Later energy intake relative to mathematically-modeled circadian time is associated with higher percentage body fat

Abstract

Objective

Later circadian timing of energy intake is associated with higher body fat percentage. Current methods for obtaining accurate circadian timing are labor and cost intensive, limiting practical application of this relationship. We investigated whether the timing of energy intake relative to a mathematically-modeled circadian time, derived from easily-collected ambulatory data, would differ between participants with a lean or overweight/obesity body fat percentage.

Methods

Participants (n=87) wore a light and activity measuring device (actigraph) throughout a cross-sectional 30-day study. For 7 consecutive days within these 30 days, participants used a time-stamped-picture phone application to record energy intake. Body fat percentage was recorded. Circadian time was defined using melatonin onset from in-laboratory collected repeat-saliva sampling or using light and activity or activity data alone entered into a mathematical model.

Results

Participants with overweight/obesity body fat percentages ate 50% of their daily calories significantly closer to model-predicted melatonin onset from light and activity data (0.61h closer) or activity data alone (0.86h closer; both log-rank p<0.05).

Conclusion

Use of mathematically-modeled circadian timing resulted in similar relationships between the timing of energy intake and body composition as that observed using in-laboratory collected metrics. These findings may facilitate use of circadian timing in time-based interventions.

Introduction

Synchronization of behaviors (e.g., eating, sleeping) with internal circadian timing is essential for optimal functioning of almost every physiological process.^1,2 People, however, may choose or be required to adopt schedules and behavioral patterns that are misaligned with their internal circadian timing system, which can increase risk of adverse cardiometabolic outcomes.^3,4

Later timing of energy intake (i.e., calories), independent of other factors, is associated with higher body mass.^5,6 Restricting calories to earlier daytime hours appears to promote healthy weight, glucose tolerance, blood pressure, and lipid levels,^7,8 suggesting therapeutic benefit of developing interventions targeting the timing of eating. Interventions that restrict the timing of energy intake based on clock time may not be effective for all individuals, however, as large inter-individual differences exist in circadian timing.⁹ The gold-standard biomarker for circadian timing is dim-light melatonin onset (DLMO), which represents the start of the biological night. Energy intake relative to DLMO is associated with higher body fat.^10,11 Thus, knowing an individual’s circadian timing is likely necessary to optimize time-based eating interventions. DLMO, however, is difficult to measure, as it requires time- and resource-intensive methods, and results are usually not available for weeks.

The primary objective of the current study was to determine if the timing of caloric midpoint (time at which 50% of daily calories are consumed)¹⁰ relative to a model-predicted DLMO (pDLMO) derived from actigraphy data (Brown et al., model¹²) would significantly differ between participants with lean or overweight/obesity body fat percentage. We also examined the relationship between body fat and sleep/wakefulness timing to identify if energy intake relative to these behaviors would suffice. Identification of relationships among timing of energy intake, pDLMO, and percentage body fat would represent a metric that could be utilized in the development of future therapies to reduce body weight and/or decrease body fat.

Methods

Data were collected as part of a 30-day study of relationships among daily behaviors, sleep, and circadian timing; inclusion criteria and primary outcomes are published.^10,11,13 Written informed consent was obtained from participants (n=87, 36 females; aged (average±stdev)19.0±1.1 years; body fat percentage 22.2±9.5%; body mass index 22.9±3.7 kg/m²). Partners Healthcare and Massachusetts Institute of Technology Institutional Review Boards approved the protocol.

Caloric Timing

Participants were instructed to record all food and beverages that they consumed for 7 consecutive days using the time-stamped photograph phone application MealLogger™ (Wellness Foundry, New York). When recording, participants added descriptions to the photograph (e.g., dressings, condiments) and included an item of known size (e.g., a spoon) to help accurately assess nutrient content and portion-size. If clarification was required, staff communicated with participants within 24 hours. If ≤2 meals were denoted for a day, participants were asked to confirm that everything was documented. Participants were also asked to confirm if all intake was recorded at the end of the 7-day assessment. Energy intake was scored separately by two nutritionists; discrepancies were resolved by a third nutritionist. Caloric content was calculated using the University of Minnesota Nutrition Data System for Research software,^14,15 with calories consumed <15-minutes apart combined into one event;¹⁶ ≥4 days of recording were required to be included in analysis.

DLMO, Sleep Timing, pDLMO, and body fat percentage

For one evening during the 30-day protocol, participants visited the laboratory to collect saliva under dim-light conditions to be assayed for melatonin, and to collect body fat percentage via bio-electrical impedance (detailed previously^10,11,13).

