Food-specific morning and night intake scores are associated with impaired glucose tolerance during pregnancy

Abstract

Background:

Impaired glucose tolerance (IGT) during pregnancy is associated with numerous short and long-term adverse health outcomes. Diet is a key factor influencing glucose tolerance, yet there is little data on the relationship between specific foods or intake timing and IGT.

Objective:

We examined whether food category intakes and their timing are associated with IGT.

Methods:

We used data from the Temporal Research in Eating, Nutrition, and Diet during Pregnancy (TREND-P) study, which recruited 144 pregnant persons at gestational age <18 weeks. Participants provided up to 28 days of food records (before and after photos for each eating occasion and text notes), and access to their electronic health records (EHR). We introduced Morning and Night Intake Scores (weighted measures reflecting the proportion of intakes consumed during nocturnal or morning periods) that incorporate exact intake timing rather than using a single threshold. We used logistic regression to estimate associations (odds ratio, OR and 95% confidence intervals, CI) between temporal intake features and IGT (determined from EHR lab results), controlling for diet quality, age, total daily energy intake, physical activity, and pre-pregnancy BMI, and conducted sensitivity analysis.

Results:

We found a significant positive association between mean daily red meat intake and IGT (OR=3.68, 95%CI: 1.68, 8.75), and a negative association with IGT for night egg intake (OR=0.95, 95%CI: 0.93, 0.98). These associations remained after sensitivity analysis. Associations between macronutrient and energy intake and IGT were not statistically significant.

Conclusions:

Our novel approach to modeling intake timing uncovered a new negative association between night egg intake and IGT, and a positive association between red meat intake and IGT. Macronutrient and energy intake were not significant predictors of IGT, showing the importance of capturing specific food intake and timing. Future research is needed to determine if these observed associations are causally linked to IGT.

Introduction

Hyperglycemia is among the most common pregnancy complications (1). Gestational diabetes mellitus (GDM), glucose intolerance that develops during pregnancy, is linked to numerous short- and long-term adverse health outcomes (2). Impaired glucose tolerance (IGT) is characterized by less severe hyperglycemia but also associated with adverse outcomes like maternal postpartum glucose intolerance (3); increased risk of preterm birth, breech presentation, macrosomia, and hypoglycemia in newborns (4,5); and decreased lung function in children (6).

Diet is a critical factor related to prenatal glucose response. Dietary patterns (7), diet quality (8), and macronutrient distribution (9) have all been associated with gestational glucose tolerance. Timing and frequency of dietary intakes may also play a role, with studies finding that night eating syndrome and shorter night-fasting intervals negatively impact glucose control (10,11). However, chrononutrition studies often categorize eating times (11,12), rather than accounting for actual timing (e.g., late night eating at 9pm versus 1am), and the timing of these categories are inconsistent across studies (13,14). Additionally, studies tend to focus on total energy intake (10,14), rather than intakes of specific foods/food groups, despite evidence that the order in which meal elements are consumed can impact glucose response (15).

Collection of accurate timing data is challenging and could explain the lack of studies considering granular meal timing. Twenty four-hour recalls are commonly used, but unreliable for providing absolute timing of intakes (16). Food records can capture meal timing (17), but have a higher participant burden and are often collected for short durations. With few days of data per participant, food records may not capture the true breadth of individual foods in a person’s diet. In the current analysis, we used data previously collected through the Temporal Research in Eating, Nutrition, and Diet during Pregnancy (TREND-P) study (18). TREND-P collected up to 28 days of intakes logged in real time (using before and after photos and text notes for each eating occassion) during pregnancy (18). While participants must remember to log each eating occasion, the before and after photos provide time bounds, and are well-suited for examining chrononutrition. In this study we aimed to determine whether food category intakes and their timing were associated with IGT during pregnancy. We developed novel intake timing scores to use the granular data of TREND-P for chrononutrition analysis.

Methods

Study Sample

The TREND-P study was designed to examine associations between dietary intakes during pregnancy and hyperglycemia (18). Details of the TREND-P study were described elsewhere (18). Briefly, pregnant individuals were recruited from prenatal clinics affiliated with New York University (NYU) Langone Health. Eligibility criteria were gestational age <18 weeks, at least 18 years of age, able to read and write in English, and planning to receive prenatal care and deliver at NYU. Individuals were excluded if they had pre-existing type 1 or type 2 diabetes, had a history of eating disorders, or were diagnosed with GDM during dietary data collection. This study was approved by the Institutional Review Board at NYU Grossman School of Medicine.

