an ecological momentary intervention system, despite not outperforming the EMA model in terms of overall accuracy.
Taken together, findings from the present study indicate that psychophysiological data about negative affect states collected by a commercial smartwatch device demonstrate comparable predictive performance as EMA affect data for the prediction of real‐time BE episodes in the naturalistic environment among a diverse sample of adults with transdiagnostic BE. Findings from this study add to the extensive literature base establishing negative affect as a critical momentary risk factor for BE (Walenda et al. 2021). The comparable performance of psychophysiological sensor data in the prediction of BE episodes compared to EMA provides additional evidence that psychophysiological metrics measuring affective arousal (i.e., heart rate and electrodermal activity) capture divergent patterns in the 60 min prior to BE and typical meals and snacks, although the present study does not provide insight into the specific nature of these patterns.
Findings lay the groundwork for future development of a low‐burden just‐in‐time adaptive intervention targeting BE that arises in response to negative affect and which integrates psychophysiological sensor data to inform intervention delivery. Before development of this tool, efforts should be made to further improve predictive accuracy of the sensor data‐only model, which might be achieved by inclusion of additional sensors, inclusion of additional sensor features, and/or increasing overlap of the moving windows in which features were computed. This just‐in‐time adaptive intervention system may be beneficial for prompting effective emotion regulation and urge management skill use when individuals with binge‐eating disorder and bulimia nervosa are at imminent risk for BE, thereby reducing instances of BE and promoting recovery. One previous study that identified that just‐in‐time adaptive interventions significantly improved treatment outcomes for individuals with BE and elevated difficulties in emotion regulation and impulsivity over enhanced cognitive‐behavioral therapy alone (Presseller et al. 2022), suggesting that a sensor‐integrated just‐in‐time adaptive intervention system targeting BE risk following negative affectivity may be particularly helpful for these individuals. End‐user involvement in design of this digital tool will be important to ensure that feasibility and usability is realized and that the tool demonstrates uptake in real‐world clinical settings.
4.1. Study Strengths and Limitations
The study was strengthened by recruitment of a diverse population in terms of sex, race, and ethnicity compared to many samples of individuals with BE, increasing generalizability of the pattern of findings to the population of individuals with binge‐spectrum eating disorders. Because person‐specific models were developed and the training and test data were split within (rather than between) subjects, it is not expected that the specific developed models would perform comparably in new participants. Although our sample was demographically diverse, the majority of our sample was comprised of individuals with bulimia nervosa‐spectrum eating disorders. Future work should evaluate how well these findings generalize to individuals with binge‐eating disorder. Further, participants were not blinded to their smartwatch data and the sample subsequently may reflect participants who are interested in health monitoring. EMA adherence was relatively low in this study compared to previous studies demonstrating EMA adherence among adults with BE of 76%–86% (Berg et al. 2015, 2013; Schaefer et al. 2020; Smyth et al. 2007). Lower EMA adherence may have been related to the study's extended monitoring period (4 weeks, compared to 1–2 weeks in most EMA studies) and the relatively modest compensation participants received ($15/week for completion of EMA and wearing the smartwatches). The extended (4 week) duration of the study was selected to ensure that at least 150 observations from each class (binge eating and typical meals and snacks) were collected during the training data period. Despite low adherence, we collected more than 150 observations in the training data in each group, indicating full power for the development of the group‐level machine learning models; however, poor EMA adherence may have impacted the number of observations of BE for individual participants, which limits confidence in the individual models developed for these participants. All of the developed models included only negative affect‐related predictors (i.e., self‐report of current negative affect levels, psychophysiological features of affective arousal), as sensors included in the Fitbit Sense 2 devices primarily capture autonomic arousal. Findings from this study therefore are likely not to generalize to other uses of EMA to measure binge eating risk factors (e.g., dietary restraint), which these types of sensors are likely not well‐equipped to measure. Inclusion of gold standard metrics of heart rate variability (e.g., root mean square of successive differences, high frequency heart rate variability, and low frequency heart rate variability) was not possible given the format of heart rate data output by the Fitbit Sense 2 model. It is likely that inclusion of these metrics could improve the accuracy of the developed models. Similarly, only tonic (average) skin conductance level data were available from the galvanic skin response sensor within the Fitbit Sense 2; addition of skin conductance response frequency may also have improved the accuracy of our models. Finally, it is possible that use of EMA to monitor eating behavior and affect over the study period increased participants' awareness of patterns maintaining their binge eating, thereby impacting the factors maintaining binge eating by the end of the study period.
