Imprecision nutrition? Intraindividual variability of glucose responses to duplicate presented meals in adults without diabetes

Abstract

Background

Continuous glucose monitors (CGMs) are used to characterize postprandial glucose responses and provide personalized dietary advice to minimize glucose excursions. The efficacy of such advice depends on reliable glucose responses.

Objectives

To explore within-subject variability of CGM responses to duplicate presented meals in an inpatient setting.

Methods

CGM data were collected from two inpatient feeding studies in 30 participants without diabetes, capturing 1189 responses to duplicate meals presented ∼1 wk apart from four dietary patterns. One study used two different CGMs (Abbott Freestyle Libre Pro and Dexcom G4 Platinum) whereas the other study used only Dexcom. We calculated the incremental area under the curve (iAUC) for glucose for each 2-h postmeal period and compared within-subject, within-CGM responses to duplicate presented meals using linear correlations, intra-class correlation coefficients (ICC), and Bland–Altman analyses. Individual variability of interstitial glucose responses to duplicate meals were also compared with different meals using standard deviations (SDs).

Results

There were weak-to-moderate positive linear correlations between within-subject iAUCs for duplicate meals [Abbott r = 0.46, 95% confidence interval (CI): 0.38, 0.54, P < 0.0001 and Dexcom r = 0.45, 95% CI: 0.39, 0.50, P < 0.0001], with low within-participant reliability indicated by ICC (Abbott 0.28, Dexcom 0.17). Bland–Altman analyses indicated wide limits of agreement (LoA) (Abbott −29.8 to 28.4 mg/dL and Dexcom −29.4 to 32.1 mg/dL) but small bias of mean iAUCs for duplicate meals (Abbott −0.7 mg/dL and Dexcom 1.3 mg/dL). The individual variability of interstitial glucose responses to duplicate meals was similar to that of different meals evaluated each diet week for both Abbott [SD_week1 11.7 mg/dL (compared with duplicate P = 0.01), SD_week2 10.6 mg/dL (P = 0.43), and SD_duplicate 10.1 mg/dL] and Dexcom [SD_week1 10.9 mg/dL (P = 0.62), SD_week2 11.0 mg/dL (P = 0.73), and SD_duplicate 11.2 mg/dL].

Conclusions

Individual postprandial CGM responses to duplicate meals were highly variable in adults without diabetes. Personalized diet advice on the basis of CGM measurements requires more reliable methods involving aggregated repeated measurements.

This trial was registered at clinicaltrials.gov as NCT03407053 and NCT03878108.

Introduction

Postprandial interstitial glucose responses as measured by continuous glucose monitors (CGM) are highly variable between individuals without diabetes. Not only do people exhibit variable glucose excursions after eating the same food, the relative ranking of foods in their ability to increase postprandial glucose can vary between people [1,2]. Such observations provide the rationale for personalizing diet advice to minimize glucose excursions by attempting to prescribe meals that result in reliably low postprandial glucose in each person [2,3]. A fundamental assumption of such precision dietary advice is that glucose responses to the same meal on repeated occasions within an individual are much less variable than their responses to different meals. However, this assumption has not been rigorously tested.

We aimed to investigate the reliability of within-subject postprandial CGM responses to duplicate presented meals consumed ∼1 wk apart by participants residing at the NIH Clinical Center Metabolic Clinical Research Unit during two inpatient feeding studies whose primary results on ad libitum energy intake have been previously reported [4,5].

Methods

Experimental design

We performed an exploratory analysis of data from two clinical research protocols completed at the NIH Clinical Center in Bethesda, MD, approved by the institutional review board of the National Institute of Diabetes and Digestive and Kidney Diseases and are registered at clinicaltrials.gov (NCT03407053 and NCT03878108). For NCT03407053, volunteers were recruited through the NIH Office of Patient Recruitment (OPR) beginning 31 January, 2018. Screening and enrollment began 1 March, 2018, and the last participant was discharged on 19 November, 2018. For NCT03878108, volunteers were recruited through the NIH OPR beginning 28 February, 2019. Screening and enrollment began 15 April, 2019, and the last participant was discharged on 4 March, 2020. For both studies, the randomization of diet order was conducted by the NIH Clinical Center Nutrition Department. Participants provided written informed consent and eligibility criteria for both studies were 1) ages 18–50 y; 2) BMI >18.5 kg/m²; and 3) weight stable (<5% change in past 6 mo).