Data Collection and Processing

Participants collected dietary data (text logs and pictures) during two, 14-day rounds; the first round started after enrollment (mean gestational age = 17.6 weeks, standard deviation, SD = 1.60) and the second round began approximately four weeks after completion of the first round (mean gestational age = 24.5 wk, SD = 1.72). The intent was for dietary data collection to be completed prior to prenatal GDM screening. Participants were provided with a smartphone and customized study app for recording food logs with before and after pictures of each eating occasion. They also received an Empatica E4 smartwatch for recording motion, temperature, heartrate, and electrodermal activity. Participants were instructed to wear the smartwatch during waking hours and to log all non-water food and beverage intake during the study.

Of the 176 participants who enrolled in the study, 32 participants were excluded due to: drop-out prior to data collection (n=7); difficulty with data collection (n=15); non-singleton pregnancy (n=3); miscarriage (n=2); extreme non-adherence to study protocol (n=2); unreliable dietary intake timing information (n=2); and missing glucose test results from the electronic health record (EHR, n=1). Our final analytic sample was 144 participants, depicted in the participant flowchart in Figure 1.

Figure 1.

Participant flowchart for the Temporal Research in Eating, Nutrition, and Diet during Pregnancy (TREND-P) study.

At the conclusion of each round of data collection, a research coordinator conducted a debrief interview with each participant to clarify dietary elements that were unclear from the photos and log notes (e.g., the type of meat used in a mixed dish, the type of milk added to coffee). Each participant also completed a questionnaire regarding socio-demographic (e.g., race and ethnicity, education, use of government programs) and behavioral (e.g., physical activity, tobacco use, prenatal vitamins) characteristics. Dietary intake during the study was determined by two clinical nutrition graduate students using before and after photos of the meals along with notes from the study app and debrief interview. Photo timestamps were used to determine the timing of eating occasions. We consider a day of dietary intake to begin at 4:00 am and continue to 3:59 am the following morning (19) as the hour from 3:30–4:30 am had the fewest eating occasions in the study. Nutritional information for each item in the TREND-P study was obtained using the ESHA Food Processor Nutrition Analysis software (version 11.1, ESHA Research). Every food item was also assigned a food code in the Food and Nutrient Database for Dietary Studies (FNDDS) (20). FNDDS codes were used in this study to obtain food categories and their respective serving sizes through the Food Patterns Equivalents Database (FPED) (21) and the level of industrial processing through the NOVA system, which categorized each FNDDS code for degree of industrial processing (22).

Glycemia Outcomes

Glucose test measures and medical diagnoses were obtained from participants’ EHR. As part of routine prenatal care, participants were screened for GDM (mean gestational age = 26.1 wk, SD = 3.5) using a non-fasting one-hour, 50-gram glucose challenge test (GCT). Participants with a GCT ≥135 mg/dL (n=40) received a follow-up fasting 2- (n=1) or 3-hour (n=37) 100-gram oral glucose tolerance test (OGTT); two participants did not have an OGTT documented in the EHR despite an abnormal GCT. We assessed glucose intolerance using the Carpenter-Coustan criteria: ≥95 mg/dL at baseline (fasting), ≥180 mg/dL at 1 hour, ≥155 mg/dL at 2 hours, and ≥140 mg/dL at 3 hours. Participants who met criteria at 2 or more time points (n=8) or who had a GDM diagnosis documented in the EHR (n=10) were categorized as having GDM (n=10). Participants with GCT values ≥135 mg/dL and who met criteria at fewer than 2 time points on the OGTT were categorized as having IGT (n=30) (6,8). For analyses, cases of IGT and GDM were combined (herein referred to as IGT, n=40), as both characterize a state of glucose intolerance and are associated with similar pregnancy risks (23). The combined category also reduced the sparsity of the data, which improved the stability and precision of model estimates while protecting against overfitting.