4.2. Future Directions
Future research that follows from the present study should aim to continue understanding the psychophysiological arousal that is associated with negative affect among individuals with BE and its momentary impact on risk for BE. For instance, future work should more granularly examine the nature of psychophysiological changes associated with negative affect states compared to positive affect states among individuals with BE (e.g., direction and magnitude of changes in heart rate and electrodermal activity, duration of changes, etc.) to better understand the impact of affective valence on psychophysiological arousal in this population. Discrete and common components with simultaneous component analysis (DISCO‐SCA) should be applied to examine the common and discrete information about affect that is collected via EMA and wearable sensors to inform the interpretation of these results. Similar research should be conducted to understand how psychophysiological metrics differ in the 60 min prior to BE episodes compared to typical meals and snacks, to elucidate psychophysiological signatures of BE risk and identify mechanisms by which physiological arousal contributes to BE maintenance.
Additional research should be conducted to improve machine learning model performance. Specifically, researchers should examine whether inclusion of heart rate variability metrics and other sensor data (e.g., accelerometry data, sleep data, location data) alongside heart rate and electrodermal activity further improves model accuracy for predicting BE episodes, given several studies linking these data to BE (da Luz et al. 2023; Dougherty et al. 2024; Trace et al. 2012) and other types of dysregulated eating behavior (Crochiere et al. 2021; Elliston et al. 2020). Finally, ensemble learning approaches may exhibit better performance than single models alone and should be explored for the prediction of BE.
Findings from the present study set the stage for future development and evaluation of a sensor‐integrated just‐in‐time adaptive intervention for BE. Development of this tool should be conducted with input from end users (i.e., patients with BE and eating disorder clinicians), especially regarding design. The capacity for the sensor‐integrated just‐in‐time adaptive intervention for BE to improve clinical outcomes among individuals with BE when used alongside enhanced cognitive‐behavioral therapy should be evaluated with a randomized controlled trial. Finally, the feasibility and acceptability of this intervention system should be evaluated, including with particular attention to the impact that smartwatch use might have on eating disorder symptom severity, given concerns raised about the possibility that these devices might worsen eating disorder symptoms (Moody et al. 2025).
Public significance statements are as follows:
Individuals with bulimia nervosa or binge eating disorder often binge eat when experiencing negative emotions.
This study used surveys and smartwatch data measuring negative emotions to predict binge eating episodes in daily life.
Smartwatch data were able to predict binge eating episodes with similar accuracy to survey data, offering a low‐burden opportunity to detect binge eating risk in daily life.
Author Contributions
Emily K. Presseller: conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, visualization, writing – original draft, writing – review and editing. Philip A. Gable: conceptualization, funding acquisition, methodology, writing – review and editing. Fengqing Zhang: conceptualization, funding acquisition, writing – review and editing, and formal analysis. Stephanie M. Manasse: investigation, project administration, supervision, writing – review and editing. Adrienne S. Juarascio: conceptualization, funding acquisition, investigation, methodology, project administration, resources, and supervision.
Funding
Research reported in this publication was supported by the National Institute of Mental Health under award number F31MH131262. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Conflicts of Interest
The authors declare no conflicts of interest.
Presseller, E. K. , Gable P. A., Zhang F., Manasse S. M., and Juarascio A. S.. 2026. “Detecting Binge Eating Risk With Naturalistic Data and Machine Learning: A Comparative Observational Study.” International Journal of Eating Disorders 59, no. 8: 1800–1813. 10.1002/eat.70107.
Action Editor: Jake Linardon
Data Availability Statement
Data, materials, and code are available upon reasonable request to the corresponding author.
References
- Allen, K. L. , Byrne S. M., and McLean N. J.. 2012. “The Dual‐Pathway and Cognitive‐Behavioural Models of Binge Eating: Prospective Evaluation and Comparison.” European Child & Adolescent Psychiatry 21, no. 1: 51–62. 10.1007/s00787-011-0231-z. [DOI] [PubMed] [Google Scholar]
- American Heart Association . 2015. “All About Heart Rate (Pulse).” https://www.heart.org/en/health‐topics/high‐blood‐pressure/the‐facts‐about‐high‐blood‐pressure/all‐about‐heart‐rate‐pulse.
- American Psychiatric Association . 2013. Diagnostic and Statistical Manual of Mental Disorders (DSM‐5). 5th ed. American Psychiatric Association. [Google Scholar]
- Arend, A. K. , Kaiser T., Pannicke B., et al. 2023. “Toward Individualized Prediction of Binge‐Eating Episodes Based on Ecological Momentary Assessment Data: Item Development and Pilot Study in Patients With Bulimia Nervosa and Binge‐Eating Disorder.” JMIR Medical Informatics 11: e41513. 10.2196/41513. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Avicenna Research . 2025. “Avicenna. In.”