Both studies were within-subject, random-order crossover designs where participants were exposed to two diets for 14 d each on 7-d rotating menus, consuming each meal twice (once during week 1 and once during week 2). Duplicate presented meals were included from four distinct dietary patterns. One was a minimally processed plant-based low-fat diet, and another was a minimally processed animal-based very-low carbohydrate ketogenic diet. The remaining two patterns had moderate macronutrient compositions, but one was high in ultraprocessed foods, whereas the other was rich in minimally processed foods. This enabled a comparison of ≤21 duplicate presented meals within each of the four dietary patterns (≤42 comparisons if participants completed one study, ≤84 if participants completed both studies). Participant flowcharts by CGM monitor are displayed in Supplemental Figure 1. The daily menus had three meals (breakfast, lunch, and dinner), where the order within day was always fixed and the time of day when each meal was consumed was similar between the first and second weeks. Furthermore, the sequence of the daily menu was the same on the first and second weeks except in cases when the respiratory chamber day (whose menu was fixed within each diet pattern) had to be scheduled on a different day because of availability.

Meals were provided to participants alone in their inpatient rooms and photographs of the meals were published previously [4,5]. Participants were instructed to eat as much or as little food as they wanted and asked to not intentionally change their weight throughout the study. All foods were weighed to the nearest 0.1 g before and after consumption, and energy intake was calculated using ProNutra software (v.3.4, Viocare). A limitation of these studies with respect to our analyses of postmeal glucose responses is that they included bottled water and snacks available throughout the day, but the timing of their consumption was not recorded.

Interstitial glucose concentrations were obtained from two brands of monitor: Abbott Freestyle Libre Pro (Abbott) and Dexcom G4 Platinum (Dexcom). In NCT03407053, some participants wore both Abbott and Dexcom, and in NCT03878108 study participants wore Dexcom. The Abbott device records glucose every 15 min and the Dexcom every 5 min. For accurate postprandial analysis, only duplicate meals with measured start time and sufficient data availability were included. Five participants completed both studies. For Abbott, the mean (range) of duplicate meals within-participant was 28 (12–34) from 14 participants providing 398 total comparisons and for Dexcom, the mean (range) of duplicate meals within-participant was 26 (5–61) from 30 participants providing 791 total comparisons. Data were aligned to the nearest 15-min (Abbott) or 5-min (Dexcom) CGM reading for the calculation of postmeal responses. Baseline glucose was assigned as the first time-point after the meal was provided. We also investigated using a baseline calculated as the average glucose measured over the 30 min before meal. The 2-h postprandial glucose incremental area under the curve (iAUC) was calculated for each meal using the trapezoid method, with dips below baseline assigned a negative value for iAUC (a calculation consistent with previous studies and similar to “netAUC” from [6]). Values of iAUC were reported as time-averaged glucose concentrations across the 120-min postprandial period (for example, iAUC/120 min).

Statistical methods

Statistical analyses and data visualization were performed in R (v4.2.3), GraphPad Prism (v9.5.0), and SAS (v9.4). Standard major axis regression was used to plot trends between meals 1 and 2 using lmodel2 in R. Simple linear correlation was calculated using Pearson’s correlation coefficient (r), with <0.4 interpreted as weak, 0.4 to 0.8 interpreted as moderate, and >0.8 interpreted as strong correlation. Repeatability was estimated by calculating the intra-class correlation coefficient (ICC) for glucose iAUCs. ICC was calculated using the following formula: ICC = participant variance / (participant variance + residual variance), which was generated from a linear mixed-effects model with participant and residual error as random effects and meal and eating occasion as fixed effects [7]. ICC values below 0.5 are considered as indicating poor reliability between measures [8]. Bland–Altman analyses were conducted accounting for multiple observations per individual [9].

To examine individual interstitial glucose variability in response to duplicate meals as compared with different meals, we partitioned the total iAUC variance between diet pattern (diet; ultra-processed, minimally processed, low-carbohydrate, and low-fat), meal type (meal; breakfast, lunch, and dinner), menu day (menu; days 1–7 on rotating menu), and duplicate meals (duplicate) between successive weeks to compute the standard deviation (SD) of duplicate iAUC responses as follows:

$${\text{SD}}_{\text{duplicate}}=\sqrt{\left[\frac{{\sum }_{\text{diet}}{\sum }_{\text{meal}}{\sum }_{\text{menu}}{\sum }_{\text{duplicate}}{\left({\text{iAUC}}_{\text{diet},\text{meal},\text{menu},\text{duplicate}}-⟨{\text{iAUC}}_{\text{diet},\text{meal},\text{menu}}⟩\right)}^2}{\text{df}}\right],}$$

where df is the degrees of freedom, subscripts indicate that the calculations are restricted to within the same diet, meal type, and menu day, and <iAUC_{diet,meal,menu}> is the average iAUC of the duplicate meals. Similarly, we computed individual SDs of the iAUC responses to different meals on each week of 7-d rotating menus as follows:

$${\text{SD}}_{\text{week}}=\sqrt{\left[\frac{{\sum }_{\text{diet}}{\sum }_{\text{meal}}{\sum }_{\text{menu}}{\left({\text{iAUC}}_{\text{diet},\text{meal},\text{menu}}-⟨{\text{iAUC}}_{\text{diet},\text{meal}}⟩\right)}^2}{\text{df}}\right]}$$

where df is the degrees of freedom and <iAUC_diet,meal> is the average iAUC over the 7-d wk of menus. The SD of energy intake was calculated in the same way as iAUC. One-way ANOVA was used to compare SDs of week 1 with duplicate meal responses, and SDs of week 2 with duplicate meal responses. Energy intake of meals in weeks 1 and 2 were compared using a paired t test with significance was accepted as P ≤ 0.05.