Morning and Night Intake Scores

To leverage the detailed timing of eating occasions in the TREND-P dataset, we calculated morning intake scores (MIS) and night intake scores (NIS) as illustrated in Figure 2. These scores can be used with energy, macronutrients, micronutrients, or specific food items. Based on previous literature, morning intakes ended at 11 am and night intakes started at 7 pm (24). The percentage of intakes was calculated for each eating occasion by dividing the intake value by the total accumulated intakes for the individual. These percentages of intakes were then multiplied by their respective morning and night hour (for MIS the multiplier ranged from 0 (11 am) to 7 (4 am) while the NIS multiplier ranged from 0 (7 pm) to 9 (3:59 am), see Figure 2) and summed to calculate the MIS and NIS. When these scores are used in linear models, the coefficients represent the change in response (for quantitative responses) or change in log odds (for binary responses) associated with a 1 percent shift of intakes one hour earlier in the morning for MIS or one hour later at night for NIS. During model building, when a model included MIS or NIS (as the exposure variable), the mean daily intake of that variable was also included in the model to ensure the timing score was not a proxy for total intake.

Figure 2.

Visualization of morning intake score (MIS) and night intake score (NIS) calculations. The figure shows six eating occasions (represented by the food drawing) and shows how the hour of consumption and percentage of daily intake the food represents are used to determine final scores. Here morning consumption receives a higher weight than night due to the larger intake amount.

Dietary Variables of Interest

We considered two groups of dietary variables, modeled separately. The first group represented dietary intakes and their respective timing scores (i.e., MIS and NIS). The second group comprised food category intakes and their timing scores. The dietary intake variables included:

• macronutrient intake (grams/day of protein, fats, and carbohydrates) and timing scores;

• total daily energy intake timing scores;

• temporal elements of the diet, including the average time (hour of day) of the first and last daily eating occasions, the average eating window (time of last meal – time of first meal), and the average number of eating occasions per day;

• degree of industrial processing, using the proportion of calories from minimally/unprocessed foods (NOVA category 1) and ultra-processed foods (UPF, NOVA category 4) (22) and UPF timing scores;

• dietary patterns extracted from the data using exploratory factor analysis (EFA).

Every food item in the FNDDS was mapped to 28 food categories and their respective serving sizes in the Food Pattern Equivalents Database (FPED) (21), which is representative of the Dietary Guidelines for Americans. Using the mean daily energy intake for each of the categories for each participant, we conducted EFA with varimax rotation after data normalization to extract dietary patterns. We used the scree plot, eigenvalue > 1, and interpretability as considerations for the number of factors. The food category variables were based on the FPED categories of foods used for the EFA dietary pattern extraction. Specifically, the mean daily intakes for each of the 28 food categories and their respective timing scores makeup the food category variables.

Confounders

Confounders included in all model development were age, early pregnancy body mass index (BMI), physical activity, total daily energy intake (kilocalories), and alternate healthy eating index during pregnancy (AHEI-P) score (a measure of overall diet quality) (8). Early pregnancy BMI (kilograms/meters²) was calculated using weight and height recorded at the first prenatal visit in the EHR (mean gestational age, SD = 9.3 weeks, 2.6). Physical activity was based on data from the triaxial accelerometer on the Empatica E4 smartwatch. We used an acceleration summary metric, the Monitor Independent Movement Summary (MIMS) algorithm (25), to extract physical activity (number of MIMS) for each minute of watch wear time. Keras et al. (26) validated summary measures of physical activity against ActiGraph activity counts and found that MIMS had better mean participant-specific correlations and mean absolute percentage error than Euclidean norm minus one, mean amplitude deviation, and activity intensity. For each participant, MIMS for the duration of the study were then summed and divided by the number of hours of watch wear time for our physical activity measure (MIMS/hour). For macronutrient models, we did not adjust for total daily energy intake due to multicollinearity and for dietary pattern models, we did not adjust for AHEI-P because the healthy, plant-based, diet was strongly correlated with AHEI-P (r = 0.83).