Throughout the protocol, participants wore an actigraph (Motionlogger, AMI, Ardsley, NY) to collect minute-by-minute light levels and activity counts (zero-crossing mode). Sleep onset and offset were determined using actigraph data.^10,11,13 To calculate pDLMO for the day of DLMO collection, data were removed if any minute-by-minute data were not recorded/transmitted by the actigraph during the 24-hour episode preceding the DLMO collection (n=17) or if the summed total light or activity recordings for that 24-hour episode deviated >3 standard deviations (n=6) from the mean of our previously published dataset^10,11.These criteria excluded n=23, leaving n=87 for analysis. Actigraphy data were averaged into 30-minute bins to be used as inputs to a published model¹² in Python that uses a neural network to classify a time series of data as ending before or after DLMO. Based on these predictions, the model interpolates the time of DLMO. For this analysis, we used the neural network architecture containing two hidden layers with dropout for two different models: Model 1 (light and activity) and Model 2 (activity). Participants were divided into an 80:20 train:test split to fit the weights of the model, using Adam optimization as detailed previously.¹²

Analysis

For each participant we: (i) used the fitted model to calculate pDLMO for each participant for the day of DLMO collection; (ii) calculated the median sleep and waking times during the week of food collection; and (iii) classified participants with lean or overweight/obesity body fat percentages, using body fat percentage cutoffs of 20% for men and 30% for women, according to published thresholds.¹⁷ Bland-Altman plots, correlations, and root mean square error (RMSE) measured agreement between DLMO and the pDLMO models. We used independent Student’s t-tests to compare DLMO and pDLMO between groups, Log-Rank tests to compare Kaplan-Meier survival curves for DLMO, pDLMO models, and median sleep times, and Pearson correlation for the relationship between caloric midpoint and body fat. Statistics were performed using GraphPad Prism-9.2 (GraphPad Software, San Diego, CA).

Results

There was an ~8-h variation in the timing of DLMO across the study sample. pDLMO from both Model 1 and Model 2 significantly correlated with DLMO with high agreements (Figure 1) and a RMSE of 01:53h for Model 1 and 02:02h for Model 2 was measured (Figure 1). Participants with lean (n=55) or overweight/obesity (n=32) body fat percentage had no significant differences in clock time of DLMO, pDLMO, and energy timing or intake (all p>0.23, Table 1).

Figure 1.

Comparison of predicted dim-light melatonin onset (pDLMO) with in-laboratory collected DLMO (gold standard for circadian phase timing). The top row reports Bland-Altman and correlation plots of (A) Model 1 (n=87) and the bottom row (B) Model 2 (n=87) of the differences between predictions and DLMO. Participants with lean body fat percentage are denoted with black circles and participants with overweight/obesity body fat percentage are denoted with gray circles. Dashed lines for Bland-Altman plots depict the median and 95% limits of agreement and the dashed line for correlation plots depict the line of unity. The solid line depicts correlation between DLMO and pDLMO and p-values are derived from Pearson correlation.

Table 1:

Comparison of body fat, melatonin, energy intake, and the time to pass caloric midpoint (average time at which 50% of daily calories were consumed) relative to different outcome variables in individuals with a lean or overweight/obesity body composition.”

	Lean (n=55)	Overweight/obesity (n=32)	p-value
Body Fat
All (%, SEM)	16.5 (0.8)	32.0 (1.1)	p<0.0001
Male (%, SEM)	14.5 (0.6)	25.2 (1.0)	p<0.0001
Female (%, SEM)	24.0 (1.2)	33.8 (1.0)	p<0.001
Melatonin
In-Laboratory DLMO (hh:mm, SEM)	23:25 (00:13)	22:59 (00:17)	p=0.23
Model 1: Light and Activity pDLMO (hh:mm, SEM)	22:55 (00:19)	23:06 (00:17)	p=0.69
Model 2: Activity pDLMO (hh:mm, SEM)	22:55 (00:20)	22:51 (00:17)	p=0.90
Food Timing and Energy Intake
Caloric Midpoint Clock Time (hh:mm, SEM)	16:08 (00:20)	16:41 (00:23)	p=0.30
Caloric Midpoint Relative to DLMO (h, SEM)	−7.4 (0.3)	−6.2 (0.3)	p<0.01
Caloric Midpoint Relative to Light and Activity pDLMO (h, SEM)	−6.9 (0.3)	−6.3 (0.3)	p=0.23
Caloric Midpoint Relative to Activity pDLMO (h, SEM	−6.9 (0.4)	−6.1 (0.4)	p=0.12
Daily Energy Intake (kcal, SEM)	1672.3 (68.4)	1658.7 (84.0)	p=0.90
Kaplan-Meier Survival Log-Rank tests
Clock time of caloric midpoint (HR, 95% CI)	0.8 (0.5 to 1.2)		p=0.32
In-Laboratory DLMO (HR, 95% CI)	2.4 (1.4 to 4.0)		p=0.001
Sleep Onset (HR, 95% CI)	1.3 (0.8 to 2.1)		p=0.25
Sleep Offset (HR, 95% CI)	0.8 (0.5 to 1.2)		p=0.22
Model 1: Light and Activity pDLMO (HR, 95% CI)	1.8 (1.1 to 2.9)		p=0.02
Model 2: Activity pDLMO (HR, 95% CI)	1.6 (1.0 to 2.7)		p=0.045

SEM, standard error of the mean; DLMO, dim-light melatonin onset; pDLMO, predicted dim-light melatonin onset; HR, Hazard Ratio; CI, Confidence Interval

When caloric midpoint was referenced to DLMO using Log-Rank tests, participants with overweight/obesity body fat percentage passed their caloric midpoint 1.2h closer to DLMO than participants with lean body fat percentage (p<0.01, Table 1). When caloric midpoint was referenced to clock timing, however, there was no significant difference between groups (p=0.32, Figure 2A, Table 1), similar to results previously reported.¹⁰ There were no significant differences between groups when caloric midpoint was referenced to median sleep or wake times (both p>0.22, Table 1). When caloric midpoint was referenced relative to the pDLMO model output, participants with overweight/obesity body fat percentage passed their caloric midpoint closer to pDLMOs both when using Model 1 and Model 2 than individuals with lean body fat (0.61h and 0.86h closer, respectively; both p<0.05, Figure 2B, C). A later clock time of caloric midpoint and a caloric midpoint closer to pDLMO were associated with a higher body fat percentage (p<0.01, Figure 2D–F)

Figure 2.

Kaplan-Meier survival curves of lean (black line) and overweight/obesity (gray dashed line) participants’ time to pass caloric midpoint (average time at which 50% of daily calories were consumed) relative to clock time and predicted dim-light melatonin onset (pDLMO) (panels A-C) and the relationship between clock timing, pDLMO, and body fat percentage (panels D-F). Model 1 (lean n=55, overweight/obesity n=32) consists of a model using light and activity data and Model 2 (lean n=55, overweight/obesity n=32) consists of a model using activity data only. P values from survival curves were derived from log-rank tests (left column) and from Pearson correlation (right column).

Discussion

The timing of energy intake represents a potentially modifiable behavior to combat obesity and improve health. The current highly intensive process of obtaining accurate circadian timing, however, places barriers to using circadian markers in clinical settings. We found that using mathematical models to calculate pDLMO for circadian timing using actigraphy with or without light data resulted in identifying differing eating patterns in individuals who had lean versus overweight/obesity body fat percentages. Specifically, we found that individuals with overweight/obesity body fat percentages consumed their calories closer to pDLMO, similar to findings obtained using the gold-standard, in-laboratory collected DLMO.^10,11 Importantly, we did not find significant relationships between body fat percentage and the timing of energy intake relative to sleep onset and offset, highlighting the need for estimates of circadian biomarkers rather than comparison to daily sleep behaviors. These data suggest that easily attainable ambulatory data may be used to estimate circadian timing, and that the relationship of eating behaviors with these estimates may be physiologically meaningful and thus clinically useful.

We did not find significant differences between groups in clock timing of energy intake, but did find significant differences when analyzed relative to a physiological measure of circadian time (DLMO) as in previous reports from this dataset.^10,11 Using DLMO is burdensome, expensive, and does not allow for circadian timing to be measured in real-time. Due to the relative ease of collecting ambulatory wearable data, many mathematical models have been proposed using actigraphy as input to calculate a pDLMO with a range of accuracies.¹⁸ We applied the Brown et al. model¹² because it predicts phase well in people living at home, although models under more tightly-controlled conditions yield greater accuracy.^18,19 An open question is how accurate such methods need to be for clinical utility, which is likely application-dependent; here, we showed that for the purposes of recommending an eating schedule at the group-level (i.e., suggest shifting caloric midpoint relative to pDLMO earlier), the accuracy produced by these models may be sufficient. Other applications (i.e., shifting circadian phase¹⁸ or individual recommendations) may require more precise predictions. Our findings do suggest that modeling may be a viable alternative to assessing circadian phase for ‘chronotherapies’.²⁰

There are several limitations to consider. While both models were successful in estimating DLMO in populations living on daytime schedules, they are less successful when circadian timing and behaviors are misaligned,¹² such as would occur in a shift-working population. Furthermore, there are tradeoffs between collecting multiple modalities and the accuracy that those modalities provide, as observed in the RMSE for each model.

Conclusion

Examining the timing of energy intake relative to mathematically-modeled pDLMO resulted in a similar relationship between participants who were lean and overweight/obesity as that observed using in-laboratory collected data. These findings may enable clinical therapeutics to consider circadian timing in time-based eating interventions by using easily collected ambulatory data.

What is already known about this subject?

• Later circadian timing of eating is associated with higher body mass index and body fat percentage in younger people.

• Current methods for accurate measurement of circadian timing are expensive; require the participant or facility to create special low-light level conditions and then collect, store, ship, and assay multiple samples for melatonin concentration; are usually measured within a single day; and are not available in near-real time.

What are the new findings in your manuscript?

• Activity and light data collected from wrist-worn devices that are non-invasive, require minimal participant effort, and are used for multiple days, can be analyzed using mathematical models to predict circadian time of an individual.

• Results from our mathematical model-based calculations of circadian phase approximate those using the “gold-standard” in-laboratory melatonin onset; namely, individuals with overweight/obesity body fat percentages eat closer to a marker of circadian time than those with lean body fat percentages. This corresponds to eating later in the biological day.

How might your results change the direction of research or the focus of clinical practice?

• These findings may enable clinicians and researchers to include circadian timing when implementing time-based eating (i.e., circadian-based time-restricted eating) and other interventions by applying mathematical models that use easily collected data. This approach could potentially increase effectiveness, compliance, and/or scalability of interventions.