During the inpatient stay of each study, venous glucose measurements were obtained during oral glucose tolerance tests (OGTTs) in both studies and mixed meal tolerance tests (MMTTs) in NCT03878108 (Dexcom only) while wearing CGMs. The time intervals for OGTTs and MMTTs were 0, 10, 20, 30, 60, 90, and 120 min. This allowed for an assessment of individual variability between simultaneously measured venous and CGM determined iAUCs in response to OGTTs and MMTTs to provide an index of how much of the CGM iAUC variability in response to duplicate meals might be because of the variability in the iAUC determined by CGM as compared with simultaneous venous measurements. So, we evaluated the SD of the difference between CGM and venous iAUCs in response to the same meal test and compared this with the SD of the duplicate CGM iAUC responses.

To potentially identify predictors of individual variability in postprandial glucose response to duplicate meals, we used forward stepwise linear regression to estimate the contribution of measured behavioral variables using the stepAIC function of the R package `MASS`. The response variable was difference between duplicate meal postprandial glucose responses assessed as either iAUC (mg/dL) or total area under the curve (tAUC; mg/dL). Predictor variables included differences in baseline interstitial glucose (mg/dL; for iAUC only), difference in time taken to consume the meals (minutes), number of days between meals (days), difference in energy intake from snack consumption (kcal), type of meal (breakfast, lunch, or dinner), difference in consumed meal-specific macronutrients energy (protein, fat, and carbohydrate in kcal), difference in exercise sessions ≤30 min before the start of a meal (n), and difference in exercise sessions during the 2-h postprandial period (n). During the study, participants were instructed to complete three 20-min light-to-moderate intensity exercise sessions on a bicycle ergometer with standardized wattage and speed (between 30 and 40% heart rate reserve). Meal duplicates with inaccurate meal or exercise timing were omitted from the regression analyses.

Results

Within-subject glucose responses to the same presented meals eaten on separate occasions were highly variable

We investigated 30 participants whose characteristics are shown in Table 1, who were presented with duplicate meals on two consecutive weeks exactly 7 d apart for 85% of Abbott measurements and 81% of Dexcom measurements. The macronutrient composition of consumed duplicate meals is presented in Supplemental Figure 2. The energy and macronutrient intake of consumed meals for Abbott is displayed in Supplemental Table 1 and for Dexcom is displayed in Supplemental Table 2. Figure 1A plots the iAUC responses to meals consumed on week 2 compared with the duplicate meals on week 1 measured in the same participants using the Abbott device. Figure 1B plots analogous data obtained using the Dexcom device. For both devices, there were weak-to-moderate positive linear correlations between the within-subject iAUC responses to duplicate meals across all dietary patterns [Abbott r = 0.46, 95% confidence interval (CI): 0.38, 0.54, P < 0.0001 and Dexcom r = 0.45, 95% CI: 0.39, 0.50, P < 0.0001]. Linear correlations were similar when meals were split into breakfast, lunch, and dinner [mean (95% CI) Abbott breakfast r = 0.35 (0.20, 0.49), lunch r = 0.52 (0.38, 0.64), dinner r = 0.48 (0.34, 0.60); Dexcom breakfast r = 0.48 (0.38, 0.58), lunch r = 0.44 (0.33, 0.53), dinner r = 0.40 (0.30, 0.49), all P < 0.0001]. The mean iAUC responses to breakfast, lunch, and dinner meals across each dietary pattern are displayed in Supplemental Table 3.

TABLE 1.

Baseline characteristics of participants included in analyses.

	Abbott All (n = 14)	Abbott Female (n = 8)	Abbott Male (n = 6)	Dexcom All (n = 30)	Dexcom Female (n = 15)	Dexcom Male (n = 15)
Age (y)	31 ± 8	30 ± 8	32 ± 8	30 ± 7	30 ± 7	31 ± 7
Body mass (kg)	73.0 ± 13.3	73.1 ± 14.8	72.7 ± 12.4	80.6 ± 19.4	77.3 ± 21.9	83.9 ± 16.6
BMI (kg/m2)	25.5 ± 5.2	27.1 ± 6.0	23.4 ± 3.5	27.6 ± 6.3	28.2 ± 7.2	27.0 ± 5.5
Body fat (%)	28.9 ± 11.1	35.2 ± 9.7	20.5 ± 6.5	31.9 ± 9.9	37.8 ± 8.3	25.9 ± 7.6

Data are mean ± SD.

FIGURE 1.