Statistical Analysis

We used the method of purposeful selection from Hosmer et al. (27) to build our logistic regression models. The method allows for the consideration of many independent variables while avoiding overfitting of the model and retaining coefficient stability. Purposeful selection involved an initial regression of the exposure variables individually (with confounders) to determine which variables should be included together in a preliminary model (p-value < 0.25, called candidate variables). We then went through iterations of model refinement for the preliminary model, removing individual variables based on highest p-value to minimize model Akaike Information Criterion (AIC). Variables that were not identified as candidate variables in the initial individual models were then reconsidered individually for possible entry into the model (i.e., if AIC was reduced). To protect against potential coefficient instability from too many variables in the model, we created three preliminary sub models (using the process just described for preliminary models) for each of the food category intake variables: mean daily intakes, MIS, and NIS. The significant variables (p-value < 0.05) for the three preliminary sub models were then included together in a single model where variables with the highest p-value were removed to minimize AIC. After validating each variable’s conformity to the assumption of linearity with the logit (using splines and likelihood ratio tests (27)), we had a resultant preliminary model. This preliminary model was then evaluated for potential interaction terms, which were only retained with p-value < 0.05. Finally, we assessed accuracy of the final model using Area Under the Receiver Operating Characteristic curve (AUROC). Model coefficient p-values were adjusted for multiple comparisons with the Benjamini-Hochberg procedure with α = 0.05 considered significant.

After achieving a final model, we checked for multicollinearity using the variation inflation factor (VIF). To assess the sensitivity of our models, we then executed the entire purposeful selection process on a reduced dataset with only the first week of data collection for each round.

All data analysis was conducted in R version 4.3.1. MIMS values were calculated using the MIMSunit package (28).

Results

Study Sample Characteristics

Tables 1–3 report descriptive characteristics of the sample (Table 1), general intake variables and macronutrient timing scores (Table 2), and food category intakes and timing scores (Table 3). IGT occurred in 27.8% of participants (n=30 for IGT, n=10 for GDM). The mean early pregnancy BMI was 24.8 kg/m² (95% confidence interval, CI: 24.0, 25.6), 22.2% (n=32) and 11.1% (n=16) had overweight and obesity, respectively. Mean daily energy intake was 1865 kilocalories (95% CI: 1787, 1943) and the mean AHEI-P score was 59.0 (95% CI: 57.1, 60.9). The mean NIS for macronutrient intakes were very similar (45.8, 45.8, and 46.6 for protein, carbohydrates, and fat). The food categories with the highest mean MIS were whole grains (77.1) and eggs (77.0). The food categories with the highest mean NIS were other starches (55.1), meat (54.4), and potatoes (52.6).

Table 1.

Descriptive characteristics for participants in the Temporal Research in Eating, Nutrition, and Diet during Pregnancy (TREND-P) study (n=144).

Characteristics	n (%)
Self-Reported Race and Ethnicity
Non-Hispanic Caucasian or White	74 (51.4)
Non-Hispanic Asian	29 (20.1)
Hispanic	22 (15.3)
Non-Hispanic Black or African American	12 (8.3)
Multi-Race	4 (2.8)
Non-Hispanic Middle Eastern	2 (1.4)
Non-Hispanic Native American or Alaskan Native	1 (0.7)
Education
Graduate/Professional School	83 (57.6)
College Graduate	49 (34.0)
Some College	9 (6.3)
GED or High School Graduate	3 (2.1)
Insurance
Private	112 (77.8)
Public	25 (17.4)
Other	5 (3.4)
Prefer not to answer	2 (1.4)
Employment Status
Full-time	111 (77.1)
Part-time	14 (9.7)
Not Employed	13 (9.0)
Prefer not to answer	6 (4.2)
Multiparous	57 (39.6)
	Mean [95% CI]
Body Mass Index	24.8 [24.0, 25.6]
Age (years)	32.5 [31.8, 33.2]
Physical Activity (mean MIMS1/hour)	593.9 [562.9, 624.8]
Total Daily Energy Intake	1865 [1787, 1943]
AHEI-P1 Score	59.0 [57.1, 60.9]

¹AHEI-P – Alternative Healthy Eating Index during Pregnancy; MIMS – Monitor Independent Movement Summary;

Table 3.

Descriptive statistics for food category intakes and timing scores. MIS and NIS do not have units.

