A personalized approach to understanding food cravings and intake: a study protocol

Abstract

Background

Studies on food craving and consumption often overlook the interconnectedness of risk factors, assuming uniform mechanisms that drive individuals to (over)consume food. This project seeks to address this gap by leveraging a precision health framework to explore whether multimodal clustering can predict weight and eating outcomes after six months, providing a more nuanced understanding of individual variability.

Methods

The project will include a longitudinal study, encompassing several sub-studies where self-report, electrophysiological, and time series dynamic data will be collected at three time points. At baseline, participants will complete comprehensive assessments, including an electroencephalography (EEG) experiment and a one-week experience sampling study (ESM). Machine learning techniques will be employed to uncover distinct participant clusters, characterized by unique patterns of food consumption and weight changes over six months. Markers that best differentiate these profiles will be identified with explainable AI techniques, which aim to make machine learning model outputs understandable by highlighting the key features or patterns driving predictions, enabling personalized insights into key factors contributing to eating behaviors and weight management.

Discussion

By exploring the variability of mechanisms influencing food consumption, eating regulation, and weight gain, we aim to uncover subgroups of individuals who are most affected by specific influences, such as stress, emotion regulation difficulties, or sleep deprivation. This project will advance theoretical understanding by integrating multimodal data and emphasizing idiographic methods to capture individual variability. Findings will provide a foundation for future research on precision approaches to eating behaviors and may offer insights into personalized strategies for prevention and management of both normative and disordered eating patterns.

Background

Obesity and overweight are associated with significant health risks for the individual [1] as well as economic costs to society [2]. The World Health Organization (WHO) has identified obesity as a critical public health concern and has emphasized the necessity of prevention efforts [3], given the relative ineffectiveness of existing treatment approaches [4]. Despite extensive public health initiatives, the prevalence of obesity and overweight continues to rise [5], highlighting the urgent need for more effective and innovative prevention strategies. These should especially focus on adolescents and young adults [3], considering body mass index (BMI) during early adulthood is positively related to BMI over the lifetime [6].

While most research and public health strategies focus on environmental and biological determinants, such as marketing, and food availability or genetics [7, 8], less attention has been given to the psychological drivers of eating behavior. The cognitive, emotional, and behavioral processes that contribute to eating behavior—especially episodes of overeating—play an important role in the development and persistence of obesity but are often overlooked in prevention models [9]. While individuals who are overweight and obesity are generally aware of the need to eat less and exercise more, most struggle to sustain these changes long-term, with over 75% regaining lost weight [4]. This highlights the need to better understand the (neuro)psychological mechanisms underlying eating behavior, such as food cue reactivity (FCR) and craving, emotion regulation, personality traits, social context, and sleep – factors that may increase susceptibility to overeating and, over time, contribute to unhealthy weight gain or obesity.

Existing research often examines these factors in isolation, yielding inconsistent findings and limited applicability to real-world settings. Additionally, the studies that focus on the psychology of eating usually focus on comparing overweight and normal-weight individuals, presuming homogeneity within groups. Recent evidence challenges this assumption [10], with a highly influential review paper concluding that “.the pattern emerging from studies comparing obese individuals and binge-eaters with controls is most remarkable for its variability and inconsistency” [11, p. 283]. For example, neuroimaging studies of food reward have yielded widely divergent results when comparing individuals with obesity to normal-weight controls: some report heightened activation in reward-related brain regions in the former group, whereas others observe attenuated or negligible differences (for a detailed review see [10]). Such mixed findings likely reflect variations in cognitive and motivational context (e.g. whether attention is focused on a food’s palatability vs. its health implications) rather than stable, inherent differences between BMI-defined groups. ​This variability underscores the importance of exploring patterns beyond traditional BMI-based categories.

The proposed project will address these gaps by studying a large sample of young adults (with no current or past eating disorders) with a varying range of BMI. Using machine learning, we will analyze multimodal data — including neurocognitive markers, self-reports, experience sampling (i.e., repeated real-time assessments of experiences and behaviors in daily life), and behavioral measures—to identify data-driven clusters of participants with shared patterns across these domains. The planned approach addresses the limitations of traditional group-level methods, providing insights into eating-related behaviors and outcomes aligned with the principles of personalized or precision medicine.