Correlation and agreement of postprandial interstitial glucose responses to duplicate presented meals. (A) Correlation of iAUC of postprandial glucose responses to duplicate presented meals using Abbott. (B) Correlation of iAUC of postprandial glucose responses to duplicate presented meals using Dexcom (trendline for A and B is major axis regression with 95% CIs). (C) Bland–Altman plot of the iAUC differences between duplicate meals compared with the average of both measurements using Abbott. (D) Bland–Altman plot of the iAUC differences between duplicate meals compared with the average of both measurements using Dexcom (solid line indicates mean bias and dashed lines indicate 95% limits of agreement). Duplicate axes are presented for conversion between United States units (mg/dL) and Système international d'unités (mmol/L).

iAUC, incremental area under the curve; CI, confidence interval.

ICCs were 0.28 for Abbott and 0.17 for Dexcom, indicating that there was a low tendency for glucose responses to be similar in duplicate meals in the same participant. Across all duplicate meals and subjects, Bland–Altman plots are shown in Figures 1C and 1D indicating a low mean bias between iAUC responses to duplicate meals (Abbott −0.7 mg/dL and Dexcom 1.3 mg/dL), but there was a large variability indicated by the wide 95% LoA for both CGMs (Abbott −29.8 to 28.4 mg/dL and Dexcom −29.4 to 32.1 mg/dL).

Supplemental Figure 3A and B plots the differences in interstitial glucose responses to duplicate meals for each individual participant using the Abbott and Dexcom devices, respectively, and show highly variable interstitial glucose responses when the same participant consumed duplicate meals on separate weeks, regardless of the CGM device. However, the iAUC bias was relatively low for most participants when averaged across different duplicate meals. Supplemental Figure 3C and D plot the same data separated by individual duplicate meals as measured using the Abbott and Dexcom devices, respectively, and indicate highly variable individual interstitial glucose responses to duplicate meals, regardless of the CGM device. Nevertheless, the iAUC bias was relatively low for most meals when averaged across participants. None of the results in this section were meaningfully altered by using baseline glucose calculated as the mean value in the 30 min before each meal (Supplemental Table 4).

Similar within-subject variability of interstitial glucose excursions to eating duplicate compared with different presented meals

Surprisingly, we found that everyone’s interstitial glucose response variability to duplicate meals was similar to the variability in their interstitial glucose responses to different meals. Figure 2A plots the SD of the interstitial glucose responses in each individual participant to different meals eaten in either week 1 or week 2 along with the SD of their interstitial glucose responses to duplicate meals for the Abbott device. Figure 2B plots analogous data from the Dexcom device. Regardless of device, the variability in the interstitial glucose response to duplicate meals was similar to each participant’s interstitial glucose response variability to different meals [Abbott: SD_{week 1} = 11.7 mg/dL (compared with duplicate P = 0.01), SD_{week 2} = 10.6 mg/dL (P = 0.43), SD_duplicate = 10.1 mg/dL and Dexcom: SD_{week 1} = 10.9 mg/dL (P = 0.62), SD_{week 2} = 11.0 mg/dL (P = 0.73), SD_duplicate = 11.2 mg/dL].

FIGURE 2.

Mean and individual participant SD of postprandial glucose responses between different meals across week 1, different meals across week 2, and duplicate presented meals between weeks. (A) Abbott (n = 14), (B) Dexcom (n = 30). Duplicate y-axes are presented for conversion between United States units (mg/dL) and Système international d'unités (mmol/L). iAUC, incremental area under the curve; SD, standard deviation.

Comparing venous with CGM glucose excursions in response to oral glucose and mixed meal tolerance tests

Mean ± SEM iAUC responses to OGTTs were similar between venous (39.5 ± 3.8 mg/dL) and Abbott CGM (39.5 ± 3.9 mg/dL, P = 0.99; Figure 3A) and were moderately correlated (r = 0.73, P < 0.0001; Figure 3B). Mean ± SEM iAUC responses to OGTTs and diet-specific MMTTs were similar between venous (38.1 ± 2.8 mg/dL) and Dexcom CGM (36.8 ± 2.9 mg/dL, P = 0.53; Figure 3D) and were also moderately correlated (r = 0.72, P < 0.0001; Figure 3E). Note that interstitial glucose responses to very low carbohydrate meals were low for all participants. The average SD of within-subject differences between simultaneous venous and CGM measurements in response to meals provides an index of the imprecision of the CGM. For the Abbott device, the average SD of the iAUC difference between duplicate presented meals was not significantly different from the index of CGM precision (P = 0.28; Figure 3C). For Dexcom, the index of CGM precision was significantly greater than the average SD of the iAUC difference between duplicate presented meals (P = 0.01; Figure 3F). These results suggest that CGM imprecision may contribute to highly variable quantification of interstitial glucose responses to duplicate presented meals.

FIGURE 3.