Food Category Variables (serving measure)	Daily Intake (servings/day)	Morning Intake Score (MIS)	Night Intake Score (NIS)
	mean [95% CI]	mean [95% CI]	mean [95% CI]
Citrus, Melons, Berries (cups)	0.4 [0.35, 0.45]	54.9 [44.2, 65.6]	35.71 [27.64, 43.78]
Fruit Juice (cups)	0.18 [0.15, 0.22]	44.4 [33.0, 55.8]	37.97 [27.59, 48.35]
Other Fruits (cups)	0.64 [0.56, 0.73]	52.0 [43.3, 60.8]	42.7 [30.5, 54.9]
Potatoes (cups)	0.27 [0.24, 0.3]	9.6 [6.2, 13.0]	52.59 [44.16, 61.03]
Other Starches (cups)	0.08 [0.06, 0.09]	8.4 [2.5, 14.3]	55.06 [41.3, 68.82]
Tomatoes (cups)	0.26 [0.23, 0.28]	10.3 [6.4, 14.1]	49.9 [42.12, 57.68]
Other Red/Orange Vegetables (cups)	0.13 [0.11, 0.14]	7.7 [3.9, 11.8]	49.85 [39.59, 60.12]
Dark Green Vegetables (cups)	0.31 [0.27, 0.34]	9.3 [6.0, 12.7]	40.02 [32.16, 47.88]
Other Vegetables (cups)	0.65 [0.59, 0.71]	12.7 [8.7, 16.7]	50.69 [41.71, 59.68]
Legumes (cups)	0.13 [0.11, 0.15]	8.6 [4.0, 13.1]	47.81 [37.26, 58.37]
Whole Grains (ounces)	0.97 [0.85, 1.09]	77.1 [65.3, 88.9]	37.55 [28.56, 46.55]
Refined Grains (ounces)	5.6 [5.3, 5.9]	36.1 [30.7, 41.5]	45.27 [39.03, 51.5]
Meat (ounces)	1.05 [0.92, 1.18]	8.0 [1.9, 14.0]	54.35 [45.47, 63.22]
Poultry (ounces)	1.42 [1.25, 1.59]	4.8 [2.4, 7.3]	51.91 [43.58, 60.24]
Seafood High Omega-3 (ounces)	0.31 [0.24, 0.39]	5.6 [1.1, 10.0]	35.81 [24.71, 46.92]
Seafood Low Omega-3 (ounces)	0.46 [0.37, 0.55]	2.65 [0.63, 4.68]	43.08 [32.34, 53.83]
Cured Meat (ounces)	0.39 [0.33, 0.46]	29.9 [21.9, 37.9]	35.53 [26.88, 44.19]
Organ Meat (ounces)	0.01 [0.0, 0.02]	3.06 [0.0, 7.36]	2.57 [0.0, 7.59]
Eggs (ounces)	0.67 [0.61, 0.72]	77.0 [65.7, 88.3]	27.52 [20.49, 34.55]
Soy (ounces)	0.22 [0.14, 0.29]	22.6 [14.7, 30.5]	40.57 [27.42, 53.72]
Nuts and Seeds (ounces)	0.78 [0.66, 0.9]	59.7 [48.0, 71.4]	32.33 [23.79, 40.88]
Milk (cups)	0.53 [0.47, 0.59]	85.3 [74.2, 96.4]	45.57 [36.61, 54.54]
Yogurt (cups)	0.15 [0.13, 0.18]	72.0 [57.5, 86.4]	33.07 [20.68, 45.46]
Cheese (cups)	0.69 [0.62, 0.76]	22.3 [17.6, 27.0]	45.53 [37.68, 53.39]
Oils (grams)	26.3 [24.7, 27.9]	33.4 [27.8, 38.9]	44.4 [37.9, 50.8]
Solid Fats (grams)	30.9 [29.2, 32.6]	45.6 [39.3, 51.8]	49.3 [43.2, 55.4]
Added Sugars (teaspoons)	9.09 [8.24, 9.94]	55.3 [48.0, 62.6]	50.8 [44.5, 57.0]
Alcoholic Drinks (number of drinks)	0.01 [0.0, 0.02]	0.0 [0.0,0.0]	4.28 [0.0, 9.28]

^*Food categories and servings were obtained from the Food Patterns Equivalents Database (FPED).

Table 2.

Descriptive statistics of the general intake variables and timing scores (macronutrient intakes only).