- Berg, K. C. , Cao L., Crosby R. D., et al. 2017. “Negative Affect and Binge Eating: Reconciling Differences Between Two Analytic Approaches in Ecological Momentary Assessment Research.” International Journal of Eating Disorders 50, no. 10: 1222–1230. 10.1002/eat.22770. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Berg, K. C. , Crosby R. D., Cao L., et al. 2013. “Facets of Negative Affect Prior to and Following Binge‐Only, Purge‐Only, and Binge/Purge Events in Women With Bulimia Nervosa.” Journal of Abnormal Psychology 122, no. 1: 111–118. 10.1037/a0029703. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Berg, K. C. , Crosby R. D., Cao L., et al. 2015. “Negative Affect Prior to and Following Overeating‐Only, Loss of Control Eating‐Only, and Binge Eating Episodes in Obese Adults.” International Journal of Eating Disorders 48, no. 6: 641–653. 10.1002/eat.22401. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Berg, K. C. , Peterson C. B., Frazier P., and Crow S. J.. 2012. “Psychometric Evaluation of the Eating Disorder Examination and Eating Disorder Examination‐Questionnaire: A Systematic Review of the Literature.” International Journal of Eating Disorders 45, no. 3: 428–438. 10.1002/eat.20931. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Blackledge, J. T. , and Hayes S. C.. 2001. “Emotion Regulation in Acceptance and Commitment Therapy.” Journal of Clinical Psychology 57, no. 2: 243–255. 10.1002/1097-4679(200102)57:2<243::aid-jclp9>3.0.co;2-x. [DOI] [PubMed] [Google Scholar]
- Boser, B. E. , Guyon I. M., and Vapnik V. N.. 1992. “A Training Algorithm for Optimal Margin Classifiers. Fifth Annual Workshop on Computational Learning Theory.”
- Bradley, M. M. , Greenwald M. K., Petry M. C., and Lang P. J.. 1992. “Remembering Pictures: Pleasure and Arousal in Memory.” Journal of Experimental Psychology. Learning, Memory, and Cognition 18, no. 2: 379–390. 10.1037//0278-7393.18.2.379. [DOI] [PubMed] [Google Scholar]
- Bray, B. , Sadowski A., Bray C., Bradley R., and Zwickey H.. 2022. “Clinical Aspects of Binge Eating Disorder: A Cross‐Sectional Mixed‐Methods Study of Binge Eating Disorder Experts' Perspectives.” Frontiers in Psychiatry 13: 1087165. 10.3389/fpsyt.2022.1087165. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Breiman, L. 2001. “Random Forests.” Machine Learning 45: 5–32. 10.1023/A:1010933404324. [DOI] [Google Scholar]
- Brockmeyer, T. , Skunde M., Wu M., et al. 2014. “Difficulties in Emotion Regulation Across the Spectrum of Eating Disorders.” Comprehensive Psychiatry 55, no. 3: 565–571. [DOI] [PubMed] [Google Scholar]
- Cools, J. , Schotte D. E., and McNally R. J.. 1992. “Emotional Arousal and Overeating in Restrained Eaters.” Journal of Abnormal Psychology 101, no. 2: 348–351. 10.1037//0021-843x.101.2.348. [DOI] [PubMed] [Google Scholar]
- Crochiere, R. J. , Zhang F. Z., Juarascio A. S., Goldstein S. P., Thomas J. G., and Forman E. M.. 2021. “Comparing Ecological Momentary Assessment to Sensor‐Based Approaches in Predicting Dietary Lapse.” Translational Behavioral Medicine 11, no. 12: 2099–2109. 10.1093/tbm/ibab123. [DOI] [PubMed] [Google Scholar]
- Crosby, R. D. , Wonderlich S. A., Engel S. G., Simonich H., Smyth J., and Mitchell J. E.. 2009. “Daily Mood Patterns and Bulimic Behaviors in the Natural Environment.” Behaviour Research and Therapy 47, no. 3: 181–188. 10.1016/j.brat.2008.11.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- da Luz, F. Q. , Sainsbury A., Salis Z., et al. 2023. “A Systematic Review With Meta‐Analyses of the Relationship Between Recurrent Binge Eating and Sleep Parameters.” International Journal of Obesity (London, England) 47, no. 3: 145–164. 10.1038/s41366-022-01250-9. [DOI] [PubMed] [Google Scholar]
- Dakanalis, A. , Timko C. A., Carra G., et al. 2014. “Testing the Original and the Extended Dual‐Pathway Model of Lack of Control Over Eating in Adolescent Girls. A Two‐Year Longitudinal Study.” Appetite 82: 180–193. 10.1016/j.appet.2014.07.022. [DOI] [PubMed] [Google Scholar]
- Deloitte Access Economics . 2020. “Social and Economic Cost of Eating Disorders in the US: A Report for the Strategic Training Initiative for the Prevention of Eating Disorders and the Academy for Eating Disorders.” https://www.hsph.harvard.edu/striped/report‐economic‐costs‐of‐eating‐disorders/.