Comparisons between postprandial venous and CGM responses to standardized meal tests. (A) Mean ± SEM and individual participant comparisons of venous iAUC responses to OGTTs with Abbott (n = 26). (B) Correlation between venous and Abbott iAUC (Pearson’s r). (C) Individual participant SD of postprandial glucose responses to duplicate presented meals using CGM (duplicates = 12-34) and venous compared with CGM comparisons (duplicates = 2) for Abbott. (D) Mean ± SEM and individual participant comparisons of venous iAUC responses to OGTTs and MMTT with Dexcom (n = 87). (E) Correlation between venous and Dexcom iAUC (Pearson’s r). (F) Individual participant SD of postprandial glucose responses to duplicate presented meals using CGM (duplicates = 8–61) and venous compared with CGM comparisons (duplicates = 2–6) for Abbott. Paired t-tests used for A,C,D,F. iAUC, incremental area under the curve; CGM, continuous glucose monitor; LC, low-carbohydrate meal tests; LF, low-fat meal tests; MMTT, mixed meal tolerance tests; OGTT, oral glucose tolerance test; SD, standard deviation.

Potential factors affecting variability of interstitial glucose responses to duplicate presented meals

We explored numerous factors that may have contributed to explaining the interstitial glucose response variability to duplicate meals. First, there was a moderate negative linear correlation between differences in baseline glucose and differences in iAUC [Abbott r = −0.46, 95% CI: −0.53 to −0.38, P < 0.0001 and Dexcom r = −0.50, 95% CI: −0.55 to −0.45, P < 0.0001], suggesting that baseline glucose concentrations may have contributed to the low repeatability of glucose iAUC. Repeating analyses with tAUC rather than incremental moderately increased ICC for both Abbott (0.57) and Dexcom (0.20).

Because food intake was ad libitum in our studies, we investigated whether differences in meal energy intake between duplicate meal weeks affected our findings. For Abbott, mean (95% CI) meal energy intake was 774 (743–804) kcal in week 1 and 743 (710–775) kcal in week 2 (P = 0.001). For Dexcom, mean (95% CI) meal energy intake was 788 (761–814) kcal in week 1 and 765 (739–790) kcal in week 2 (P = 0.002). Energy intake between duplicate meals was strongly positively correlated (Abbott r = 0.83, 95% CI: 0.79, 0.86, P < 0.0001, Dexcom r = 0.83, 95% CI: 0.81, 0.85, P < 0.0001) and there was a weak positive correlation between differences in energy intake and differences in glucose iAUC (Abbott r = 0.22, 95% CI: 0.12, 0.31, P < 0.0001 and Dexcom r = 0.09, 95% CI: 0.02, 0.16, P = 0.009). However, repeating our analyses using only duplicate meals where energy intake was within 100 kcal between meals did not materially affect our results regarding the iAUC correlations (Abbott n = 203, r = 0.44, 95% CI: 0.32, 0.54, P < 0.0001 and Dexcom n = 376, r = 0.40, 95% CI: 0.31, 0.48, P < 0.0001), or ICC (Abbott 0.28 and Dexcom 0.18), or Bland–Altman analyses (Abbott bias −0.8 mg/dL, LoA −29.4 to 27.8 mg/dL and Dexcom bias 2.2 mg/dL, LoA −32.0 to 36.5 mg/dL). Using only meals where energy intake was within 100 kcal, the variability in the interstitial glucose response to duplicate meals remained similar to each participant’s interstitial glucose response variability to different meals [Abbott: SD_{week 1} = 10.7 mg/dL (P = 0.29), SD_week2 = 10.4 mg/dL (P = 0.48), SD_duplicate = 9.8 mg/dL and Dexcom: SD_week1 = 12.0 mg/dL (P = 0.87), SD_{week 2} = 10.2 mg/dL (P = 0.10), SD_duplicate = 11.8 mg/dL). Conversely, the variability in energy intake was lower for duplicate meals than response variability to different meals (Supplemental Figure 4).

In addition to the three daily meals provided, participants were also given snacks that could be consumed at any time of day. To examine whether our results may have been affected by differences in snack intake between days with duplicate meals, we filtered the data such that snack intake was <200 kcal on both duplicate meal days, resulting in 138 duplicates meals available for Abbott and 302 for Dexcom. For Abbott, mean (95% CI) meal energy intake was 795 (742–848) kcal in week 1 and 753 (700–807) kcal in week 2 (P = 0.01) and mean (95%) snack intake was 38 (28–48) kcal/d in week 1 and 31 (23–40) kcal/d in week 2 (P = 0.31). For Dexcom, mean (95% CI) meal energy intake was 715 (681–749) kcal in week 1 and 684 (649–718) kcal in week 2 (P = 0.002) and mean (95%) snack intake was 30 (23–36) kcal/d in week 1 and 22 (17–28) kcal/d in week 2 (P = 0.04). Repeating our analyses using only meals where snack intake was <200 kcal did not materially affect our results regarding the iAUC correlations (Abbott r = 0.48, 95% CI: 0.34, 0.60, P < 0.0001 and Dexcom r = 0.54, 95% CI: 0.45, 0.61, P < 0.0001), or ICC (Abbott 0.33 and Dexcom 0.26), or Bland–Altman analyses (Abbott bias −1.4 mg/dL, LoA −30.2 to 27.5 mg/dL and Dexcom bias 1.0 mg/dL, LoA −28.0 to 30.0 mg/dL). Using only meals where daily snack intake was <200 kcal, the variability in the interstitial glucose response to duplicate meals remained similar to each participant’s interstitial glucose response variability to different meals [Abbott: SD_{week 1} = 10.8 mg/dL (P = 0.49), SD_{week 2} = 10.0 mg/dL (P = 0.86), SD_duplicate = 9.7 mg/dL and Dexcom: SD_{week 1} = 10.2 mg/dL (P = 0.33), SD_{week 2} = 7.9 mg/dL (P = 0.55), SD_duplicate = 8.8 mg/dL]. None of the results in this section were meaningfully altered by using baseline glucose calculated as the mean value in the 30 min before each meal (Supplemental Table 4).