		Morning Intake Score (MIS)	Night Intake Score (NIS)
General Intake Variables (unit of measure)	Mean [95% CI]	mean [95% CI]	mean [95% CI]
Eating Window (hours)	10.7 [10.4, 11.0]
First Eating Occasion (hour of day)	9.6 [9.4, 9.8]
Last Eating Occasion (hour of day)	20.0 [19.7, 20.2]
Number of Meals (n)	5.5 [5.2, 5.8]
Health Plant-based Dietary Pattern (Factor Score)	0.02 [−0.14, 0.18]
High Fat, Sugars, Refined Grains Dietary Pattern (Factor Score)	−0.01 [−0.17, 0.15]
Meat and Vegetables Dietary Pattern (Factor Score)	0 [−0.16, 0.17]
Percent of Calories from unprocessed/minimally processed foods	37.0 [35.4, 38.7]
Percent of Calories from ultra-processed foods	44.4 [42.8, 45.9]
Protein (g)	76.6 [73.2, 80.0]	38.7 [34.0, 43.3]	45.8 [40.0, 51.7]
Carbohydrates (g)	223.5 [212.7, 234.3]	44.2 [39.3, 49.1]	45.8 [40.0, 51.6]
Fat (g)	77.0 [73.5, 80.5]	39.4 [34.4, 44.3]	46.6 [40.7, 52.5]

^*Means are across participants based on daily mean values. MIS and NIS do not have units.

Dietary Pattern Extraction

We identified three dietary patterns (factor loadings in Supplemental Figure S1): Factor 1 represents a Healthy, Plant-based diet; Factor 2 represents a High Fats, Sugars, and Refined Grains diet; and Factor 3 represents a Meat and Vegetables diet. For each participant, their average daily servings for each food category were multiplied by each factor loading. These values were summed for each factor to create the factor score used in our model.

Model Development

From the general intake variables, five were initially included in the preliminary model based on their p-values in individual regression models (i.e., percent of calories from minimally/unprocessed foods; percent of calories from UPF; meat and vegetables dietary pattern; high fats, sugars, and refined grains dietary pattern; time of day of first eating occasion). The model results for each are shown in Figure 3A and final model results are shown in Figure 3B. No interactions were significant in the model and only time of day of the first eating occasion and percent of calories from minimally/unprocessed foods were retained in the model. However, we did not find associations between IGT and first or last eating occasion, eating window, or macronutrient timing scores.

Figure 3.

Model development results for the general dietary intake variables. In panel A, each row of data represents a model of the individual variable regressed on IGT with confounders. All variables above the purple line were moved into the preliminary model. Panel B is the final model for the general dietary intake variables.

Eighteen food category variables had p < 0.25 in their initial individual models as seen in Supplemental Figures S2–S4. These were used in their respective preliminary submodels, the results of which are shown in Supplemental Figures S5–S7. Figure 4A shows the final model after combining and refining submodels to minimize AIC. Two food category variables had significant associations with IGT (p < 0.05). Mean daily red meat intake was positively associated with IGT; each daily one-ounce increase in red meat intake was associated with 3.68 (95% CI: 1.68–8.75) times greater odds of IGT. Eggs NIS was negatively associated with IGT; each percentage of egg intake one hour later at night was associated with 1.05 (95% CI: 1.02 – 1.08) times lower odds of IGT. Since 13.5% of all egg intakes in our sample were at night, overall each hour of later consumption was associated with 1.9 (95% CI: 1.3–2.7) times lower odds of IGT. The model AUROC for predicting IGT was 0.836. In the sensitivity analysis, Potatoes and Other Fruits had an adjusted p-value of exactly 0.05, which does not meet our criteria for significance (p < 0.05) and neither achieved significance in the primary model. While Potatoes were significant in one of the preliminary submodels (Supplemental Figure S5), this did not persist once submodels were combined.

Figure 4.

Final food category model results with sensitivity analysis. Panel A is the final model results for the food category variables with the full dataset. Panel B is the final model when using the reduced dataset for the entire modeling process. Confidence intervals are cut off at 4 to preserve readability.