- Dissanayake, T. , Rajapaksha Y., Ragel R., and Nawinne I.. 2019. “An Ensemble Learning Approach for Electrocardiogram Sensor Based Human Emotion Recognition.” Sensors (Basel) 19, no. 20: 4495. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Dougherty, E. N. , Bottera A. R., and Haedt‐Matt A. A.. 2024. “Dysphoric Mood May Explain the Relation Between Sleep Reactivity and Binge Eating but Not Purging.” Journal of American College Health 73: 1–5. 10.1080/07448481.2024.2336998. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Elliston, K. G. , Schuz B., Albion T., and Ferguson S. G.. 2020. “Comparison of Geographic Information System and Subjective Assessments of Momentary Food Environments as Predictors of Food Intake: An Ecological Momentary Assessment Study.” JMIR mHealth and uHealth 8, no. 7: e15948. 10.2196/15948. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Engel, S. G. , Wonderlich S. A., Crosby R. D., et al. 2013. “The Role of Affect in the Maintenance of Anorexia Nervosa: Evidence From a Naturalistic Assessment of Momentary Behaviors and Emotion.” Journal of Abnormal Psychology 122, no. 3: 709–719. 10.1037/a0034010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Engelhard, M. M. , Navar A. M., and Pencina M. J.. 2021. “Incremental Benefits of Machine Learning‐When Do we Need a Better Mousetrap?” JAMA Cardiology 6, no. 6: 621–623. 10.1001/jamacardio.2021.0139. [DOI] [PubMed] [Google Scholar]
- Fairburn, C. G. 2008. Cognitive Behavior Therapy and Eating Disorders. 1st ed. Guilford Press. [Google Scholar]
- Fairburn, C. G. , Cooper Z., and O'Connor M.. 2014. “Eating Disorder Examination (Edition 17.0D).”
- Faul, F. , Erdfelder E., Lang A. G., and Buchner A.. 2007. “G*Power 3: A Flexible Statistical Power Analysis Program for the Social, Behavioral, and Biomedical Sciences.” Behavior Research Methods 39, no. 2: 175–191. 10.3758/bf03193146. [DOI] [PubMed] [Google Scholar]
- Frazier, T. W. , Strauss M. E., and Steinhauer S. R.. 2004. “Respiratory Sinus Arrhythmia as an Index of Emotional Response in Young Adults.” Psychophysiology 41, no. 1: 75–83. 10.1046/j.1469-8986.2003.00131.x. [DOI] [PubMed] [Google Scholar]
- French, M. N. , and Chen E. Y.. 2021. “Emotion and Psychophysiological Responses During Emotion‐Eliciting Film Clips in an Eating Disorders Sample.” Frontiers in Psychology 12: 630426. 10.3389/fpsyg.2021.630426. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Goldstein, S. P. , Zhang F., Thomas J. G., Butryn M. L., Herbert J. D., and Forman E. M.. 2018. “Application of Machine Learning to Predict Dietary Lapses During Weight Loss.” Journal of Diabetes Science and Technology 12, no. 5: 1045–1052. 10.1177/1932296818775757. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Golland, Y. , Keissar K., and Levit‐Binnun N.. 2014. “Studying the Dynamics of Autonomic Activity During Emotional Experience.” Psychophysiology 51, no. 11: 1101–1111. 10.1111/psyp.12261. [DOI] [PubMed] [Google Scholar]
- Gouizi, K. , Bereksi Reguig F., and Maaoui C.. 2011. “Emotion Recognition From Physiological Signals.” Journal of Medical Engineering & Technology 35, no. 6–7: 300–307. 10.3109/03091902.2011.601784. [DOI] [PubMed] [Google Scholar]
- Guo, Y. , Graber A., McBurney R. N., and Balasubramanian R.. 2010. “Sample Size and Statistical Power Considerations in High‐Dimensionality Data Settings: A Comparative Study of Classification Algorithms.” BMC Bioinformatics 11: 447. 10.1186/1471-2105-11-447. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hasnul, M. A. , Aziz N. A. A., Alelyani S., Mohana M., and Aziz A. A.. 2021. “Electrocardiogram‐Based Emotion Recognition Systems and Their Applications in Healthcare‐A Review.” Sensors (Basel) 21, no. 15: 5015. 10.3390/s21155015. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Hilbert, A. , Rief W., Tuschen‐Caffier B., de Zwaan M., and Czaja J.. 2009. “Loss of Control Eating and Psychological Maintenance in Children: An Ecological Momentary Assessment Study.” Behaviour Research and Therapy 47, no. 1: 26–33. 10.1016/j.brat.2008.10.003. [DOI] [PubMed] [Google Scholar]
- Holmes, M. , Fuller‐Tyszkiewicz M., Skouteris H., and Broadbent J.. 2014. “Tests of an Extension of the Dual Pathway Model of Bulimic Symptoms to the State‐Based Level.” Eating Behaviors 15, no. 2: 280–285. 10.1016/j.eatbeh.2014.03.011. [DOI] [PubMed] [Google Scholar]
- Hosmer, D. W. , Lemeshow S., and Sturdivant R. X.. 2013. “Assessing the Fit of the Model.” In Applied Logistic Regression, 3rd Edition ed. Wiley. [Google Scholar]
- Hudson, J. I. , Hiripi E., Pope H. G. Jr., and Kessler R. C.. 2007. “The Prevalence and Correlates of Eating Disorders in the National Comorbidity Survey Replication.” Biological Psychiatry 61, no. 3: 348–358. 10.1016/j.biopsych.2006.03.040. [DOI] [PMC free article] [PubMed] [Google Scholar]
- James, G. , Witten D., Hastie T., and Tibshirani R.. 2017. An Introduction to Statistical Learning With Applications in R. Springer. 10.1007/978-1-4614-7138-7. [DOI] [Google Scholar]
- Jang, E. H. , Byun S., Park M. S., and Sohn J. H.. 2019. “Reliability of Physiological Responses Induced by Basic Emotions: A Pilot Study.” Journal of Physiological Anthropology 38, no. 1: 15. 10.1186/s40101-019-0209-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Jang, E. H. , Park B. J., Park M. S., Kim S. H., and Sohn J. H.. 2015. “Analysis of Physiological Signals for Recognition of Boredom, Pain, and Surprise Emotions.” Journal of Physiological Anthropology 34: 25. 10.1186/s40101-015-0063-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Juarascio, A. , Srivastava P., Presseller E., Clark K., Manasse S., and Forman E.. 2021. “A Clinician‐Controlled Just‐In‐Time Adaptive Intervention System (CBT+) Designed to Promote Acquisition and Utilization of Cognitive Behavioral Therapy Skills in Bulimia Nervosa: Development and Preliminary Evaluation Study.” JMIR Form Res 5, no. 5: e18261. 10.2196/18261. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Juarascio, A. S. , Crochiere R. J., Tapera T. M., Palermo M., and Zhang F.. 2020. “Momentary Changes in Heart Rate Variability Can Detect Risk for Emotional Eating Episodes.” Appetite 152: 104698. 10.1016/j.appet.2020.104698. [DOI] [PubMed] [Google Scholar]
- Juarascio, A. S. , Parker M. N., Hunt R., Murray H. B., Presseller E. K., and Manasse S. M.. 2021. “Mindfulness and Acceptance‐Based Behavioral Treatment for Bulimia‐Spectrum Disorders: A Pilot Feasibility Randomized Trial.” International Journal of Eating Disorders 54, no. 7: 1270–1277. 10.1002/eat.23512. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Juarascio, A. S. , Presseller E. K., Srivastava P., Manasse S. M., and Forman E. M.. 2022. “A Randomized Controlled Trial of CBT+: A Clinician‐Controlled, Just‐In‐Time, Adjunctive Intervention for Bulimia‐Spectrum Disorders.” Behavior Modification 47: 1454455221109434. 10.1177/01454455221109434. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Juarascio, A. S. , Srivastava P., Presseller E. K., et al. 2022. “Using Continuous Glucose Monitoring to Detect and Intervene on Dietary Restriction in Individuals With Binge Eating: The SenseSupport Withdrawal Design Study.” JMIR Formative Research 6: e38479. 10.2196/38479. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kanjo, E. , Al‐Husain L., and Chamberlain A.. 2015. “Emotions in Context: Examining Pervasive Affective Sensing Systems, Applications, and Analyses.” Personal and Ubiquitous Computing 19: 1197–1212. [Google Scholar]
- Kessler, R. C. , Berglund P. A., Chiu W. T., et al. 2013. “The Prevalence and Correlates of Binge Eating Disorder in the World Health Organization World Mental Health Surveys.” Biological Psychiatry 73, no. 9: 904–914. 10.1016/j.biopsych.2012.11.020. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kim, H. , Kwon S., Park S., Kang C., and Kim H.. 2025. “Feasibility and Adherence to Ecological Momentary Assessment Among Community‐Dwelling Adults With Suicide Risk.” BMC Nursing 24, no. 1: 807. 10.1186/s12912-025-03432-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kittel, R. , Braudhardt A., and Hilbert A.. 2015. “Cognitive and Emotional Functioning in Binge‐Eating Disorder: A Systematic Review.” International Journal of Eating Disorders 48: 535–554. [DOI] [PubMed] [Google Scholar]
- Kockler, T. D. , Santangelo P. S., and Ebner‐Priemer U. W.. 2018. “Investigating Binge Eating Using Ecological Momentary Assessment: The Importance of an Appropriate Sampling Frequency.” Nutrients 10, no. 1: 105. 10.3390/nu10010105. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Kreibig, S. D. 2010. “Autonomic Nervous System Activity in Emotion: A Review.” Biological Psychology 84, no. 3: 394–421. 10.1016/j.biopsycho.2010.03.010. [DOI] [PubMed] [Google Scholar]
- Kuhn, M. , and Johnson K.. 2019. Feature Engineering and Selection: A Practical Approach for Predictive Models. Chapman and Hall/CRC. [Google Scholar]
- Kukk, K. , and Akkermann K.. 2022. “The Interplay Between Binge Eating Risk Factors: Toward an Integrated Model.” Journal of Health Psychology 27, no. 1: 92–102. 10.1177/1359105320947806. [DOI] [PubMed] [Google Scholar]
- Langener, A. M. , and Jacobson A. M.. 2025. “Just in Time or Just a Guess?” In Addressing Challenges in Validating Prediction Models Based on Longitudinal Data. OSF Preprints. [Google Scholar]
- Lavender, J. M. , Utzinger L. M., Cao L., et al. 2016. “Reciprocal Associations Between Negative Affect, Binge Eating, and Purging in the Natural Environment in Women With Bulimia Nervosa.” Journal of Abnormal Psychology 125, no. 3: 381–386. 10.1037/abn0000135. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Lavender, J. M. , Wonderlich S. A., Engel S. G., Gordon K. H., Kaye W. H., and Mitchell J. E.. 2015. “Dimensions of Emotion Dysregulation in Anorexia Nervosa and Bulimia Nervosa: A Conceptual Review of the Empirical Literature.” Clinical Psychology Review 40: 111–122. 10.1016/j.cpr.2015.05.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Linardon, J. 2018. “Meta‐Analysis of the Effects of Cognitive‐Behavioral Therapy on the Core Eating Disorder Maintaining Mechanisms: Implications for Mechanisms of Therapeutic Change.” Cognitive Behaviour Therapy 47, no. 2: 107–125. 10.1080/16506073.2018.1427785. [DOI] [PubMed] [Google Scholar]
- Lunardon, A. , Menardi G., and Torelli N.. 2021. “Random Over‐Sampling Examples.” In CRAN. https://cran.r‐project.org/web/packages/ROSE/ROSE.pdf.
- Marchewka, A. , Zurawski L., Jednorog K., and Grabowska A.. 2014. “The Nencki Affective Picture System (NAPS): Introduction to a Novel, Standardized, Wide‐Range, High‐Quality, Realistic Picture Database.” Behavior Research Methods 46, no. 2: 596–610. 10.3758/s13428-013-0379-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- McMain, S. , Korman L. M., and Dimeff L.. 2001. “Dialectical Behavior Therapy and the Treatment of Emotion Dysregulation.” Journal of Clinical Psychology 57, no. 2: 183–196. 10.1002/1097-4679(200102)57:2<183::aid-jclp5>3.0.co;2-y. [DOI] [PubMed] [Google Scholar]
- Messerli‐Burgy, N. , Engesser C., Lemmenmeier E., Steptoe A., and Laederach‐Hofmann K.. 2010. “Cardiovascular Stress Reactivity and Recovery in Bulimia Nervosa and Binge Eating Disorder.” International Journal of Psychophysiology 78, no. 2: 163–168. 10.1016/j.ijpsycho.2010.07.005. [DOI] [PubMed] [Google Scholar]
- Moody, S. , Ross L., Opitz M.‐C., et al. 2025. “Associations Between the Use of Fitness and Diet Tracking Technology and Disordered Eating Behaviour: A Systematic Review.” European Eating Disorders Review 33, no. 6: 1288–1313. 10.1002/erv.70006. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ono, M. , Schneider S., Junghaenel D. U., and Stone A. A.. 2019. “What Affects the Completion of Ecological Momentary Assessments in Chronic Pain Research? An Individual Patient Data Meta‐Analysis.” Journal of Medical Internet Research 21, no. 2: e11398. 10.2196/11398. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Peat, C. M. , Berkman N. D., Lohr K. N., et al. 2017. “Comparative Effectiveness of Treatments for Binge‐Eating Disorder: Systematic Review and Network Meta‐Analysis.” European Eating Disorders Review 25, no. 5: 317–328. 10.1002/erv.2517. [DOI] [PubMed] [Google Scholar]
- Presseller, E. K. , Lampe E. W., Zhang F., et al. 2023. “Using Wearable Passive Sensing to Predict Binge Eating in Response to Negative Affect Among Individuals With Transdiagnostic Binge Eating: Protocol for an Observational Study.” JMIR Res Protoc 12: e47098. 10.2196/47098. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Presseller, E. K. , Wilkinson M. L., Trainor C., Lampe E. W., and Juarascio A. S.. 2022. “Self‐Regulation Deficits Moderate Treatment Outcomes in a Clinical Trial Evaluating Just‐In‐Time Adaptive Interventions as an Augmentation to Cognitive‐Behavioral Therapy for Bulimia‐Spectrum Eating Disorders.” International Journal of Eating Disorders 55, no. 5: 709–716. 10.1002/eat.23695. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ranzenhofer, L. M. , Engel S. G., Crosby R. D., et al. 2014. “Using Ecological Momentary Assessment to Examine Interpersonal and Affective Predictors of Loss of Control Eating in Adolescent Girls.” International Journal of Eating Disorders 47, no. 7: 748–757. 10.1002/eat.22333. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ranzenhofer, L. M. , Engel S. G., Crosby R. D., et al. 2016. “Real‐Time Assessment of Heart Rate Variability and Loss of Control Eating in Adolescent Girls: A Pilot Study.” International Journal of Eating Disorders 49, no. 2: 197–201. 10.1002/eat.22464. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Ranzenhofer, L. M. , Solhjoo S., Lloyd E. C., et al. 2021. Shifted Autonomic Indices Prior to Loss‐Of‐Control Eating: A Possible Biomarker Eating Disorders. Research Society Annual Meeting, Virtual. [Google Scholar]
- Ruscitti, C. , Rufino K., Goodwin N., and Wagner R.. 2016. “Difficulties in Emotion Regulation in Patients With Eating Disorders.” Borderline Personality Disorder and Emotion Dysregulation 3, no. 3: 3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Schaefer, L. M. , Smith K. E., Anderson L. M., et al. 2020. “The Role of Affect in the Maintenance of Binge‐Eating Disorder: Evidence From an Ecological Momentary Assessment Study.” Journal of Abnormal Psychology 129, no. 4: 387–396. 10.1037/abn0000517. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Sehm, M. , and Warschburger P.. 2017. “The Dual‐Pathway Model of Binge Eating: Is There a Need for Modification?” Appetite 114: 137–145. 10.1016/j.appet.2017.03.028. [DOI] [PubMed] [Google Scholar]
- Shi, Q. , Thornton L. M., Kilshaw R., et al. 2024. “Relationship Between Intensive Passive Data Signals and Patterns of Binge‐Eating Behaviors: From a Dynamical‐System Approach.” Clinical Psychological Science 13, no. 3: 558–581. 10.1177/21677026241280728. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Smith, K. E. , Mason T. B., Schaefer L. M., et al. 2020. “Examining Intra‐Individual Variability in Food‐Related Inhibitory Control and Negative Affect as Predictors of Binge Eating Using Ecological Momentary Assessment.” Journal of Psychiatric Research 120: 137–143. 10.1016/j.jpsychires.2019.10.017. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Smyth, J. M. , Wonderlich S. A., Heron K. E., et al. 2007. “Daily and Momentary Mood and Stress Are Associated With Binge Eating and Vomiting in Bulimia Nervosa Patients in the Natural Environment.” Journal of Consulting and Clinical Psychology 75, no. 4: 629–638. 10.1037/0022-006X.75.4.629. [DOI] [PubMed] [Google Scholar]
- Smyth, J. M. , Wonderlich S. A., Sliwinski M. J., et al. 2009. “Ecological Momentary Assessment of Affect, Stress, and Binge‐Purge Behaviors: Day of Week and Time of Day Effects in the Natural Environment.” International Journal of Eating Disorders 42, no. 5: 429–436. 10.1002/eat.20623. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Stice, E. 2001. “A Prospective Test of the Dual‐Pathway Model of Bulimic Pathology: Mediating Effects of Dieting and Negative Affect.” Journal of Abnormal Psychology 110, no. 1: 124–135. 10.1037//0021-843x.110.1.124. [DOI] [PubMed] [Google Scholar]
- Svaldi, J. , Griepenstroh J., Tuschen‐Caffier B., and Ehring T.. 2012. “Emotion Regulation Deficits in Eating Disorders: A Marker of Eating Pathology or General Psychopathology?” Psychiatry Research 197, no. 1–2: 103–111. 10.1016/j.psychres.2011.11.009. [DOI] [PubMed] [Google Scholar]
- Tate, A. D. , Fertig A. R., de Brito J. N., et al. 2024. “Momentary Factors and Study Characteristics Associated With Participant Burden and Protocol Adherence: Ecological Momentary Assessment.” JMIR Form Res 8: e49512. 10.2196/49512. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Thompson, E. R. 2007. “Development and Validation of an Internally Reliable Short‐Form of the Positive and Negative Affect Schedule (PANAS).” Journal of Cross‐Cultural Psychology 38, no. 2: 227–242. [Google Scholar]
- Trace, S. E. , Thornton L. M., Runfola C. D., Lichtenstein P., Pedersen N. L., and Bulik C. M.. 2012. “Sleep Problems Are Associated With Binge Eating in Women.” International Journal of Eating Disorders 45, no. 5: 695–703. 10.1002/eat.22003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Valderas, M. T. , Bolea J., Laguna P., Bailon R., and Vallverdu M.. 2019. “Mutual Information Between Heart Rate Variability and Respiration for Emotion Characterization.” Physiological Measurement 40, no. 8: 084001. [DOI] [PubMed] [Google Scholar]
- van Genugten, C. R. , Schuurmans J., Lamers F., et al. 2020. “Experienced Burden of and Adherence to Smartphone‐Based Ecological Momentary Assessment in Persons With Affective Disorders.” Journal of Clinical Medicine 9, no. 2: 322. 10.3390/jcm9020322. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Walczak, S. , and Cerpa N.. 2003. “Artificial Neural Networks.” In Encyclopedia of Physical Science and Technology, edited by Meyers R. A., 3rd Edition ed. Elsevier Science. [Google Scholar]
- Walenda, A. , Bogusz K., Kopera M., Jakubczyk A., Wojnar M., and Kucharska K.. 2021. “Emotion Regulation in Binge Eating Disorder.” Psychiatria Polska 55, no. 6: 1433–1448. 10.12740/PP/OnlineFirst/122212. [DOI] [PubMed] [Google Scholar]
- Weinbach, N. , Sher H., and Bohon C.. 2018. “Differences in Emotion Regulation Difficulties Across Types of Eating Disorders During Adolescence.” Journal of Abnormal Child Psychology 46, no. 6: 1351–1358. 10.1007/s10802-017-0365-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Williams, M. T. , Lewthwaite H., Fraysse F., Gajewska A., Ignatavicius J., and Ferrar K.. 2021. “Compliance With Mobile Ecological Momentary Assessment of Self‐Reported Health‐Related Behaviors and Psychological Constructs in Adults: Systematic Review and Meta‐Analysis.” Journal of Medical Internet Research 23, no. 3: e17023. 10.2196/17023. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Williams‐Kerver, G. A. , Schaefer L. M., Hazzard V. M., et al. 2021. “Baseline and Momentary Predictors of Ecological Momentary Assessment Adherence in a Sample of Adults With Binge‐Eating Disorder.” Eating Behaviors 41: 101509. 10.1016/j.eatbeh.2021.101509. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Yiu, A. , Christensen K., Arlt J. M., and Chen E. Y.. 2018. “Distress Tolerance Across Self‐Report, Behavioral and Psychophysiological Domains in Women With Eating Disorders, and Healthy Controls.” Journal of Behavior Therapy and Experimental Psychiatry 61: 24–31. 10.1016/j.jbtep.2018.05.006. [DOI] [PubMed] [Google Scholar]
- Zou, H. , and Hastie T.. 2005. “Regularization and Variable Selection via the Elastic Net.” Journal of the Royal Statistical Society 67: 301–320. [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
Data, materials, and code are available upon reasonable request to the corresponding author.