Explaining variance in response to duplicate meals using known behavioral variables

For Abbott, the difference in baseline glucose, carbohydrate content of meals, energy intake from consumption of snacks, and the presence of exercise during the postprandial period were identified as predictor variables for the difference in iAUC between duplicate meals (Table 2). For tAUC, predictor variables were the difference energy intake from snacks, the difference in carbohydrate content, and difference in time to consume meals (Table 2). For Dexcom, the difference in baseline glucose, difference in carbohydrate content of meals, the presence of exercise during the postprandial period, energy intake from consumption of snacks, and difference in time to consume meals were identified as predictor variables for the difference in iAUC between duplicate meals (Table 2). For tAUC, the presence of exercise during the postprandial period and the difference in carbohydrate content were identified as predictor variables (Table 2). Although these identified predictor variables significantly contributed to individual variability in postprandial glucose responses to duplicate meals, they explained only a small amount (<33%) of the variability as determined by the coefficients of determination (R²) values in the resulting linear regression models (Table 2).

TABLE 2.

Forward stepwise linear regression of iAUC and tAUC interstitial glucose responses to duplicate meals eaten on separate days using Abbott and Dexcom CGMs.

Abbott (n = 264)	Difference in iAUC (mg/dL)
Variable	Coefficient	Std. Error	P value	AIC
(Intercept)	−1.73	0.82	0.036	1458.21
Difference in baseline glucose (mg/dL)	−0.43	0.05	<0.001	1390.03
Difference in carbohydrate content (kcal)	0.04	0.01	<0.001	1372.54
Difference in snack energy intake (kcal)	−0.01	0.00	0.007	1367.77
Difference in postprandial exercise sessions (n)	−2.51	1.40	0.073	1366.49
R2	0.315
Abbott (n = 264)	Difference in tAUC (mg/dL)
Variable	Coefficient	Std. Error	P value	AIC
(Intercept)	−2.27	1.01	0.026	1488.23
Difference in snack energy intake (kcal)	−0.01	0.00	0.008	1483.35
Difference in carbohydrate content (kcal)	0.03	0.01	0.006	1478.18
Difference in time to consume meal (min)	−0.11	0.08	0.146	1478.03
R2	0.060
Dexcom (n = 598)	Difference in iAUC (mg/dL)
Variable	Coefficient	Std. Error	P value	AIC
(Intercept)	1.99	0.56	<0.001	3351.32
Difference in baseline glucose (mg/dL)	−0.42	0.03	<0.001	3159.67
Difference in carbohydrate content (kcal)	0.02	0.01	<0.001	3148.05
Difference in postprandial exercise sessions (n)	−3.36	0.96	0.001	3136.90
Difference in snack energy intake (kcal)	−0.00	0.00	0.022	3133.46
Difference in time to consume meal (min)	−0.07	0.05	0.156	3133.42
R2	0.317
Dexcom (n = 598)	Difference in tAUC (mg/dL)
Variable	Coefficient	Std. Error	P value	AIC
(Intercept)	2.96	1.08	0.006	3928.96
Difference in postprandial exercise sessions (n)	−6.63	1.84	<0.001	3918.82
Difference in Carbohydrate Content (kcal)	0.02	0.01	0.024	3915.69
R2	0.028

Abbreviation: AIC, Akaike information criterion; CGM, continuous glucose monitor; iAUC, incremental area under the curve; tAUC; total area under the curve.

Predictor variables included differences in baseline interstitial glucose (mg/dL; for iAUC only), difference in time taken to consume the meals (min), number of days between meals (d), difference in energy intake from snack consumption (kcal), type of meal (breakfast, lunch, or dinner), difference in consumed meal-specific macronutrients energy (protein, fat, and carbohydrate in kcal), difference in exercise sessions ≤30 min before the start of a meal (n), and difference in exercise sessions during the 2-h postprandial period (n).