For sensitivity analysis, we conducted the complete model development process using a reduced dataset with only the first week of each round of data collection. Results are shown in Figure 4B. Mean daily intake of red meat and eggs NIS retained significance and direction of association (positive and negative, respectively). Two other foods achieved statistical significance during sensitivity analysis that were not in the primary model: mean daily intake of eggs (p = 0.046) and nuts and seeds (p = 0.046), both of which were negatively associated with IGT. The AUROC for the sensitivity analysis model was 0.845.

Discussion

In this prospective cohort study, we found significant associations between food categories and food category timing scores with IGT. Specifically, we found that mean daily intake of red meat was associated with greater odds of IGT, while egg NIS was associated with lower odds of IGT. Other dietary composition variables, such as energy and macronutrient intakes, were not associated with IGT. These findings support existing dietary guidance and provide greater insight about the relationship between specific foods and glycemia during pregnancy.

Existing research exploring associations between dietary characteristics and hyperglycemia during pregnancy has found mixed results, with some studies reporting no associations between diet quality score and maternal glucose response (29), and others reporting a positive association between less healthy dietary patterns (high in refined grains, fat, and added sugars with low intake of fruits and vegetables) and higher odds of GDM (7). We did not find an association with overall dietary patterns and IGT, which may be due to better characterization of usual diet (up to 28 days per participant). Prior results for intake timing generally showed negative associations for late eating times: night eating syndrome was associated with insulin resistance (10) and higher fasting glucose during pregnancy (14). Additionally, skipping breakfast before and during early pregnancy was associated with increased odds of GDM (12). However, these studies did not examine specific foods consumption in relation to timing, which our results suggest may be an influential factor since we did not find associations between IGT and general chrononutrition variables.

Previous studies often discretized intakes (e.g., quantiles for dietary patterns (7) and quality indexes (29), breakfast meals (12), night consumption (14), and night eating syndrome cutoff (10)), which may explain differences across results. Such strategies involve binning observations using subjective cut-points or thresholds that may lack consensus or vary across populations. Yet it is unlikely that intake timing has no effect before a boundary and a full effect shortly after (e.g., 6:45 pm versus 7:15 pm with night eating defined as after 7pm). Another common approach is ranking observations to compare outcomes between the first and last quantiles, but this can reduce statistical power (30). Further, previous studies commonly used either an FFQ (9,12,29), a night eating questionnaire (10), or two 24-hour recalls (7). However, food recalls are unreliable for capturing meal timing (16) while FFQs do not capture timing and both are subject to recall bias (16). In contrast, TREND-P captured detailed timing and intake data in real time and using the specific timing and intake data in our analyses instead of discretized versions may account for the discrepancy in results. Loy et al. (14) conducted a similar study to ours using 4-day paper and photo-based dietary records. However, they binned eating occasions into two 12-hour periods (07:00–18:59 and 19:00–06:59) and the choice of boundaries for these periods may have contributed to the difference in results. Our new intake timing measures, NIS and MIS, do not use strict cutoffs and weight more extreme times more strongly (e.g., intakes at 2:00am contributing more to a NIS than intakes at 8:00pm). This makes our analysis less sensitive to the choice of boundaries for each time period, and ensures more extreme behaviors contribute more to each score.

The associations between mean daily intakes of red meat and eggs NIS and IGT persisted in the sensitivity analysis. Other studies have found an increased GDM risk with higher intakes of red and processed meat (31) and with prenatal heme iron intakes (animal-based foods), while total, non-heme, and supplemental iron intakes were not associated with GDM (32). Excess iron intakes and high body stores lead to oxidative stress that may impair pancreatic beta cell function and, consequently, insulin secretion (33). Heme iron is more bioavailable than non-heme iron, which may explain the obeserved differences in GDM risk. Given that iron is a critical nutrient need during pregnancy (32), reductions in intakes of red meat should be compensated with iron from alternate sources (e.g., poultry, legumes, dark green vegetables, and fortified foods). In non-pregnant adult populations, regular egg consumption has been shown to improve glycemic control. In one randomized controlled trial, daily consumption of one large egg significantly reduced fasting blood glucose levels by 4.4% (34) and in an observational study, intakes of five or more eggs per week were associated with lower mean fasting glucose levels and a reduced risk of developing impaired fasting glucose or type 2 diabetes (35). During pregnancy, evidence for egg consumption is mixed, with studies reporting both higher (36,37) and lower (38,39) odds of GDM. Our results add nuance to this discussion, as we found timing of egg intake to be important, particularly at night.