Food cue reactivity and cravings

Modern obesogenic environments are saturated with food cues that signal the continuous availability of high-calorie foods. Exposure to these cues activates FCR [12], a central appetitive state encompassing psychological (i.e., craving), physiological (i.e., salivation), and neurocognitive (e.g., attention allocation) responses [13]. Cravings, a key component of FCR, are defined as an intense urge to eat a specific food [14]. They drive overconsumption of calorically dense foods [15] and subsequently contribute to an increased risk for overweight and obesity [16]. Cravings and food intake are not only regulated by metabolic needs, but are also affected by several other factors, such as self-regulatory processes, the physiological state of the person, emotional states, and habitual processes. The interplay of these mechanisms can lead to repeated episodes of overeating, which, over time, may contribute to weight gain and difficulties in eating regulation. Below, we briefly summarize the current state-of-the-art research on each of these contributing mechanisms and emphasize important gaps in the literature that require further study, which will be addressed in the project.

Craving and consumption as a regulatory failure

Craving can be defined as an affective state that can be modulated with the use of regulation strategies [17]. Emotion regulation strategies can modulate cravings by redirecting attention to specific aspects of food, such as its health benefits or the long-term consequences of overconsumption [18, 19]. Studies show that such strategies engage the brain’s self-control networks [20, 21], reducing neural responses to food cues and decreasing motivated attention toward food [22]. Conversely, focusing on immediate gratification of eating enhances neural activity in reward-processing regions [20, 21], prolonging motivated attention toward food [23, 24]. Despite promising evidence that neural activity predicts weight changes [25] and craving for healthy foods [26], existing studies often rely on retrospective self-reports, which can be biased [27]. Further research is needed to clarify how neural correlates of (un)successful regulation relate to cravings and food intake in real-world contexts.

Craving and consumption as a response to internal States

Internal states, such as emotions, stress, and sleep deprivation, can also become associated with food consumption [13]. Negative affect, in particular, has been shown to increase food cravings and consumption [28, 29], often as a maladaptive strategy for emotion regulation [30]. Beyond mean levels of affect, emotion differentiation—the ability to identify and label emotions precisely [31]—plays a critical role in regulating eating behavior . Lower levels of emotion differentiation could hinder attempts for successful regulation of one’s emotions, which would, in turn, increase cravings and food consumption. Research on the role of emotion differentiation in craving and eating behavior is scarce. To date, only a few studies examined these relationships with mixed results [28, 32–34]. Nevertheless, the preliminary results show that low emotion differentiation is a risk factor for experiencing higher food cravings, potentially leading to increased food intake over the long run.

Similarly, stress also plays a critical role in eating behaviors and has been identified as a risk factor for obesity, as it both decreases energy expenditure and increases energy intake [35, 36]. Neuroscientific evidence highlights substantial overlap between the processing of stress and food cues, indicating a shared pathway that predisposes individuals to overconsumption [37]. Stress is associated with an increased intake of unhealthy foods and decreased intake of healthy foods [38, 39] as well as greater levels of emotional and binge eating [39]. These effects, however, seem to be dependent upon the time frame of the stress response. Chronic stress (relative to acute stress), is associated with increased intake of unhealthy food [40], maladaptive eating behaviors [41], and increased neural response to food cues [42].

Sleep is another critical internal state influencing food cravings and consumption. Poor sleep quality and duration amplify neural responses to food cues, bias attention toward high-caloric foods, and impair emotion regulation [43–45]. Reduced self-regulation mediates the relationship between poor sleep quality and increased cravings, predisposing individuals to unhealthy eating patterns and weight gain [46]. Moreover, sleep deprivation biases an individual’s attention toward negative emotion-inducing stimuli, potentiating the effect of negative mood on craving [47] whilst simultaneously disrupting its effective regulation [48]. Given the widespread problem of rising population-wide sleep deprivation, understanding the interplay between sleep and other internal states is essential for addressing maladaptive eating behaviors.

Craving and consumption as a response to the social context

Beyond internal states, social context plays a pivotal role in shaping eating behaviors. Social interactions influence food intake through mechanisms such as social modeling and interpersonal emotion regulation [49, 50]. The association between social context and food intake appears to be highly dependent on the specific characteristics of the environment. On the one hand, oxytocin, a hormone linked with social engagement, reduces neural responding to food cues [51]. In line with this, positive social support was shown to mitigate emotional eating and stress-induced food intake [52]. On the other hand, social interactions can also promote overeating in certain contexts [53]. Understanding the nuanced effects of social environments on eating behaviors is crucial for explaining individual variability in eating behavior and informing tailored interventions.

The role of maladaptive eating traits in craving and consumption

Finally, individual differences in maladaptive eating traits, such as stress-related, emotional, external, disinhibited, and restrictive eating, significantly affect food consumption patterns. They refer to an individual’s dispositional tendency to alter their eating behavior in response to various factors, such as stress (i.e., stress eating), emotions (i.e., emotional eating), and external cues in the environment (i.e., external eating), regardless of their internal hunger and fullness cues. Examples of maladaptive eating traits also include disinhibited eating, a general tendency toward loss of control over food intake, and restrictive eating, a propensity to limit and control food consumption. These traits are often present on a continuum, overlapping with normative behaviors, and are strongly linked to changes in eating behavior [54, 55]. For instance, restrictive eating moderates the relationship between negative affect and increased food intake, with heightened FCR observed in high trait emotional eaters under negative affect [30, 56, 57]. Similarly, stress eating and disinhibited eating exacerbate the effects of stress and poor sleep on food intake and long-term weight gain [58, 59]. Exploring the interplay between maladaptive traits and broader psychological mechanisms is necessary to gain insights into the variability in eating behaviors. It is important to note that the present study focuses on non-clinical variations in eating behavior; individuals with a current or past diagnosis of an eating disorder are excluded from participation.

Individual differences across factors

While the project examines a broad array of factors, it is critical to recognize that individuals respond differently to these influences. For instance, some participants may be more impacted by stress, while others may exhibit heightened sensitivity to sleep deprivation. Similarly, the role of emotional states, such as negative affect or poor emotion differentiation, may vary across individuals. By capturing this variability, the project aims to identify distinct subgroups (specifically, clusters generated by machine learning) of individuals, each influenced by a unique combination of factors. These findings have the potential to advance our understanding of the interplay between various mechanisms driving food cravings and consumption, paving the way for more nuanced and precise insights into eating behavior.

Proposed research approach

This project will integrate multimodal data streams, including self-reports, EEG recordings, and experience sampling methodology (ESM; repeated real-time assessments of participants’ experiences and behaviors in daily life) data, enabling a comprehensive analysis of the neurocognitive, behavioral, and emotional mechanisms underlying eating behaviors. At baseline, these multimodal data will be used to identify distinct clusters of individuals (see Fig. 1) using advanced machine learning techniques [60]. These clusters will provide personalized predictors of eating behaviors and weight outcomes at three- and six-months, offering a novel alternative to traditional BMI-based models, which typically categorize individuals into predefined weight groups (e.g., normal weight, overweight, obesity) and examine differences in psychological and behavioral factors between these groups. We hypothesize that machine-learning based clusters, which integrate multimodal individual-level data, will outperform BMI-based models in predicting food intake, eating regulation, and weight change over time1. Specifically, we expect that these clusters will reveal distinct subgroups characterized by unique combinations of neurocognitive, behavioral, and psychological factors. As part of our analysis, we will also use explainable AI, which refers to methods that make machine learning model predictions interpretable and highlight which features most influence outcomes [61], to identify markers that have a maximal diagnostic power to differentiate each cluster. These insights may inform future studies by highlighting specific variables that most significantly contribute to the individual differences in weight and food-related outcomes. Such findings will not only build on existing research demonstrating the heterogeneity in eating behaviors and weight-related outcomes [11], but can also offer new insights into the multifactorial mechanisms underlying these behaviors. In the field of craving and FCR, studies like this are rare [7, 62]. The planned data-driven clustering approach overcomes the limitations of traditional group-level methods by capturing individual variability in eating-related behaviors and outcomes.

Fig. 1.

A graphical representation of the approach used in the project

Methods

Study design and procedure

The planned study uses a prospective, mixed method design, encompassing EEG recordings and self-reports collected both in the laboratory and in naturalistic settings . Participants will be assessed over a six-month period; at baseline (T1), after three months (T2), and after six months (T3). Data collection includes self-report questionnaires, EEG recordings, and an experience sampling (ESM) protocol (see Fig. 2).

Fig. 2.

A graphical representation of study procedures. Note. Full lines represent the sequence of activities; dotted lines represent data flow

At T1, participants will complete baseline questionnaires, undergo an EEG-based craving regulation experiment, and participate in a one-week intensive longitudinal experience sampling (ESM) study via the SEMA3 application, a mobile platform that delivers real-time surveys to participants’ smartphones [63]. During the ESM study, participants will provide real-time reports on cravings, food intake, mood, and contextual factors multiple times per day. Data collected during T1 will be used to build personalized participant clusters for subsequent comparisons and analyses (see Data analysis). At T2 and T3, participants will complete follow-up questionnaires assessing changes in body weight, food intake, eating regulation strategies, and perceived stress over the preceding three months. Data collected during T2 and T3 will serve as the primary outcome variables of interest, providing insights into changes in eating behavior and weight regulation over a six-month period.