Implications for meal ranking

Given the low reliability of iAUC responses, advice to eat meals having low postprandial glucose responses (that is, bottom tertile) on the basis of a single meal test does not necessarily result in low interstitial glucose excursions in response to the same meals in the future. Supplemental Figure 5 shows that meals in the bottom iAUC tertile on week 1 were 98% and 108% lower than the mean across all meals for Abbott and Dexcom, respectively. However, the same meals on week 2 were only 42% and 60% lower than average. Conversely, Supplemental Figure 5 also shows that advice to avoid meals with elevated iAUC, in week 1 shows the inverse. Meals in the upper iAUC tertile on week 1 were 85% and 131% higher than the mean across all meals for Abbott and Dexcom, respectively. However, the same meals on week 2 were only 45% and 53% higher than average, suggesting regression to the mean with repeated measures.

Discussion

CGM devices are becoming widely used in people without diabetes as part of commercial precision nutrition programs that provide personalized diet advice [10], and in March 2024 the United States Food and Drug Administration cleared the first over-the-counter CGM for individuals without diabetes who want to better understand how diet and exercise may impact interstitial glucose levels [11]. However, CGM responses need to be reliable to be useful [12]. A fundamental assumption of personalized or “precision” nutrition is that an individual’s responses to repeated meals are less variable than their responses to different meals. Otherwise, it would be impossible to provide reliable advice to avoid meals that result in undesirable responses. Previous work found relatively reliable postprandial CGM responses to a small number of duplicate simple meals like bread [2] or muffins [1], but such meals are not representative of multicomponent meals that are the focus of personalized dietary advice in the real-world. Surprisingly, our study found that the reliability of postprandial CGM responses to many duplicate multicomponent meals was poor and that the within-subject variability to duplicate meals was roughly as large as the variability across different meals. Perhaps this is why recent randomized trials comparing personalized nutrition interventions focused on interstitial glucose responses observed small effects for mean glucose (within 7 mg/dL, 0.39 mmol/L) and HbA1c (within 0.14%) [13], or no differences in glycemic variability and HbA1c [14] as compared with general diet advice.

We recently demonstrated that postprandial interstitial glucose responses using two different brands of CGMs simultaneously worn on different anatomical locations resulted in only moderate correlations of within-subject postprandial responses to simultaneously measured multicomponent meals (r = 0.68) and modest concordance of the meal rankings by iAUC (Kendal rank correlation = 0.43) [15]. A subsequent study using simple test meals (that is, muffins, milkshakes, and energy bars) confirmed that simultaneous within-subject postprandial iAUCs measured using different CGM devices were only moderately correlated (r = 0.61) but the rank order of these simple meals according to iAUC was more concordant (Kendall rank correlation = 0.68) than the rankings of multicomponent meals in our previous study, perhaps reflecting the formulation of simple test foods to have wide differences in glycemic load [16]. Interestingly, using identical CGMs in the same anatomical location resulted in much better agreement (r = 0.97; Kendall rank correlation = 0.87) suggesting that a given CGM device provides valid measures of postprandial interstitial glucose responses to simple test meals on a single occasion [16]. However, this does not address the reliability of within-subject responses to repeated meals.

The relatively low within-subject reliability of postprandial CGM responses to duplicate multicomponent meals in our study occurred under inpatient metabolic ward conditions where meal order within each day was standardized and was typically preceded by a previous standardized day. While less reflective of free-living conditions, such inpatient feeding studies reduce the amount of variability explained by behavioral factors, enabling better understanding of the amount of interstitial glucose variability that can be explained by ingestion of meals, providing a better indication of measurement error [17,18]. However, despite the strengths of our inpatient feeding design, our study had several limitations. First, the primary aims of the original studies were to measure ad libitum energy intake differences between dietary patterns; therefore, duplicate meals were not necessarily consumed in identical amounts, although energy intake of duplicate meals was highly correlated and repeating the analyses using only meals within 100 kcal of each other did not change interpretation. Furthermore, despite the regimented meal order and timing achieved with implementation of the 7-d rotating menus, snacks were available for consumption at any time of day, which may have differentially affected meal responses. Re-analysis using only duplicate meals on days when snack intake was below 200 kcal or when the energy intake difference between duplicate meals was <100 kcal did not materially affect our results. Another limitation of our study is that dietary interventions could have changed physiology in week 2 compared with week 1, but the Bland–Altman and participant level data do not indicate meaningful bias between weeks.

The two strongest predictor variables of postprandial glucose differences in response to duplicate meals were differences in baseline glucose and carbohydrates consumed. In addition, the impact of exercise on postprandial glucose responses was identified using both monitors but appeared to be of greater magnitude with the Dexcom. The presence of a 20-min bout of light-to-moderate intensity exercise during the postprandial period decreased glucose responses in the Dexcom, similar to previous findings [19]. The discrepancy between CGMs could be because of the anatomical location of the monitors in proximity to physically active tissue [20], and increased subcutaneous adipose tissue blood flow [21]. In our previous analysis, the discrepancy between Abbott (arm) and Dexcom (abdomen) was larger in individuals with higher body fat [15]. Regardless of CGM, the variation in postprandial glucose excursions to duplicate meals was only weakly correlated with the measured predictor variables thought to influence glucose responses in our regression model. This suggests it would be very difficult to explain the variance in CGM response between duplicate meals by collecting data on behavior in a free-living setting.