The nutrient composition of eggs, including carotenoids and vitamin D, may explain possible beneficial effects on blood glucose control (35). Carotenoids act as antioxidants and have a range of biological activities, such as reducing inflammation and oxidative stress, which lower fasting blood glucose concentrations and insulin resistance during pregnancy (40,41). Similarly, Vitamin D is a fat-soluble steroid hormone that may improve pancreatic beta cell function through mechanisms related to insulin receptor expression and calcium metabolism (42). One randomized controlled trial found that vitamin D supplementation reduced insulin resistance among individuals with GDM (43). While night eating is commonly associated with health risks (11,13,14), the NIS for eggs was associated with lower odds of IGT in the primary and sensitivity analysis. While no study has specifically investigated egg consumption at night and glucose response, this finding is supported by related research, as small, nutrient-dense, low-energy foods consumed at night might not have the same negative outcomes as mixed meals (44). This could imply that eggs have different effects than other types of nighttime meals.

Mean daily intake of nuts and seeds was negatively associated with IGT in the sensitivity analysis. Nuts and seeds are sources of polyunsaturated fatty acids, vitamins and minerals, fiber, and other beneficial plant-based compounds like polyphenols and phytosterols, all of which may contribute to glycemic control (45). There is some evidence from randomized controlled trials of nut and seed consumption leading to improved fasting glucose in adults with prediabetes (46), and in a prospective cohort study early pregnancy nut and soy intake was associated with lower GDM risk (47). Dietary patterns before pregnancy that include higher intakes of nuts and seeds are also associated with a reduced risk of GDM (48).

Our study had several limitations. First, the sample consisted of educated individuals with private insurance receiving prenatal care at an urban academic medical center, which may limit the generalizability of findings to other populations. Futher, the detailed data collection and annotation approach of TREND-P limited the sample size, so future work is needed on a larger population. We combined GDM and IGT (clinically distinct conditions) for analysis because of the low individual prevalence of each condition. This eliminated the ability to find associations specific to the individual conditions, but we believe it was justified by the similar nature of the conditions and the benefits of a more balanced dataset given the small sample size. Future work is needed to investigate whether dietary associations may differ between these conditions. While we cannot rule out model overfitting connected to our sample size, our purposeful selection modeling approach (with assumption and variation inflation checks), sensitivity analysis, and high model accuracy provide confidence in these associations. The study was also observational, not an intervention, so we cannot make any causal claims about the discovered associations. It also lacked supplement data which may have provided nuanced information for foods that provide critical nutrients (e.g. red meat and iron). Finally, the TREND-P study did not assess sleeping patterns of participants, so the role of circadian rhythms in these analyses could not be assessed.

In this study we found that general dietary composition (e.g., macronutrients) and meal timing may not fully capture dietary associations with IGT, even when using detailed timing data. Our introduction of morning and night intake scores enabled discovery of a novel association between night egg consumption and lower risk of impaired glucose tolerance. Our results show the importance of detailed dietary data for examining chrononutrition. Future intervention studies are needed to determine whether the associations observed are causal, and whether other factors such as sleep patterns may play a role.

Abbreviations:

AHEI-P

Alternate healthy eating index during pregnancy

AIC

Akaike Information Criterion

AUROC

Area Under the Receiver Operating Characteristic

EFA

Exploratory factor analysis

EHR

Electronic health record

FFQ

Food frequency questionnaire

FNDDS

Food and Nutrient Database for Dietary Studies

FPED

Food Patterns Equivalents Database

GCT

Glucose challenge test

GDM

Gestational diabetes mellitus

IGT

Impaired glucose tolerance

MIMS

Monitor Independent Movement Summary

MIS

Morning intake score

NIS

Night intake score

NYU

New York University

OGTT

Oral glucose tolerance test

TREND-P

Temporal Research in Eating, Nutrition, and Diet during Pregnancy

UPF

Ultra-processed food

VIF

Variance inflation factor

Data Availability*:

Data described in the manuscript, code book, and analytic code will be made available upon request pending application and approval.