Participants and recruitment

Participants will be young adults aged 18–35 years who meet the following inclusion criteria: 1) right-handedness (to minimize variability related to hemispheric dominance in EEG measures), 2) ownership of a smartphone capable of running the SEMA3 application, and 3) [no current or past diagnosis of eating disorders, neurological disorders, or neuromuscular diseases (checked via self-report; see the Baseline assessment section). Recruitment efforts will aim to achieve a diverse sample with regard to gender and BMI. Participants will be primarily recruited through convenience sampling and targeted outreach. Convenience sampling will focus on students and their social networks from partner institutions, with additional recruitment via social media, students’ mailing lists, the research team’s networks, and flyers. To ensure BMI diversity, targeted advertisements on social media platforms and outreach to organizations serving individuals with underweight or overweight conditions may be employed.

Sample size determination

Since the ESM component of the project involves the largest number of observations and a complex multilevel structure, we based the effect size calculations on ESM data. Parameters were derived from prior experience sampling studies conducted by our team, which found small within-person effect sizes (standardized coefficients approximately ranging from 0.10 to 0.20) in related domains such as eating behavior, craving regulation, and emotion regulation. Based on this, the expected effect size for the within-person predictor is considered small. In line with this, we aim to recruit at least 250 participants at baseline, allowing for 80% power to detect small level-1 effects in multilevel analyses, and accounting for attrition and missing data (α = 0.05; [64]).

In an effort to increase participant motivation and minimize attrition, several incentives will be offered to participants multiple times during the study. Students from partnering institutions will receive university extra course credit for their participation. All participants will get the opportunity to enter prize draws for shopping vouchers or nutritional counselling as well as receive detailed personalized feedback reports on their results. Due to the lengthy data collection process, prize draws will occur at regular intervals throughout data collection, with approximately 50 participants entering at once.

Materials and data collection

Baseline assessment (T1): multimodal data collection

This section captures all data gathered at baseline, during the first laboratory visit and subsequent week, used to generate individual profiles and machine learning clusters.

Self-report questionnaires and demographics

At T1, participants will provide demographic information, including age, gender, and education level. Additionally, we will assess their average weekly physical activity, diet (e.g., omnivore, vegetarian), eating goals, and food insecurity (operationalized as participants’ access to healthy food and measured using the Food Insecurity Experience Scale; [65]. Information on physical and psychiatric disorders, as well as medication intake that might interfere with food consumption, will also be collected via self-report. Participants will complete a checklist of physical health conditions (e.g., metabolic, cardiovascular, endocrine, gastrointestinal, and immune-related disorders) and psychiatric disorders, indicating current or past diagnoses. They will also report regular medication use that could impact appetite, eating behavior, or weight regulation. Participants’ height and weight will also be recorded at the time of their first lab visit, their BMI will be calculated using the formula (weight in kilograms divided by height in meters squared).

Additionally, they will complete the questionnaires listed below.

Food craving, behavior, and experiences. Participants will complete several questionnaires pertaining to different aspects of food cravings and food intake. More specifically, we will gather data on 1) trait food craving (Food Craving Questionnaire Trait – Reduced; FCQ-T; [66]); 2) eating behavior traits, such as emotional eating, external eating, restrained eating (Dutch Eating Behavior Questionnaire, DEBQ, [67]), stress eating (Salzburg Stress Eating Scale; SSES; [68]) and uncontrolled eating (Uncontrolled Eating dimension of the Three-Factor Eating Questionnaire R-18; TFEQ; [69]), 3) habitual food intake in the past three months (Healthy and Unhealthy Eating Behaviors Scale; HUEBS; [70]). Eating regulation success will be assessed using the Perceived Self-Regulatory Success in Dieting Scale (PSDS; [71]) and the Weight Efficacy Lifestyle Questionnaire (WEL; [72]).

Emotion regulation. Participants will complete a general measure of emotion regulation difficulties (Difficulties in Emotion Regulation Scale: DERS; [73]) and a measure pertaining to the tendency to use social emotion regulation strategies (i.e., relying on others during emotional experiences; Interpersonal Emotion Regulation Questionnaire (IERQ; [74]). Additionally, participants will complete a measure of momentary positive and negative affect immediately preceding the EEG experiment (Positive and Negative Affect Schedule; PANAS; [75]).

Sleep. Participants will complete the Pittsburgh Sleep Quality Index (PSQI; [76]), which assesses sleep quality and disturbances over the last month.

General emotional well-being. Participants will also complete a series of measures pertaining to experiencing depression, anxiety, and stress (Depression, Anxiety and Stress Scales, DASS; [77]); Perceived Stress Scale, PSS; [78]; Fatigue scale from the Profile for Mood Scales; POMS; [79]).