Repeating analyses with tAUC rather than iAUC modestly improved reliability for both Abbott and Dexcom. The moderate reliability of Abbott was similar to reliability previously reported for 4-h glucose tAUCs using venous samples from duplicate mixed meals, performed under standardized conditions twice within 28 d in adults with normal glucose tolerance [22], but ICC for Dexcom was still much lower when using tAUC.

Because of the inpatient setting, our study has limited generalizability to free-living people. However, free-living behaviors will likely further increase the within-subject variability of CGM responses to similar meals. A plethora of modifiable behavioral factors can also influence postprandial glycemic responses to the same meal within an individual and the reasons for the variable responses to repeated meals in our study are presently unknown. In our study, meals were ad libitum, but participants tended to eat similar amounts of the repeated meals and differences in energy intake did not seem to account for the differences in interstitial glucose response, as the variability was similar when only analyzing meals with similar energy intake within 100 kcal (data not shown). However, variations in the sequence of foods consumed within the repeated multicomponent ad libitum meals may have contributed to the variability because food sequence has been previously shown to result in varying glycemic responses in people with and without type 2 diabetes [[23], [24], [25], [26]]. Physical activity differences may have also played a role, as previous studies have shown that breaking up prolonged sitting with small amounts of physical activity during the postprandial period reduces postprandial glycemia [[27], [28], [29]], and even leg fidgeting may have an effect [30]. Sleep quality and bedtime has recently been associated with changes in CGM-derived measures of postprandial glucose [31], so variations in sleep quality may have contributed to differences in the studies presented. Importantly, if such behavioral factors are indeed important contributors to meal glycemic responses, then an enormous amount of data may be required to capture these behavioral determinants and reliably predict an individual’s glucose excursions and thereby provide personalized “precision” diet advice.

Participants wore CGMs during standardized meal tests (either OGTTs or diet-specific MMTTs) concurrent with venous blood measures to provide some indication of the contribution of CGM technical error to the observed variable postprandial responses, acknowledging there is a 5–6 min lag time for glucose moving from intravascular to interstitial compartments [32]. The correlation between venous and CGM was responses were similar to values recently reported, but we did not observe a mean difference between CGM and venous iAUCs, compared with the ∼20.5 mg/dL (∼1.14 mmol/L) mean difference that was reported recently [33]. The SD of venous to CGM responses were not different to duplicate meal SDs for Abbott and were higher for Dexcom, suggesting that the variance between CGM and venous measures was at least as high as the variance of duplicate meals within CGM. However, there were fewer pairs to calculate SD with the venous to CGM comparisons compared with the within CGM comparison of duplicate meals and perhaps more repeated measures are required to reliably compare the variance between CGM and venous measures within each participant.

It is worth noting that even venous glycemic iAUC responses to duplicate test meals exhibit poor reproducibility [34] and moderate correlation [35], which is exacerbated in individuals without diabetes, so the goal of personalization on the basis of individual meal glycemic responses may be difficult even with more standardized laboratory measurements.

Our participants were representative of a generally healthy population across a wide range of body mass indexes, but without diabetes or other metabolic disease. Recent cross-sectional evidence using Medtronic devices suggests day-to-day reproducibility of CGM readings is lower in younger individuals (<60 y) without prediabetes or type 2 diabetes [36]. Intriguingly, we found a low mean bias of within-subject iAUCs in response to multiple pairs of duplicate meals suggesting that it may be possible to reliably estimate within-subject postprandial responses to the same meals provided that enough repeated measurements are made. Identifying the number of repeated postprandial CGM measurements, in response to the same meals within-participants, that is required to provide reliable personalized estimates, is a critical question for future research. Our results suggest that two measurements are too few even under inpatient metabolic ward conditions.

In conclusion, our data suggest that personalized diet advice is unlikely to be reliable if it is based primarily on postprandial CGM measurements obtained using very few repeated measurements in adults without diabetes. Instead, precision nutrition requires more reliable methods involving aggregated repeated measurements.

Author contributions

The authors’ responsibilities were as follows – JG, KDH: conceptualized and designed the research, AH, JAO, KM, JG, KDH: analyzed and interpreted the data, and critically reviewed, drafted, and approved the final manuscript; and all authors: read and approved the final manuscript.

Funding

Supported by the Intramural Research Program of the National Institutes of Health, National Institute of Diabetes & Digestive & Kidney Diseases.

Data availability

Data described in the manuscript, code book, and analytic code will be made available upon request by email to the corresponding author.

Conflict of interest

The authors report no conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: