This study was a 14-day single-arm observational investigation designed to examine the relationship between real-time self-reported hunger and glucose fluctuations. The study integrated CGM and EMA technologies to capture real-time physiological and behavioral data in free-living conditions. The study protocol, including recruitment, screening, and monitoring procedures, is described in detail in the published protocol paper by Rethorst et al [20]. The protocol was approved by the Texas A&M University Institutional Review Board (IRB2021-1412F). The study procedures followed the previously published protocol, and no deviations were made.

Inclusion criteria required individuals to be between 18 to 65 years of age, have a BMI ≥25, be fluent in English, and own a smartphone. Exclusion criteria included (1) current or prior diagnosis of type 1 or type 2 diabetes, (2) pregnancy, (3) diagnosis of a gastrointestinal or eating disorder, and (4) regular (defined as daily or near-daily) use of medications or supplements known to interfere with CGM accuracy or glucose metabolism, including corticosteroids, metformin, vitamin C, aspirin, and acetaminophen. Participants were recruited by community flyers, social media advertisements, and an email contact list consisting of community members who opted to be notified about upcoming research studies. Interested individuals completed an online screening survey consisting of health history and behavioral eligibility questions. Eligible individuals were invited to attend an enrollment visit at which eligibility was confirmed through objective measurements (eg, height and weight) and verification of self-reported medical history and medication use. All participants completed an informed consent document prior to completing any screening or baseline procedures and received up to US $71 for participation. Participants were enrolled from October 5, 2022, to April 26, 2023.

As this study was based on a previously developed observational protocol with a target sample size of 100 participants, the sample size was determined based on feasibility considerations [20]. Individuals who did not respond to recruitment efforts or declined participation were not included in the study, and reasons for nonparticipation were not systematically recorded.

All data were time stamped and processed using Python (version 3.11; Python Software Foundation). As the 2 data streams (CGM and EMA) differ in sampling frequency and timing, preprocessing was necessary to align CGM readings with EMA entries and construct a unified time-resolved dataset. Glucose values were extracted at the closest time point to each EMA and averaged over successive 30-minute post-EMA intervals (ie, 0‐30, 30‐60, 60‐90, 90‐120, 120‐150, and 150‐180 minutes) to model temporal changes in glucose following self-reported hunger. GE was defined as glucose fluctuations exceeding +1.5 or –1.5 SD from an individual’s mean glucose level. This threshold was selected to focus on larger, more pronounced excursions that may have greater clinical relevance [22]. GE events were identified within each post-EMA window. Hunger responses were binary-coded (“yes”=1; “no”=0), and relevant covariates such as sex, BMI, and weekday versus weekend were merged into the analytic dataset. All preprocessing and transformation were performed using standard Python-based data science libraries [23] to align and compute CGM and EMA data [24,25].

Four mixed effects models were constructed to examine the relationship between self-reported hunger and glucose outcomes. We hypothesized that self-reported hunger would be associated with an increase in glucose level over time and an increased likelihood of a GE. All models included participant-level random intercepts to account for within-subject variability. The models included participant-level random intercepts to account for within-subject clustering; however, no explicit residual correlation structure (eg, autoregressive structure) was specified. Fixed-effect covariates included sex, BMI category, age, and weekday status. Given the relatively small number of male individuals enrolled in the study (n=13), models were not conducted separately for each sex. As illustrated in Figure 1, all models were aligned to the time stamp of the EMA response, using the closest available CGM value, which ranged from 0 to 14 minutes after each EMA prompt due to the 15-minute CGM sampling frequency. The 4 models were designed to capture distinct temporal relationships between hunger and glucose: immediate CGM, short-term mean, time-dependent trends, and finally the likelihood of GE. Missing data patterns were evaluated using Little’s test for missing completely at random.

This linear mixed model evaluated the association between within-person and between-person hunger and the CGM value recorded closest to each EMA prompt, representing the immediate glycemic response.

This model assessed the mean glucose level across the 30-minute period following each EMA prompt to determine the short-term association of hunger with glucose.

To examine time-dependent patterns of glucose regulation following self-reported hunger, this model assessed the cumulative average CGM glucose levels at 30-, 60-, 90-, 120-, 150-, and 180-minute intervals after each EMA prompt. The goal was to evaluate glucose trends over time following self-reported hunger.

A logistic mixed effects model with participant-level random intercepts was used to estimate the probability of experiencing a GE (as defined above) within multiple post-EMA intervals (30‐60, 60‐90, 90‐120, 120‐150, and 150‐180 minutes), with hunger as the primary predictor. Covariates included sex, BMI, age, and weekend status. All statistical models were implemented using the statsmodels package (version 0.14.0) in Python (version 3.11), and statistical significance was defined as P<.05. Data visualizations were created using matplotlib and seaborn [26,27]. Missing data from either EMA or CGM were ignored in the analyses, consistent with a complete case approach.

This study was approved by the Texas A&M University Institutional Review Board (IRB2021-1412F) on August 30, 2022. All participants provided informed consent before participating in this study. Participants received up to US $71 as compensation for their participation. All data were deidentified prior to analysis to ensure participant privacy and confidentiality.

A total of 1007 individuals completed the screening survey. Of these, 310 passed initial eligibility screening and completed intake forms. Subsequently, 194 were scheduled for a baseline visit. However, 110 either canceled, declined further participation, or were deemed ineligible during the phone screening or at the baseline appointment. The participant flow throughout the recruitment and screening process is presented in Figure 2.

The final analytic sample included 84 participants consisting of 13 men (15.5%) and 71 women (84.5%). Demographic and biometric characteristics, including age, BMI, sex, race, and ethnicity, are summarized in Table 1. Overall compliance was moderate to high across modalities, with average adherence rates of 73% for EMA responses and 79% for CGM wear. A test of missingness using Little’s test indicated that the data were missing completely at random (χ²₁₀=12.47; P=.25).

Model 1 predicted the CGM glucose value recorded immediately after each self-reported hunger entry via EMA. The linear mixed effects model showed that within-person hunger was significantly associated with lower immediate glucose levels (β=−12.5 to −14.9; P<.001). Between-person hunger, sex, BMI category, age, and weekday status were not statistically significant predictors. These findings are summarized in Table 2.

Model 2 assessed the average glucose level over the 30-minute period following each EMA response. The linear mixed effects model indicated a significant negative association between within-person hunger and 30-minute mean glucose levels (β=−11.0 +1.2 or −11.0 −1.2; P<.001), indicating lower glucose in the 30 minutes following self-reported hunger. No significant associations were found for between-person hunger, sex, BMI category, age, or weekday status. Full estimates for model 2 can be found in Table 2. As a sensitivity analysis, the models were reestimated using an autoregressive (1) correlation structure to account for temporal autocorrelation. The results were largely consistent with the primary analyses, with slight attenuation in effect sizes but no change in statistical significance. Within-person hunger remained significantly associated with lower glucose levels for both immediate glucose (β=−12.95; P<.001) and 30-minute mean glucose outcomes (β=−10.48; P<.001).

Model 3 examined average CGM glucose levels across multiple time points (30‐180 minutes) following each EMA prompt. Results from a linear mixed effects model indicated that glucose levels increased over time after participants reported being hungry but remained relatively stable when they were not. A significant time × hunger interaction was observed (P<.001); at 180 minutes, glucose levels were estimated to be 6.68 mg/dL higher following self-reported hunger. Female sex was marginally associated with lower glucose levels (β=−7.6; P=.05), and glucose levels were slightly lower on weekends (β=−0.5; P=.04). Age showed a marginal association, whereas BMI was not a significant predictor. These results are summarized in Table 3.

As illustrated in Figure 3, glucose trajectories diverged over time depending on hunger status. Participants who reported being hungry showed a steady rise in glucose levels from 30 to 180 minutes, while those who were not hungry maintained relatively stable or slightly declining levels throughout the same period.

Model 4 analyzed the likelihood of GE occurring within 5 post-EMA time intervals: 30‐60, 60‐90, 90‐120, 120‐150, and 150‐180 minutes Table 4. Results from the logistic mixed effects regression showed that hunger was a significant predictor of GE during specific intervals. Specifically, hunger was associated with a 23.7% increase in the likelihood of experiencing a GE within 30 to 60 minutes (P=.02), a 36.1% increase within 60 to 90 minutes (P<.001), and a 13% increase within 90 to 120 minutes (P=.01). Hunger was not a significant predictor in the remaining intervals (0‐30, 120‐150, 150‐180 minutes; P>.05). Other covariates, including sex, BMI, age, and weekend status, did not show significant associations with GE occurrence across any time interval.

This study investigated the dynamic relationship between self-reported hunger and glucose fluctuations in nondiabetic adults with overweight or obesity using real-time data from CGM and EMA. Across 4 statistical models, results consistently demonstrated that hunger is associated with meaningful short-term changes in glycemic patterns, suggesting the potential relevance of subjective behavioral signals in metabolic regulation.

Models 1 and 2 revealed that participants demonstrate significantly lower glucose levels at the time of hunger and during the immediate 30-minute period following reported hunger. These findings align with the physiological understanding of hunger as a marker of energy deficit and a signal for imminent food intake [10]. Building on these findings, model 3 provided further evidence for this interaction by identifying a sustained increase in glucose following hunger reports, which was not observed when hunger was absent. This trend likely reflects behavioral responses to hunger, suggesting that hunger may precede food intake, which could contribute to the observed increase in glucose levels [28,29]. The progressive divergence in glucose trajectories between hungry and not-hungry states illustrates how subjective experience may precede and predict measurable metabolic changes, particularly in naturalistic, free-living environments [30]. Model 4 reinforced these findings by demonstrating that hunger significantly increased the likelihood of experiencing a GE within the 30-minute to 120-minute window after EMA prompts. The strongest association occurred during the 60-minute to 90-minute window, potentially reflecting the lag between hunger detection, food intake, and resulting glycemic response.

Overall, these findings suggest that hunger may serve as a real-time behavioral indicator associated with subsequent eating behavior, which in turn is associated with downstream glycemic fluctuations. Rather than a passive subjective state, hunger may serve as a dynamic marker of metabolic need, offering a scalable and low-burden indicator for identifying periods of glycemic vulnerability, particularly among individuals at risk for metabolic dysregulation but not yet meeting diagnostic thresholds. This insight supports the integration of CGM and EMA as a useful framework for developing JITAI that respond to momentary behavioral and physiological cues. In addition, compliance across EMA and CGM suggests that the multimodal protocol was generally feasible. However, variability in adherence and device-related issues (eg, CGM detachment) highlight potential challenges for implementation in real-world settings.

This is clinically meaningful, as GE, even in nondiabetic individuals, are increasingly linked to insulin resistance, inflammation, and cardiometabolic risk [6]. These findings suggest that hunger could be explored as a potential trigger for delivering personalized prompts targeting diet, physical activity, or stress management during high-risk windows. By showing that hunger is followed by an increased likelihood of GE, this study suggests the potential to inform future strategies aimed at guiding food choices and reducing metabolic risk. For example, hunger reports could be used to trigger JITAI that deliver tailored recommendations (eg, lower glycemic index food choices, portion control, or brief behavioral strategies) prior to eating in an effort to minimize the postprandial GE. This is clinically meaningful, as greater glycemic excursions and variability have been associated with increased oxidative stress and adverse metabolic outcomes [6,31]. In this context, hunger may function as an anticipatory behavioral cue that precedes metabolic responses, distinguishing it from physiological triggers that occur closer to or after glucose changes.

Although some prior research has explored associations between mood, stress, or energy levels and glycemic fluctuations [12,19], few have explicitly examined hunger as a time-sensitive predictor. These results thus extend earlier findings by highlighting hunger as an immediate and time-sensitive behavioral cue that may precede glycemic shifts, offering insight into real-world eating patterns [14]. The observed inverse association between concurrent hunger and glucose levels is consistent with prior physiological and behavioral research suggesting that hunger reflects a state of relative energy deficit. However, this study extends this literature by examining these relationships using high-frequency, real-time data in naturalistic settings, allowing for the characterization of within-person temporal dynamics and the prediction of subsequent GEs.

Compared to prior studies that examined single-point associations or broad daily averages [14,32], this study’s high-resolution, time-aligned modeling [33] offers a more granular view of within-day glucose dynamics. Unlike studies that focused solely on dietary records or controlled feeding conditions [31,34], this analysis highlights hunger as a proximal, low-burden behavioral marker for anticipating glycemic instability in everyday settings.

Theoretically, this study advances a paradigm that conceptualizes metabolic regulation as a dynamic, bidirectional process in which subjective behavioral states may act as proximal indicators of behaviors (eg, eating) that contribute to downstream metabolic changes [20]. This perspective moves beyond traditional models that treat glucose control as an automatic or isolated process, emphasizing instead the reciprocal interplay between behavior and physiology. This conceptualization aligns with dual-process models of appetite regulation, which highlight the integration of homeostatic and hedonic signals in guiding eating behavior and metabolic outcomes [35,36].

Several limitations in this study should be noted. Hunger was assessed using a single binary item, which may not fully capture the complexity of appetite or eating drive. Furthermore, hunger was analyzed as an isolated behavioral signal, although it is likely influenced by unmeasured contextual factors such as stress, emotional state, and the surrounding food environment. Individual differences in interoceptive awareness and self-report accuracy could also influence results. The sample consisted primarily of women and may not generalize to other populations. Given that some sex-related differences in glucose levels were observed, future studies with more balanced sex representation are needed to clarify whether sex meaningfully moderates these associations. Importantly, food intake was not measured, limiting our ability to directly examine the relationship between reported hunger and subsequent eating behavior, especially when eating can occur without the presence of hunger [37]. Therefore, the observed association between hunger and subsequent GEs is likely mediated by unmeasured eating behavior rather than reflecting a direct physiological effect of hunger itself. Additionally, repeated EMA prompts may have induced reactivity, potentially influencing participants’ awareness of hunger and their self-reports. Finally, the models did not include an explicit residual correlation structure for repeated measures over time. As a result, temporal autocorrelation may not have been fully accounted for. Despite these limitations, this study offers novel insight into the dynamic interplay between behavioral states and glucose regulation in daily life and underscores the importance of considering subjective experience in metabolic research. Future studies should consider multidimensional hunger assessments, incorporate dietary assessment, and include more diverse samples to improve generalizability, and explore how such momentary signals can be operationalized to guide real-time interventions that support metabolic health across diverse populations.

In conclusion, this study highlights hunger as a dynamic and actionable behavioral signal that bridges subjective experience and physiological response. By integrating real-time hunger assessments with CGM data, these findings provide a foundation for enhancing personalized strategies to reduce metabolic risk in overweight populations.