EEG experimental paradigm and stimuli

Participants will complete an experimental food craving regulation task (see Fig. 3) while undergoing simultaneous EEG recording. The paradigm is designed to assess neurophysiological and self-reported responses to visual food stimuli, particularly the ability to regulate craving in response to food stimuli. During the task, participants will view 128 images of high and low-calorie food presented for 4000 ms each, preceded by a fixation cross (500–1000 ms) and regulation instructions (4000 ms). All pictures will be rated in three dimensions: wanting, liking, and arousal, using a 7-point Likert scale. The images were sourced from the Food-pics database [80] and supplemented with images obtained through internet searches. While the specific images are not identical across the four regulation conditions, they depict the same types of food items (e.g., ice cream, pizza, or salads) and are matched on key visual properties critical for EEG research, including color, brightness, and saturation. To minimize variability in hunger and satiety levels, participants will be instructed to abstain from eating for at least two hours prior to the EEG session and to avoid consuming heavy meals immediately beforehand. Light snacks are permitted if needed to prevent discomfort from excessive hunger. Participants will also report their current hunger level immediately before the EEG task using a self-report measure.

Fig. 3.

An example of a single trial in the EEG paradigm

Experimental design

The study will follow a 2 (Picture type: High-Calorie, Low-Calorie) x 4 (Regulation condition: Look, Now, Later Positive, Later Negative) within-subject experimental design, where participants will view blocks of food images under different regulation conditions. In the Look condition, participants will be instructed to simply look at the picture and respond to it naturally. Following the Now instruction, participants will attend to the presented food picture by imagining its taste, texture, and smell. The Later Positive and Later Negative conditions are intended to garner participants’ attention to the long-term consequences of food consumption, with the Later Positive focus tending to positive goal-congruent consequences, and Later Negative tending to negative goal incongruent consequences. After each block, participants will evaluate their success in following the regulation instructions. The sequence of the conditions will be pseudo-randomized across participants, starting with the Look condition followed by a randomized order of the other three conditions. The pseudo-randomization was chosen after piloting, which indicated difficulty returning to a neutral viewing state after prior regulation efforts. The order of high vs. low-calorie pictures will be randomized within each block.

EEG setup

EEG data will be recorded using the ActiCHamp system (ActiCHamp, Brain Products GmbH, Gliching, Germany). Participants will wear a cap fitted with 63 active actiCAP electrodes, positioned according to the 10–10 system. Signals will be recorded with Brain Vision Recorder, using the FCz position as the reference electrode and the FPz position as the ground electrode. An electrolyte gel will be applied to ensure optimal conductivity.

Experience sampling questionnaire (ESM)

The experience sampling study will consist of two survey types that will be completed over seven consecutive days, starting the day after the lab visit.

Morning questionnaire

Sleep. Participants will complete a morning survey between 8:00 and 11:00 am. They will report on the time they went to bed and got out of bed, the time they needed to fall asleep, and the time they actually slept. Additionally, they will report on sleep quality using a rating scale from 0 (very bad) to 100 (very good). The questions were adapted from The consensus sleep diary [81].

Day questionnaire

Participants will complete the second survey type five times a day every three hours starting at 10 am (10 am, 1 pm, 4 pm, 7 pm, and 10 pm).

Positive and negative affect. Participants will assess nine emotions (relaxed, calm, enthusiastic, excited, sad, bored, anxious, irritable, stressed) using a 100-point visual analog scale (0 = not at all, 100 = very much), indicating how [emotion] they felt at the moment in a randomized order. We included four positive and four negative emotions, with half of these items representing low-arousal and half representing high-arousal negative and positive affect, covering the circumplex model of affect [82].

Current craving. Participants will rate the intensity of their current craving (e.g., “How strong is your desire for specific foods at the current moment?“) on a scale from 0 (no craving) to 100 (extremely intense craving). Additionally, they will indicate what type of food they were craving the most by choosing pre-specified food type categories: salty snacks (e.g., chips, pretzels), sweets (e.g., chocolate, cookies), fatty foods (e.g., burger, pizza, fries), starchy foods (e.g., bread, pasta, rice), dairy products (e.g., yogurt), vegetables/salad (e.g., tomatoes, carrots) and fruit (e.g., apples, berries). The food categories were derived from the Yale Food Addiction Scale 2.0 [83] and adapted to include healthy food categories in line with previous research [84]. At this point, they will also be able to indicate that they do not have a craving for specific food.

Food intake. Participants will report their food intake (and the context) over the last three hours. More specifically, they will report whether they ate something in the last 3 h (YES/NO). If they choose YES, they will indicate the type of food they consumed by selecting all that applied from the list (multiple answers could be selected). The list will include the same categories as for the craving, with the addition of “other”. Furthermore, they will indicate whether they ate alone or in company (ALONE/IN COMPANY), to what extent they ate because they were hungry or because something tasty was presented. Furthermore, they will also indicate whether they experienced a loss of control while eating. All items will be rated on a VAS scale ranging from 0 (not at all) to 100 (very much). If they report not consuming food, they will answer neutral items (pertaining to their activities) matched according to the type and length of questions.

Social interactions. Participants will report on several aspects of their social interactions during the last three hours. Namely, whether they spent time with others (YES/NO), whether this interaction was pleasant (scale from 0 to 100) and also whether food was present during the interaction (YES/NO). If participants report no social interactions during the past three hours, they will answer neutral items matched according to the type and length of questions.

Follow-up assessment (T2 and T3)

These will be completed at 3- and 6-months after taking part in the EEG and ESM part of the study. At each follow-up, participants will complete general questions about their activity level and potential food-related goals during the past three months. Additionally, they will complete questionnaires assessing the constructs described below.

Healthy and unhealthy eating behavior, where participants indicate the extent to which they consumed different food items during the last three months, assessed with the Healthy and Unhealthy Eating Behaviors Scale [70].

Craving and food intake regulation success during the last three months will be assessed using the Weight Efficacy Lifestyle Questionnaire [72] where participants rate their ability to resist the desire to eat across five-domains (negative emotions, availability, social pressure, physical discomfort, and positive activities). Additionally, participants will complete the Perceived Self-Regulatory Success in Dieting Scale [71] and the uncontrolled eating dimension of the Three-Factor Eating Questionnaire (89; see baseline measures for more information).

BMI. Participants will report their current weight at T2 and T3 and indicate whether the reported weight is based on a recent actual measurement or an estimate, and provide the date of their last weigh-in. They will be encouraged, when possible, to weigh themselves shortly before completing the survey using a standardized scale.

Perceived stress during the last three months will be assessed using the Perceived Stress Scale (PSS-10; 80), a widely used and validated self-report measure of subjective stress.

Cluster analysis framework

Input variables for cluster generation

The following variables, alongside baseline measures (see Baseline Measures for details) will serve as inputs for machine learning clustering.

EEG-derived features

Subjective ratings. During the EEG experiment, participants will rate wanting, liking, and arousal after each presented image using a 7-point Likert scale. Additionally, participants will report the extent to which they were successful in following the instructions.

ERP amplitudes. In the analysis of EEG recordings, we will focus our analysis on the P200 (180–250 ms after stimulus presentation), P300 (300–500 ms after stimulus presentation), and the early (300–600 ms) and late LPP (600–4000 ms) components, which are important for processing emotionally relevant stimuli (e.g., [84, 85]) and stimuli related to eating behavior [86–88]. In combination with a visual inspection of the data, we will focus on frontal, central, and parieto-occipital electrode clusters, where the P200, P300, and LPP components are most prominent [89, 90]. Data will be averaged for all groups and conditions separately.

Experience sampling-derived features

Craving. Participants will report craving intensity and craving type (e.g., sweet, salty, fatty foods). See Experience sampling questionnaire for more details.

Food intake. Participants will report the type of food consumed (e.g., sweet, salty, fatty foods), as well as the total number of food consumption instances over a period of 7 days, the intensity of loss of control when eating as well as hunger-based eating (see Experience sampling questionnaire for more details).

Emotion differentiation. Positive and negative emotion differentiation will be computed with the intraclass correlation coefficient (ICC), using agreement with averaged raters to compute the average consistency between negative emotions across time [91]. More specifically, we will calculate the ICCs across measurement occasions within individuals. Negative values will be recoded to 0. ICCs will then be normalized using a Fisher’s r-to-z transformation and reverse scored for the purposes of better interpretability, with higher scores indicating higher differentiation, following past work [92].

Outcome variables for predictive validation (main hypothesis)

The main outcome variables that will be used to evaluate the predictive utility of the clusters identified using the input variables will be assessed during follow-up measurements, 3- and 6-months after completing the EEG paradigm and the ESM study. The measures evaluate participants’ progress in eating regulation and weight management over time. More specifically the main outcome variables will be [1] healthy and unhealthy eating behavior [2], uncontrolled eating, and [3] eating regulation success (please refer to the section Follow-up assessment (T2 and T3) for more information).

Data processing and statistical analysis

Variable transformations prior to subsequent analyses will be determined based on distribution diagnostics and outlier analysis.

Data Preparation and integration

Data analysis will integrate multiple data streams collected during the study, including self-reports, EEG recordings, and experience sampling data. Data preparation will involve preprocessing to ensure data quality and usability across modalities. For EEG data, preprocessing will include artifact rejection, filtering, and segmentation based on the experimental paradigm. Analysis of raw EEG data will be conducted using BrainVision Analyzer (version 2.3.0). Artifacts due to eye movements will be corrected via the implemented ICA correction software; components corresponding to horizontal and vertical eye movements will be selected based on their shape, timing, and topography. Additional artifact episodes (e.g., due to movement) will be excluded after a visual inspection in each trial using a semiautomated procedure with artifacts identified if the voltage difference is equal or greater than 200 µV within a trial. For experience sampling data, responses will be validated for compliance and quality. Validation will include checks of response time anomalies, and compliance with the protocol. Data will be reshaped into a long-format structure for multilevel modeling, accommodating the nested design of repeated measurements within individuals.

Machine learning-based clustering and interpretation

Unsupervised clustering algorithms (e.g., k-means, hierarchical clustering) will be applied to uncover participant subgroups. Cluster quality will be assessed using validation metrics such as silhouette score and Davies-Bouldin index. To enhance the interpretability of machine learning-derived clusters, explainable AI methods will be incorporated [93]. Initially, explanations of the patterns on the global level will be made with state-of-the-art approaches (e.g., explanation by example, interpretable proxy models, and feature importance estimation). By using these methods, we will obtain a more in-depth understanding of the specifics of each distinct profile/cluster of individuals and identify the main features of each cluster (global explanation). Next, the individual’s traits will be explained with various local explanation approaches (e.g., SHAP and LIME). This will enable a personalized characterization of each individual by identifying the main attributes by which they were assigned to a specific cluster (local explanation).

Statistical modeling of cluster outcomes

Once clusters are defined, their predictive validity will be assessed using appropriate statistical models tailored to the data structure. Three- and six-month outcomes (e.g., eating regulation success, healthy/unhealthy eating behaviors, BMI changes) will be analyzed using linear mixed-effects models to account for within-person correlations.

Cross-sectional analyses (e.g., baseline differences between clusters) will be conducted using linear regression or logistic regression, depending on the distribution and type of the outcome variable. Moderation analyses will also be performed to explore interaction effects between cluster membership and contextual variables (e.g., stress levels, sleep quality).

To assess whether machine learning-based clusters offer added value over conventional approaches, these models will be compared to traditional BMI-based models (i.e., predicting changes over time based on BMI-group membership) using metrics such as F1-score and area under the ROC curve (AUC-ROC).

Study timeline

Data collection for the study commenced in January 2025, following ethics approval and submission of the study protocol for peer review. Recruitment and data collection are ongoing and are expected to continue through June 2026, depending on the pace of participant enrollment and retention across the three timepoints.

Discussion

This project aims to investigate whether personalized participant clusters—derived through machine learning from multimodal self-reported, electrophysiological, and longitudinal data—can predict outcomes related to food intake, eating regulation, and changes in BMI after three and six months.

The project is expected to make several contributions to the field. First, by employing a machine learning-based clustering approach, the study seeks to identify multidimensional profiles of risk – data-driven subgroups of individuals characterized by specific patterns across neurocognitive, behavioral, and psychological markers that influence eating behavior. This framework represents a departure from traditional group-level methods, which rely on population averages and often fail to account for the heterogeneity in eating behaviors and outcomes. For example, clusters may reveal subgroups with distinct patterns of heightened FCR, poor emotion differentiation, or adaptive eating behaviors, offering a new perspective on how these factors interact to shape eating behavior and weight regulation.

Moreover, the longitudinal design of the study, with follow-ups at 3 and 6 months, will allow for the evaluation of the predictive utility of the identified clusters. By linking these clusters to outcomes such as eating regulation success, healthy/unhealthy eating behaviors, and BMI, the project advances theoretical understanding of the mechanisms underlying eating regulation while setting the stage for future research applications.

Limitations and challenges

The planned protocol faces potential limitations and challenges. One challenge relates to the integration of multiple data streams, such as EEG, ESM, and baseline measures, which may introduce methodological complexities. While machine learning methods are well-suited to handling such data, their results are dependent on the quality of the input data and require careful interpretation. Replication in independent datasets will be critical to confirm the robustness and generalizability of the clustering approach.

Another challenge lies in the study’s scope and intensity. The baseline protocol includes several self-report instruments, an EEG task, and a 7-day ESM protocol with multiple daily prompts. These are followed by two assessments at 3 and 6 months. While this multimodal, longitudinal design is central to the study’s aims, it also imposes a considerable burden on participants and demands substantial logistical coordination. High task demands increase the risk of attrition, partial non-compliance, and data loss — especially across the follow-up period. To address these risks, we extensively piloted all baseline components and optimized study procedures across subdomains (EEG, ESM, questionnaires). We implemented strategies to reduce participant burden, including modular scheduling, staged and performance-based incentives, and personalized reminders. These procedures have proven successful in our previous ESM studies (e.g., 15). Despite these measures, some participants may not complete all components. We will monitor attrition and data quality closely and will transparently report and statistically account for missingness in downstream analyses.

Another limitation is the reliance on retrospective self-reports at follow-up. Although participants indicate the timing and source of their weight estimates, recall biases remain possible. Similarly, although validated instruments (e.g., HUEBS, WEL) are used to assess generalized eating patterns, self-reports cannot fully capture behavioral dynamics. Still, given the practical constraints of conducting follow-ups in a large sample, these measures offer a feasible and methodologically sound approach.

Although the sample size has been calculated to ensure sufficient power, the generalizability of findings may be limited by the demographic characteristics of the sample. Recruiting a diverse sample in terms of age, socioeconomic status, and cultural background will be essential for ensuring broader applicability of the findings.

Future directions and applications

This protocol establishes a foundation for research on individual differences in the psychological and neurocognitive mechanisms that influence eating behavior. By uncovering distinct, data-driven profiles (machine learning-based clusters) of individual variability, the findings will contribute to a more nuanced understanding of how neurocognitive, behavioral, and psychological factors interact to shape eating behaviors. Future research can build on these findings to explore how identified profiles evolve over time, interact with contextual factors, and influence six-month trajectories in weight and eating outcomes. While intervention development is not the primary aim of this project, the clusters identified may generate hypotheses for future applied research. For example, targeted experimental studies could examine whether specific strategies (e.g., mindfulness-based approaches or emotion regulation training) are more effective for subgroups with particular profiles. Such insights would bridge the gap between theoretical advances and practical applications in precision health.

Conclusion

The study protocol outlines an innovative approach to understanding the neurocognitive, behavioral, and emotional underpinnings of eating behavior. The project is among the first to assess whether personalized clusters, based on readily modifiable psychological variables obtained in both laboratory and naturalistic contexts, can successfully predict weight and eating outcomes over six months. By leveraging multimodal data and machine learning, the planned project diverges from nomothetic approaches by emphasizing individual variability in eating-related behaviors. This idiographic focus has the potential to advance theoretical knowledge and lay the foundation for future precision health initiatives.

Abbreviations

EEG

Electroencephalography

ESM

Experience sampling methodology

WHO

World health organization

BMI

Body mass index

FCR

Food cue reactivity

AI

Artificial intelligence

T1

First time-point

T2

Second time-point

T3

Third time-point

FCQ-T

Food craving questionnaire – trait

DEBQ

Dutch eating behavior questionnaire

SSES

Salzburg stress eating scale

TFEQ

Three-factor eating questionnaire R-18

HUEBS

Healthy and unhealthy eating behaviors scale

WEL

Weight efficacy lifestyle questionnaire

PSDS

Perceived success in dieting scale

DERS

Difficulties in emotion regulation scale

IERQ

Interpersonal emotion regulation questionnaire

PANAS

Positive and negative affect schedule

PSQI

Pittsburgh sleep quality index

DASS

Depression, anxiety and stress scales

PSS

Perceived stress scale

POMS

Profile for mood scales

LPP

Late positive potential

ICC

Intraclass correlation coefficient

ICA

Independent component analysis

Author contributions

SZ: Conceptualization, Methodology, Writing – Original Draft Preparation, Supervision, Funding Acquisition. SK: Methodology, Software, Visualization, Writing – Reviewing and Editing. MH: Writing – Reviewing and Editing. SMB: Writing – Reviewing and Editing. ŽK: Conceptualization, Methodology, Writing – Original Draft Preparation, Project Administration.

Funding

This research project is funded by the Slovenian Research and Innovation Agency (ARIS J5-50176) awarded to SZ. The funding source was not involved in the study conceptualization, design, decision to publish and preparation of this manuscript nor will it be involved in data collection and analysis.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

The study will be carried out in accordance with the 1964 Declaration of Helsinki. Ethics approval was obtained from the local ethics committee (038-27-193/2024/7/FF/UM). Participation in the study will be voluntary, and all participants will provide written informed consent prior to commencing.

Competing interests

The authors declare no competing interests.