Racial/Ethnic-Derived Maternal Diets Predict Birth Outcomes Better than a Diet Derived from a Combined Sample among Hispanic/Latina and non-Hispanic White Pregnant Individuals in the ECHO Cohort

Abstract

Little is known about diet based on maternal fasting blood glucose (FBG) and birth outcomes in diverse populations. We hypothesized that racial/ethnic-derived FBG-based diets would predict birth outcomes better than a diet derived from the overall sample. Pregnant Hispanic/Latina (n=420) and non-Hispanic White (n=564) individuals (combined, n=984) from two Environmental influences on Child Health Outcomes (ECHO) cohorts provided ≥ 1 24-hr diet recalls. We evaluated primary (birthweight [BW]-for-age, macrosomia, large-for-gestational age, and preterm birth) and secondary (BW, gestational age [GA] at birth, low birthweight (LBW), and small-for-gestational age) birth outcomes. Reduced-rank regression with maternal FBG was used to derive dietary patterns in the combined and racial/ethnic samples. We used multivariable linear and logistic regression to estimate diet associations with birth outcomes. In each racial/ethnic group, seemingly unrelated estimation with clustering was used to test differences in diet coefficients between combined and racial/ethnic-specific models. The overall sample diet was characterized by higher intakes of refined grains and lower intakes of whole grains, solid fats, and nuts and seeds; racial/ethnic-derived diets were similar, with some exceptions. A one-standard deviation increase in the combined pattern was significantly associated with lower BW-for-age (β=−0.08, 95% confidence interval [CI]:−0.16, −0.004), BW (β=−57.5, 95% CI:−94.8, −20.2), and GA at birth (β=−0.13, 95% CI:−0.24, −0.01) and greater odds of preterm birth (OR=1.41, 95% CI:1.03, 1.94) and LBW (OR=1.62, 95% CI:1.07, 2.46), with pronounced coefficients in racial/ethnic-derived diets. Overall, racial/ethnic-derived FBG diets predicted adverse birth outcomes better than the diet derived from the overall sample in each racial/ethnic group.

Graphical Abstract

This study identified diet patterns linked to fasting glucose (FG) in Hispanic/Latina and non-Hispanic White pregnant individuals, separately, and in a combined sample of the two. We identified a diet high in refined grains and low in whole grains, solid fats, and nuts/seeds that was adversely linked to birthweight (BW)-for-age, BW, gestational-age-at-birth, low BW, and preterm birth, with stronger racial/ethnic-derived diet effects.

1. Introduction

In the US, Hispanic/Latina pregnant individuals are generally disproportionately affected by adverse birth outcomes, including higher proportions of low birthweight (LBW; <2,500 g) and preterm births, as compared to their non-Hispanic White counterparts [1, 2]. Diet during pregnancy can potentially serve as a point for intervention. Indeed, maternal diet during pregnancy has been posited to be an important modifiable risk factor for suboptimal fetal growth and, consequently, adverse birth outcomes. However, little is known about whether cultural dietary differences in food combinations in the overall diet play a meaningful role in the development of adverse birth outcomes across different racial and ethnic groups.

A recent review and meta-analysis of maternal dietary patterns, or overall diet, during pregnancy and birth outcomes highlighted that most US studies have defined diet using theory-based or data-driven approaches. Theory-based approaches typically apply a priori diet criteria that have been previously linked to health (e.g., the Healthy Eating Index [HEI]) [3], but these approaches may not result in the identification of dietary patterns that adequately reflect dietary behaviors of the study population. By contrast, data-driven approaches (e.g., exploratory factor analysis) normally derive dietary patterns using the dietary intake data of the study population directly; however, the resulting patterns may not be associated with specific health endpoints, such as adverse birth outcomes [4]. Alternatively, other methods, such as reduced-rank regression (RRR) provides the potential for uncovering underlying dietary patterns that would otherwise be missed using theory-driven approaches while improving the biological/health relevance of resulting patterns unlike data-driven methods that typically use dietary intake data alone [5]. For example, RRR identifies combinations of foods eaten together, as in an overall diet, that correlate the best with an intermediate variable (e.g., maternal fasting blood glucose [FBG]) that is biologically relevant (i.e., risk factor) to a health outcome of interest (e.g., adverse birth outcomes) [6]. As a result, RRR-derived diets based on maternal FBG would capture mechanistic pathways of diet’s influence on maternal FBG and may be further examined in relation to birth outcomes in downstream analyses.

The distribution of important metabolic biomarkers, such as FBG, varies by race/ethnicity [7]. Because diet also varies by race/ethnicity [8], deriving racial/ethnic-specific dietary patterns using racial/ethnic-specific FBG distributions may allow derivation of dietary patterns that are more strongly linked to FBG and, consequently, to birth outcomes among the racial/ethnic group in which patterns are derived. A prior RRR study derived dietary patterns using maternal FBG found significant associations mostly between the relevant RRR-derived pattern and greater newborn birth weight, fat mass, and adiposity, but failed to evaluate dietary pattern associations with clinically relevant birth outcomes or endpoints (e.g. large-for-gestational age [LGA]) [9].

Although this same study included a small proportion of Hispanic/Latina pregnant participants (20%), racial/ethnic-specific findings were not evaluated, and the results most likely reflected the dietary behaviors of the majority group. Consequently, the generalizability of these findings may not extend to diverse and socioeconomically disadvantaged populations [10] who are disproportionately affected by adverse birth outcomes and for whom interventions are most needed. Finally, because RRR-derived dietary patterns linked to maternal FBG may not necessarily have clear “healthy” or “unhealthy” dietary characteristics, evaluating the healthfulness of derived diets using a diet quality measure such as the HEI may help contextualize findings.

Therefore, in this study, we aim to derive dietary patterns (using maternal FBG) during pregnancy among non-Hispanic White and Hispanic/Latina individuals, separately, and then compared these dietary patterns with each other and with patterns derived in a racially/ethnically pooled sample of the two groups. We hypothesize that all derived dietary patterns will demonstrate differences in intake at the food/component level. Additionally, we aim to test whether RRR-derived dietary patterns are associated with diet healthfulness using HEI-2015. We hypothesize that there will be stronger relationships between HEI-2015 and derived dietary patterns with food group intakes that follow the scoring for corresponding foods in HEI-2015. Finally, we will examine whether racial/ethnic-derived dietary patterns predict birth outcomes (primary: birth-weight [BW]-for-age z-scores, preterm birth, macrosomia, and LGA; secondary: low birthweight [LBW], small-for-gestational age [SGA]) better than a diet derived from the overall sample in each racial/ethnic group. We hypothesize that racial/ethnic-derived dietary pattern estimates will be greater in magnitude than the estimates for the diet derived from the overall sample in each racial/ethnic group. Our findings will advance scientific knowledge by identifying whether racial/ethnic-specific diets predict adverse birth outcomes better in their respective racial/ethnic group than those derived in racially/ethnically mixed population.

2. Methods and Materials

2.1. Study Participants

Data were obtained from pregnant individuals enrolled in two cohorts participating in the National Institutes of Health Environmental influences on Child Health Outcomes (ECHO) Cohort, a consortium of observational cohorts of mothers and children formed to understand the role of early life exposures on children’s health and development [11]. We used available 24-hr dietary intake and blood biospecimen data from two sociodemographically distinct prospective pregnancy cohorts including the Maternal And Developmental Risks from Environmental and Social Stressors (MADRES) cohort and the Healthy Start cohort. Among the ECHO cohorts, we only included MADRES and Healthy Start in the present analysis based on: 1) the inclusion of Hispanic/Latina and non-Hispanic White pregnant individuals; and 2) similar timing of data collection (i.e., 3rd trimester) and detailed methodology of dietary intake (i.e., 24-hr diet recalls versus food frequency questionnaires), maternal FBG, birth outcomes of interest, and relevant covariates. All cohort-specific protocols were approved by their respective university institutional review boards (Healthy Start, COMIRB 09–0563; MADRES Population Core, HS-15–00498), and written informed consent was obtained from participants.

2.1.1. Healthy Start

From 2009–2014, Healthy Start recruited 1410 pregnant women from outpatient prenatal clinics at the University of Colorado Hospital. The study population includes predominantly non-Hispanic White pregnant individuals aged 16–45 years old residing in Aurora, CO [9]. Eligible individuals were ≥16 years of age; had a singleton pregnancy; had completed <24 weeks of gestation at the time of enrollment; had no pre-pregnancy health diagnoses, including diabetes mellitus, cancer, psychiatric illness, or asthma treated with steroid medications; and reported no history of stillbirth or preterm birth at <25 weeks of gestation. Participants completed two in-person prenatal clinic visits at a median of 17 weeks (range: 10–24 weeks) and 27 weeks (range: 24–32 weeks) of gestation and one visit at delivery. Because some individuals had subsequent pregnancies enrolled in Healthy Start only, we excluded Healthy Start participants who were on their second (or greater) enrolled pregnancy.

2.1.2. MADRES

From 2015–2023, the MADRES Study recruited 883 pregnant women from four prenatal care centers (one county hospital, two community health clinics, and one private obstetrics and gynecology practice) serving predominantly low-income Hispanic/Latino communities in Los Angeles, CA [12]. The study population is predominantly low-income, of Hispanic/Latino origin, and aged 18–45 years. Eligible individuals were ≥18 years of age, had a singleton pregnancy, had completed <30 weeks of gestation at the time of enrollment, were not incarcerated, and had language fluency in English or Spanish.

2.2. Birth Outcomes

In Healthy Start and MADRES, BW (grams) and infant biological sex (male, female) were abstracted from medical records. Similarly, gestational age (GA) at birth (weeks) was abstracted from medical records in Healthy Start, while in MADRES, a hierarchy of methods [13] was used based on data availability, in the following order: 1) 1^st trimester (<14-wk GA) ultrasound measurement of crown–rump length (60%); 2) 2^nd trimester (<28-wk GA) ultrasound measurement of fetal biparietal diameter (25%); 3) if measures from an early ultrasound were unavailable, the physician’s best clinical estimate abstracted from maternal medical records (14%); and 4) self-reported last menstrual period date (<1%). We then generated sex-specific BW-for-GA z-scores using a representative US reference that reflects current trends in obesity and gestational diabetes, both of which can impact fetal growth [14]. For clinical relevance, we also evaluated dichotomous BW outcomes using standard definitions: low BW (<2500 g), macrosomia (>4500 g), SGA (BW-for-GA ≤ 10%), LGA (BW-for-age ≥ 90%), and preterm birth (GA < 37 weeks).

2.3. Dietary Assessment

Participants completed ≥ 1 staff- or self-administered 24-hr dietary recalls (≥ 2 recalls overall: 87.2%) using the National Cancer Institute’s Automated Self-Administered 24-hr Recall (ASA24) tool [15] on separate days. The ASA24 tool asks participants to recall all foods and beverages consumed in the previous 24 hours. ASA24 estimates day-level nutrient and food group values using the Food and Nutrient Database for Dietary Studies (version 4.1, 2010) and the MyPyramid Equivalents Database (version 2.0, 2008) [16]. The MyPyramid Equivalents food group servings are consumer-guided daily amounts of foods and beverages from each food group designed to meet nutrient needs based on the 2015–2020 Dietary Guidelines for Americans [17]. These standard conversions facilitate the comparison of food amounts from different foods within the same food group. In the fruit and vegetable groups, for instance, a one-cup equivalent is the amount of a food considered to be equivalent to one cup of a cut-up fruit or vegetable; meanwhile, in the milk group, the one-cup equivalent is based on volume [18]. Daily intakes from each MyPyramid Equivalents food group were averaged across recalls for participants with ≥2 recalls. Participants were excluded if their average energy intake was considered implausible (<500 or >5000 kcal/d). Finally, because our goal was to understand intake from food alone, we excluded dietary supplement data.

Out of the 28 original food-group equivalents, soy products and alcoholic beverages were excluded from analysis due to low consumption (<5%) in either the combined or racial/ethnic-specific study samples [9]. Because many food-group equivalents showed highly skewed distributions, the remaining 26 food groups were either used as they were or aggregated into the following 12 major foods groups: dairy, fruit, vegetables, white potatoes, whole grains, refined grains, meat, eggs, nuts and seeds, oils, solid fats, and added sugars (Supplementary Table 1). Then, the average intakes of the 12 food groups were log-transformed by replacing zero values with an arbitrarily small value of 0.001 [9].

To quantify diet healthfulness of RRR-derived dietary patterns, we used the HEI-2015, which is a diet quality indicator that captures how closely an eating pattern or combination of foods adheres to the 2015–2020 Dietary Guidelines for Americans recommendations [3, 19]. HEI-2015 includes nine components centered on adequacy (total fruit, whole fruit, total vegetables, greens and beans, whole grains, dairy, total protein foods, seafood and plant proteins, and fatty acids) and four focused on moderation (refined grains, sodium, added sugars, and saturated fats), for a total of 13 dietary components, all of which are equally weighted. Using simple HEI scoring, HEI-2015 scores were calculated to assess overall diet quality on the individual level; scores ranged from 0–100, with higher scores indicating greater diet quality [20].

2.4. Maternal Fasting Blood Glucose (FBG)

We chose maternal FBG as the sole intermediate/response variable in RRR based on: 1) previous work demonstrating associations between maternal FBG and adverse birth outcomes at different stages of pregnancy including the 3^rd trimester [21], a period during which dietary assessment was performed similarly in both Healthy Start and MADRES; and 2) to ensure derived dietary patterns specifically capture maternal FBG variation, since including multiple intermediate/response variables in RRR would result in dietary patterns that explain the shared variance across all intermediate variables, potentially obscuring associations unique to maternal FBG.

In Healthy Start, maternal FBG was measured using only fasting blood samples collected by trained staff. In MADRES, FBG values were ascertained using a combination of fasting blood samples collected by staff and electronic medical records. Here, fasting is defined as not having ingested anything except for water for at least 8 hours. In both studies, samples were immediately transported to each study’s respective laboratory site for processing [9, 12]. While dietary assessment was normally administered following blood collection on the same day in Healthy Start, in MADRES, in some cases dietary assessment was performed several days prior (59%) if not after (41%) blood sample collection. For these instances, we calculated differences in dates between blood collection and dietary assessment and excluded MADRES study participants with extreme negative differences (≥4 standard deviations [SDs]) in dates.

2.5. Covariates

Electronic medical records and staff-administered questionnaires during in-clinic visits were used to ascertain relevant maternal characteristics including Latina/Hispanic ethnicity and race (Hispanic/Latina, non-Hispanic White), age, pre-pregnancy body mass index (BMI), maternal education, co-habitation status, parity (1^st, 2^nd, 3^rd or more), and total daily energy intake (kilocalories) calculated using ASA24 and averaged for individuals with ≥2 dietary recalls. Total physical activity (metabolic equivalents [METs]/week) was captured using the Pregnancy Physical Activity Questionnaire in the third trimester of pregnancy [22]. Because the proportion of individuals who reported prenatal smoking was small (n=43, 4.37%) in our sample, we did not include maternal smoking during pregnancy as a covariate in our models.

2.6. Statistical Analyses

Of the 1410 participants in the Healthy Start study and 883 participants in MADRES, we excluded participants who self-identified as belonging to a racial/ethnic category other than non-Hispanic White or Hispanic/Latina, were on their second (or more) pregnancy while enrolled in the Healthy Start cohort, had missing or incomplete ASA24 data, had implausible energy intakes, had missing covariates, or had an extreme difference in dates (> 4 SDs) between dietary assessment and blood biospecimen collection (Supplementary Figure 1). Because individuals with a history of gestational diabetes mellitus (GDM) tend to be metabolically different and have healthier diets than those reporting no history of GDM [23], we additionally excluded individuals who self-reported GDM or had missing GDM data. These exclusions resulted in 420 Hispanic/Latina and 564 non-Hispanic White study participants, for a total of 984 study participants in the combined sample.

Before performing RRR, we log-transformed dietary predictors and maternal FBG to improve distribution normality. Because the number of dietary patterns derived cannot exceed the number of intermediate variables used in RRR [24], one dietary pattern was obtained by entering the 12 MyPyramid food group intake predictors (dairy, fruit, vegetables, white potatoes, whole grains, refined grains, meat, eggs, nuts and seeds, oils, solid fats, and added sugars) into an RRR model designed to simultaneously maximize the variability explained in maternal FBG during pregnancy (intermediate/response variable) in the combined, non-Hispanic White-specific, and Hispanic/Latina-specific samples, separately. For ease of interpretation, RRR food group loadings from the combined sample and from each racial/ethnic-specific analysis were used to generate a heat map. Because the pattern structure of RRR tends to be more complicated than those in other data-driven approaches (e.g., principal component analysis), the factor loading coefficients, or correlations of each food component with the corresponding derived pattern, are generally weaker [24]. As in previous work, we considered food groups with absolute loadings >0.10 to meaningfully contribute to a corresponding derived dietary pattern [25]. For a given food group, a higher absolute factor loading coefficient indicates a stronger weighted correlation (accounts for all other food group model inputs) with the relevant dietary pattern. We also include unweighted Pearson correlation coefficients between intakes of each MyPyramid Equivalents food group and 1) maternal FBG and 2) RRR-derived dietary patterns (Supplementary Table 2). Standardized dietary pattern scores for each individual were generated by multiplying the loading coefficient of each food group by the corresponding food group intake and summing across foods. A one-SD dietary pattern score increase reflects greater adherence to the dietary pattern in question. Dietary pattern scores were standardized and divided into tertiles for ease of interpreting findings. To evaluate the diet healthfulness of the derived dietary patterns, we performed analysis of variance (ANOVA) tests with Bonferroni adjustment (statistical significance set at p < 0.01) to compare mean HEI-2015 scores by dietary pattern tertiles in the combined and racial/ethnic-specific samples, and we evaluated Pearson correlations between each dietary pattern (scores) and HEI-2015.

Maternal characteristics were compared between Hispanic/Latina and non-Hispanic White participants using t-tests and Pearson chi-squared tests as appropriate. To evaluate the role of diet on infant BW and GA, mean differences in infant BW and GA were tested using multivariable linear regression models. For clinical birth outcomes (LBW, macrosomia, SGA, LGA, and preterm birth), multivariable logistic regression was employed. Consistent with previous work in Healthy Start and MADRES [9, 26], all multivariable models were adjusted for maternal age (years), education (< high school graduate, ≥ high school graduate), pre-pregnancy BMI (calculated using weight [kg] divided by height [m²]), parity (1^st, 2^nd, 3^rd or more), total physical activity (MET-min/week), co-habitation status (living with a partner, not living with a partner), total daily energy intake (kilocalories), and study site (Los Angeles, CA; Aurora, CO). For analyses including diet derived in the combined sample, the power needed to detect a one-SD difference in BW-for-age Z-scores with a sample size of 440 in each race/ethnicity group (Hispanic/Latina, non-Hispanic White) and an alpha of 0.05 was estimated at 0.881 [9].

We excluded GA (at delivery) from the model adjustment set for birth weight (grams) because GA may mediate associations between maternal diet during pregnancy and infant birth outcomes and may introduce bias if included [27]. We therefore evaluated GA and BW-for-GA as distinct birth outcomes.

To determine whether racial/ethnic-specific dietary patterns predicted birth outcomes better than the pattern derived in the combined sample in each racial/ethnic group, we used seemingly unrelated estimation (SUEST), which combines estimation results (i.e., parameter estimates and related (co)variance matrices) under one parameter vector and simultaneous (co)variance matrix and allows the evaluation of cross-model hypotheses [28]. We used SUEST with the cluster option using study location (i.e., Los Angeles, CA; Aurora, CO) as the cluster variable to directly test differences between the dietary pattern coefficients between models using the dietary pattern derived from the combined sample and models using the racial/ethnic-specific derived dietary pattern in each racial/ethnic group, adjusting for the same covariate adjustment set described above.

Finally, to evaluate the potential for selection bias in our study, we compared select demographic and reproductive health characteristics by the inclusion/exclusion status of participants. All analyses were performed in Stata version 16.1 (Stata Corp). We considered statistical significance at p < 0.05.

3. Results

Participant characteristics by inclusion/exclusion status of the eligible study participants in each pregnancy cohort study are shown in Supplementary Table 3. In Healthy Start, pregnant individuals who were included in the analysis were significantly more likely to have newborns with higher BW and GA and lower risk for LBW and preterm birth, to be older, to cohabitate, to have a high school education or greater, and to have a lower pre-pregnancy BMI, as compared to excluded individuals. Meanwhile, in MADRES, study participants who were included in the analysis were significantly less likely to cohabitate.

Table 1 shows participant characteristics by race/ethnicity. Compared to non-Hispanic White pregnant individuals (n = 564, 57.3%), Hispanic/Latina individuals (n = 420, 42.7%) were significantly older at study entry (30.4 versus 27.9 years), had more than twice the proportion of preterm births (8.3% versus 4.4%), and were more likely to have less than a high school education (58.1% versus 10.5%), report not living with a partner (35% versus 83.0%), report having had three or more births (32.6% versus 15.4%), have higher pre-pregnancy BMI (28.6 versus 24.5), higher maternal FBG measurements (85.7 versus 77.3 mg/dL), and lower mean HEI-2015 scores (50.9 versus 55.6). Overlayed histograms of log-transformed maternal FBG of each racial/ethnic group are displayed in Supplementary Figure 2.

Table 1.

Participant characteristics in Non-Hispanic White (n=564) and Hispanic/Latina (n=420) pregnant individuals and a combined sample of the two racial/ethnic groups

Characteristic	Combined	Non-Hispanic White	Hispanic/Latina	p b
	n=984	n=564	n=420
Maternal age, y	29 ± 5.8	26.9 ± 6.2	30.4 ± 5	<0.001
Birth outcomes
Birthweight (g)	3293.8 ± 478	3313.1 ± 479	3267.9 ± 475	0.142
Gestational age at birth (wks)	39.4 ± 1.5	39.6 ± 1.5	39.1 ± 1.5	<0.001
Birthweight-for-age (z-score)	−0.3 ± 1	−0.34 ± 1	−0.26 ± 1	0.237
Low birthweight (<2500 g) (%)	40 (4.1)	24 (4.3)	16 (3.8)	0.726
Macrosomia (>4500 g) (%)	67 (6.8)	40 (7.1)	27 (6.4)	0.683
Small-for-gestational age (%)	84 (8.5)	56 (9.9)	28 (6.7)	0.070
Large-for-gestational age (%)	105 (10.7)	54 (9.6)	51 (12.1)	0.197
Preterm birth (<37 weeks) (%)	60 (6.1)	25 (4.4)	35 (8.3)	0.011
Sociodemographics
Newborn sex
Female (%)	473 (48.1)	263 (46.6)	210 (50)	0.295
Male (%)	511 (51.9)	301 (53.4)	210 (50)
Cohabitation (%)
Yes	615 (62.5)	468 (83)	147 (35)	<0.001
No	369 (37.5)	96 (17)	273 (65)
Education (%)
< High school	615 (62.5)	59 (10.5)	244 (58.1)	<0.001
High school or beyond	369 (37.5)	505 (89.5)	176 (41.9)
Study (%)
Healthy Start	750 (76.2)	542 (96.1)	208 (49.5)	<0.001
MADRES	234 (23.8)	22 (3.9)	212 (50.5)
Dietary recalls (%)
One	126 (12.8)	43 (7.6)	83 (19.8)	<0.001
Two or more	858 (87.2)	521 (92.4)	337 (80.2)
Reproductive health
Parity
1st	455 (46.2)	300 (53.2)	155 (36.9)	<0.001
2nd	305 (31.0)	177 (31.4)	128 (30.5)
3rd or more	224 (22.8)	87 (15.4)	137 (32.6)
Pre-pregnancy BMI	26.3 ± 6.2	24.5 ± 5.1	28.6 ± 6.7	<0.001
Fasting blood glucose (mg/dL)	80.9 ± 14.1	77.3 ± 8.5	85.7 ± 18.1	<0.001
Gestational age at blood biospecimen collection	27.4 ± 6.3	27.2 ± 3.3	27.7 ± 8.8	0.275
Total physical activity (MET-min/wk)3	184.6 ± 102.6	165.4 ± 84.7	210.4 ± 117.8	<0.001
Energy intake (kcals)	2005.9 ± 577.9	2077 ± 503.1	1917.3 ± 648.9	<0.001
Healthy Eating Index (2015)	53.6 ± 10.7	55.6 ± 10.8	50.9 ± 9.9	<0.001
Gestational age at dietary assessment	28.2 ± 3.0	27.3 ± 2.6	29.4 ± 3.1	<0.001

Abbreviations: MADRES, Maternal And Developmental Risks from Environmental and Social Stressors; MET, metabolic equivalent.

^aValues are means ± SDs or n (%).

^bp-values were calculated using analysis of variance for continuous variables and chi-square tests for categorical variables.

Figure 1 displays a heat map of food group loadings from RRR for dietary patterns derived using the combined sample and each racial/ethnic group. In the combined sample, refined grains (0.18) had the highest positive loading, while solid fats (−0.23), whole grains (−0.16), and nuts and seeds (−0.12) had the highest negative loadings. Among non-Hispanic White participants, whole grains (−0.19), fruit (−0.18), and solid fats (−0.11) had the highest negative loadings. Among Hispanic/Latina individuals, refined grains (0.17) and fruit (0.11) had the highest positive loadings, while solid fats (−0.20), white potatoes (−0.13), and oils (−0.11) had the highest negative loadings.

Figure 1. Comparison of food group factor loadings of each dietary pattern derived in reduced rank regression using maternal fasting blood glucose during pregnancy in Hispanic/Latina, non-Hispanic white, and combined samples^a,b,c.

^a Loadings are positive (red hue) and negative (green hue) correlations between 12 food groups and derived dietary patterns. Loadings > 0.10 in absolute are considered to meaningfully contribute to the derived dietary pattern [24, 25]

^b Variation in log-transformed glucose (mg/dL) explained by food groups in the combined sample and in each racial/ethnic sample

^c Pearson correlations between each dietary pattern (score) and log-transformed glucose (mg/dL).

*** p<0.001; ** p<0.01; * p<0.05.

Abbreviations: NH-White, non-Hispanic White

One dietary pattern was derived in each of the three study samples using RRR: one overall dietary pattern in the combined sample and one in each racial/ethnic group. On average, the dietary pattern from the combined sample explained 13.2% of the variability among food groups and 6.9% of the variability in the intermediate variable (maternal FBG). Among non-Hispanic White individuals, the derived racial/ethnic-specific dietary pattern explained, on average, 13.4% of the variability among food groups and 5.2% of the variability in maternal FBG. Among Hispanic/Latina individuals, racial/ethnic-specific dietary findings explained, on average, 6.8% of the variability among food groups and 3.6% of the variability in maternal FBG. In the combined sample, the derived dietary pattern was significantly and positively correlated with maternal FBG (r = 0.26, p < 0.001). To a lesser extent, this same pattern was also significantly and positively associated with maternal FBG among Hispanic/Latina (r = 0.11, p = 0.021) and non-Hispanic White (r = 0.16, p = 0.002) individuals. Lastly, racial/ethnic-specific dietary patterns were also significantly and positively correlated with maternal FBG in their respective racial/ethnic group (non-Hispanic White: r = 0.23, p < 0.001; Hispanic/Latina: r = 0.19, p = 0.001).

Figures 2 and 3 show HEI-2015 scores by derived dietary patterns (tertiles) in the combined and racial/ethnic-specific samples. In general, mean HEI-2015 scores were significantly lower in the highest tertile than the lowest tertile (57.6 versus 49.6, ANOVA p<0.001; r = −0.30, p<0.001) of the dietary pattern derived using data from the combined sample (Figure 2). This trend was similar for this same diet in both the non-Hispanic White (Figure 3A: ANOVA p<0.001; r = −0.31, p<0.001) and Hispanic/Latina (Figure 3C: ANOVA p=0.006; r = −0.16, p = 0.001) samples. In non-Hispanic White individuals, mean HEI-2015 scores were significantly lower in the highest dietary pattern than the lowest dietary pattern (Figure 3B: 57.6 versus 49.6, ANOVA p<0.001; r = −0.56, p<0.001) for the racial/ethnic-specific derived dietary pattern, while among Hispanic/Latina individuals, HEI-2015 scores were significantly higher in the highest tertile than the lowest tertile of the corresponding racial/ethnic-specific dietary pattern (Figure 3D: 52.3 versus 48.6, ANOVA p = 0.003; r = 0.11, p = 0.021).

Figure 2. Healthy eating index (2015) scores were significantly lower with each increasing tertiles of the dietary pattern derived in the combined sample (n=984) of Hispanic/Latina (n=420) and non-Hispanic white pregnant (n=564) individuals^a,b.

^a Dietary patterns were derived in reduced rank regression using maternal fasting blood glucose during pregnancy in Hispanic/Latina and Non-Hispanic White participants, separately, and in a combined sample of both racial/ethnic groups

^b Analysis of variance test, p <0.001; Pearson correlation: r = −0.30, p <0.001

Figure 3. Among non-Hispanic White (Figures A and B, n=564) and Hispanic/Latina (Figures C and D, n=420) pregnant groups, separately, healthy eating index (2015) scores were significantly lower with each increasing tertiles of both the dietary pattern derived in the combined sample and their respective racial/ethnic-specific derived dietary pattern. ^a,b,c.

^a Dietary patterns were derived in reduced rank regression using maternal fasting blood glucose during pregnancy in Hispanic/Latina and non-Hispanic White participants, separately, and in a combined sample of both racial/ethnic groups

^b Analysis of variance test: Figure 3A: p < 0.001; Figure 3B: p < 0.001; Figure 3C: p = 0.006; Figure 3D: p = 0.003

^c Pearson correlations: Figure 3A: r = −0.31, p < 0.001; Figure 3B: r = −0.56, p < 0.001; Figure 3C: r = −0.16, p = 0.001; Figure 3D: r = 0.11, p = 0.021

Table 2 displays the multivariate associations between derived dietary patterns and birth outcomes from multivariate linear and logistic regression models in the combined sample and in each racial/ethnic group. In the combined sample, a one-SD score increase in the derived dietary pattern was statistically significantly associated with lower BW (β = −57.5, 95% CI: −94.8, −20.2), GA at birth (β = −0.13, 95% Cl: −0.24, −0.01), and BW-for-age z-score (β = −0.08, 95% Cl: −0.16, −0.004) and greater odds of LBW (OR = 1.62, 95% CI: 1.07, 2.46) and preterm birth (OR = 1.41, 95% CI: 1.03, 1.94).

Table 2.

Multivariate dietary pattern associations with birth outcomes in Non-Hispanic White (n=564), Hispanic/Latina (n=420), and combined samples (n=984) of pregnant individuals^a

Birth outcomes	Combined (n=984)b	Non-Hispanic White (n=564)c			Hispanic/Latina (n=420)c
	Combined DPd	Combined DP	NHW DPe	SUESTf	Combined DP	Hispanic/Latina DPe	SUEST
	β (95% CIs)	β (95% CIs)	β (95% CIs)	p	β (95% CIs)	β (95% CIs)	p
Continuous
Birthweight (grams)	−57.5 (−94.8, −20.2)	−27.2 (−29.3, −25.0)	−48.5 (−69.8, −27.1)	0.030	−52.9 (−63.9, −41.9)	−51.7 (−63.1, −40.2)	<0.001
GA at birth (wks)	−0.13 (−0.24, −0.01)	−0.03 (−0.19, 0.12)	−0.13 (−0.14, −0.11)	0.200	−0.17 (−0.28, −0.07)	−0.03 (−0.09, 0.02)	<0.001
BW-for-age (z-score)	−0.08 (−0.16, −0.004)	−0.03 (−0.12, 0.06)	−0.05 (−0.11, 0.01)	0.284	−0.05 (−0.12, 0.02)	−0.11 (−0.19, −0.03)	<0.001
	OR (95% CIs)	OR (95% CIs)	OR (95% CIs)	p	OR (95% CIs)	OR (95% CIs)	p
Dichotomous
Low birth weight (<2500 g)	1.62 (1.07, 2.46)	1.14 (0.98, 1.34)	1.21 (1.05, 1.40)	<0.001	1.90 (0.68, 5.36)	0.86 (0.78, 0.95)	0.094
Macrosomia (>4500 g)	0.81 (0.59, 1.11)	0.78 (0.61, 1.00)	0.89 (0.63, 1.24)	0.006	0.83 (0.62, 1.12)	0.58 (0.47, 0.70)	<0.001
Small-for-GA	1.07 (0.79, 1.45)	1.11 (0.87, 1.41)	1.03 (0.95, 1.13)	0.352	0.93 (0.84, 1.04)	1.05 (0.79, 1.40)	0.193
Large-for-GA	0.79 (0.61, 1.01)	0.74 (0.59, 0.94)	0.76 (0.58, 1.00)	0.094	0.89 (0.81, 0.98)	0.70 (0.65, 0.75)	<0.001
Preterm birth (< 37 wks GA)	1.41 (1.03, 1.94)	1.15 (0.80, 1.66)	1.49 (1.43, 1.56)	0.215	1.67 (0.76, 3.70)	1.11 (0.76, 1.64)	0.050

Abbreviations: BW, birthweight; DP, dietary pattern; GA, gestational age; NHW, non-Hispanic White; SUEST, seemingly unrelated estimation.

^aEstimates are expected mean differences (β [95% confidence intervals, CIs]) and percent differences (OR [95% CIs]) of continuous and dichotomous birth outcomes, respectively, associated with a one-standard deviation score increase in the relevant dietary pattern.

^bIn the combined sample, all models were adjusted for maternal age (y), parity (first, second, third or more), pre-pregnancy body mass index (weight [kg] / height [m²]), total daily energy intake (kcals), total physical activity in the 3^rd trimester of pregnancy (metabolic equivalents/week), maternal co-habitation (y/n), maternal education (< high school, ≥ high school), study location (Los Angeles, CA; Aurora, CO), and maternal fasting blood glucose (mg/dL). Bolded values were statistically significant at p < 0.05.

^cIn racial/ethnic-specific samples, all models were adjusted for the same covariates adjusted for in the combined sample, except for study location (Los Angeles, CA; Aurora, CO).

^dDietary pattern (scores) derived in the combined sample

^eDietary pattern (scores) derived in non-Hispanic White- and Hispanic/Latina-specific samples separately.

^fSeemingly unrelated estimation (SUEST) p-values, which indicate significant differences in dietary pattern coefficients from combined versus racial/ethnic-specific multivariate models using study location (Los Angeles, CA; Aurora, CO) as a cluster variable.

Among non-Hispanic White participants, a one-SD increase in the dietary pattern derived in the combined sample was statistically significantly associated with reduced BW (β = −27.2, 95% CI: −29.3, −25.0) and odds of LGA (OR = 0.74, 95% CI: 0.59, 0.94). In this same group, the racial/ethnic-specific derived dietary pattern was statistically significantly associated with lower BW (β = −48.5, 95% Cl: −69.8, −27.1) and GA at birth (β = −0.13, 95% Cl: −0.14, −0.11) and greater odds of LBW (OR = 1.21, 95% CI: 1.05, 1.40) and preterm birth (OR = 1.49, 95% CI: 1.43, 1.56). Among these findings, we observed statistically significant differences in dietary pattern coefficients only for BW and LBW, both of which demonstrated a more pronounced association with the racial/ethnic-specific dietary pattern compared to that of the combined sample (SUEST p = 0.030 and p < 0.001, respectively).

Among Hispanic/Latina individuals, a one-SD score increase in the dietary pattern derived from the combined sample was significantly associated with lower BW (β = −52.9, 95% CI: −63.9, −41.9) and GA at birth (β = −0.17, 95% CI: −0.28, −0.07) and lower odds of LGA (OR = 0.89, 95% CI: 0.81, 0.98). In this same group, a one-SD score increase in the racial/ethnic-specific derived dietary pattern was significantly associated with reduced BW (β = −51.7, 95% CI: −63.1, −40.2) and BW-for-age (β = −0.11, 95% CI: −0.19, −0.03) and lower odds of LBW (OR = 0.86, 95% CI: 0.78, 0.95), macrosomia (OR = 0.58, 95% CI: 0.47, 0.70), and LGA (OR=0.70, 95% CI: 0.65, 0.75). Among these findings, we observed statistically significant differences in dietary pattern coefficients from racial/ethnic-specific dietary pattern compared to that of the combined sample for BW-for-age, macrosomia, and LGA, all of which demonstrated relatively more pronounced associations with the racial/ethnic-specific dietary pattern compared to that of the combined sample (all SUEST p < 0.001). By contrast, the findings for the diet derived in the combined sample were stronger in magnitude than those of the racial/ethnic-specific sample only for BW and GA at birth (both SUEST p < 0.001).

4. Discussion

Among a combined sample of Hispanic/Latina and non-Hispanic White pregnant individuals, we derived and identified a dietary pattern positively linked to maternal FBG during pregnancy that was characterized by higher intakes of refined grains and lower intakes of whole grains, nuts and seeds, and solid fats. Racial/ethnic-specific dietary findings demonstrated slight intake differences for some foods (i.e., fruit, whole grains, and refined grains) between Hispanic/Latina and non-Hispanic White pregnant individuals. Overall, the diet derived using the combined sample was significantly associated with several adverse birth outcomes including lower BW, GA at birth, and BW-for-age z-scores and increased risk of LBW and preterm birth among all participants. Among racial/ethnic groups, this same diet was associated only with lower BW and GA at birth (only Hispanic/Latina group) and lower risk of LGA. We fail to reject our main hypothesis; racial/ethnic-specific dietary patterns predicted continuous birth outcomes better in their respective racial/ethnic groups than a diet derived using a racially/ethnically combined population.

Our RRR dietary pattern findings using maternal FBG during pregnancy were generally consistent with previous work for key foods known to be associated with FBG, including refined grains and whole grains [29]. A previous RRR analysis using maternal FBG and MyPyramid Equivalents food group data in Healthy Start derived a dietary pattern characterized by higher intakes of eggs, potatoes, other starchy vegetables, discretionary solid fat, citrus, melons and berries, and non-whole grains (refined grains) and lower intakes of dairy products, soy, dark-green vegetables, and whole grains [9]. The aggregation of minor into major food groups in our study made it difficult to compare loading coefficients for several food groups used in previous work (e.g., other starchy vegetables). To verify previous findings, we performed RRR with the 24 MyPyramid Equivalents food groups previously analyzed in our combined sample and found generally consistent findings regarding dietary composition linked to maternal FBG, including higher intakes of refined grains, other starchy vegetables, meat, and eggs and lower intakes of yogurt and whole grains (data not shown). However, food groups such as potatoes, discretionary solid fats, and citrus, melons, and berries that loaded highly in previous dietary pattern work had loadings close to 0 in our data (data not shown). These differences in dietary composition may be attributed to varying intake distributions of food groups, which are well known to vary by race/ethnicity. Therefore, the combination of foods that emerge as a dietary pattern in RRR will partly depend on the extent to which food groups are commonly consumed. Lastly, previous dietary pattern findings explained only 5% of the variability in maternal FBG, which was consistent with our findings (7%) and suggests that a large proportion of maternal FBG during pregnancy may be mostly explained by environmental, genetic, and lifestyle factors, not by diet during pregnancy. Another potential explanation is if the predictor variables (e.g., food group intakes) have low variation, then the explained variation in the outcome variable (e.g., maternal FBG levels) may not be a good measure of the maternal diet–FBG relationship [30]. For instance, if intakes of refined grains and whole grains show low variation in the study sample, then the explained maternal FBG variation by diet may be disproportionately low, even though diet could still be a key contributor to maternal FBG levels where sufficient variation in food groups of diet patterns is present. Indeed, the log-transformed intakes of refined grains and solid fats, for example, had narrow and tall distributions (data not shown), indicating relatively low variation. This weak link between derived diets and maternal FBG across all samples in our study may be the reason we did not observe positive associations of dietary patterns with macrosomia or LGA, both of which include hyperglycemia as a risk factor [31].

While the dietary pattern from the combined sample generally reflected the “average” food combinations observed across racial/ethnic-specific patterns, there were slight differences in intake of some foods between racial/ethnic-specific diets. For example, fruit intake appeared to be greater in the Hispanic/Latina-specific derived diet than in the non-Hispanic White-specific pattern. There were also notable differences across racial/ethnic-specific diets for refined grains and whole grains, both of which have been previously known to be associated with FBG [32, 33]. For instance, while lower intakes of whole grains were found to meaningfully contribute to the food composition of the non-Hispanic White-specific diet, higher intakes of refined grains were demonstrated to contribute meaningfully to the Hispanic/Latina-specific diet, but not vice-versa. Differences in food group intakes observed between racial/ethnic-specific RRR-derived dietary patterns may largely be explained by differences in cultural dietary practices (e.g., cooking preparation, common consumption of individual foods and combinations thereof), which may also contribute to varying glycemic effects (e.g., tortilla and beans versus white bread and pasta) and, consequently, to different RRR-derived dietary patterns in each racial/ethnic group.

These differences in diet composition across studies have implications for dietary pattern associations with birth outcomes. For instance, while the previously RRR-derived diet characterized by higher intake of carbohydrates, solid fats, and certain fruits in Healthy Start was significantly associated with increased BW, our findings suggest that a diet with higher intakes of refined grains and lower intakes of whole grains, solid fats, and nuts and seeds may increase the risk of lower BW and several other birth outcomes. These findings underscore the importance of understanding individual and joint effects of dietary components in an overall diet and their implications for birth outcomes. Because traditional regression methods do not allow for the inclusion of highly correlated predictors, future research should leverage flexible machine learning statistical tools, such as Bayesian kernel machine regression, to address high correlations among dietary components to identify key individual and/or joint dietary factors driving observed dietary pattern associations with birth outcomes. Nonetheless, the current findings on dietary patterns are generally in accordance with previous scientific reviews concluding that adherence to dietary patterns generally characterized by relatively higher intakes of vegetables, fruits, whole grains, low-fat dairy, and lean protein foods is significantly associated with lower risk of adverse birth outcomes including preterm birth, whereas adherence to diets generally characterized by relatively higher intakes of refined grains, processed meat, and foods high in saturated fat or sugar is associated with lower BW and trends towards greater risk of preterm birth [4, 34].

The Dietary Guidelines for Americans (DGA) recommends following a healthy dietary pattern at every life stage including pregnancy [35]. The DGA characterizes a healthy dietary pattern as having the following nutrient-dense foods and beverages (no or little added sugars, saturated fat, and sodium), within calorie limits, and in recommended amounts: 1) vegetables of all types (e.g., dark green, red, orange, starchy); 2) whole fruit; 3) grains (1/2 of which are whole grain); 4) fat-free or low-fat dairy (e.g., milk, yogurt, cheese) or non-dairy alternatives; 5) protein foods (e.g., lean meats, eggs, nuts and seeds); and 5) oils (e.g., vegetable oils and oils in seafood, nuts, and other foods). Additional DGA recommendations specific to pregnancy include not consuming any alcoholic beverages, avoiding certain fish species (e.g., mackerel, swordfish) due to toxic exposure to methylmercury, and taking a daily supplement to help meet recommended intakes for folate/folic acid, iron, iodine, and vitamin D that diet alone may not provide [36]. Because RRR-derived diet patterns are intended to be predictive of maternal FBG, we do not expect resulting patterns to follow the DGA. Indeed, the dietary pattern derived in the combined sample showed higher intakes of refined grains and lower intakes of whole grains, nuts and seeds, and solid fats, which do not generally adhere to DGA recommendations.

Our dietary findings related to HEI-2015 verified our expectations regarding the relationship between RRR-derived diets and the DGA recommendations. For instance, the dietary pattern derived in the combined sample was negatively associated with HEI-2015 in the combined and racial/ethnic-specific samples. Based on the diet derived from the combined sample, our findings suggest that a combination of higher intakes of refined grains and lower intakes of whole grains, nuts and seeds, and solid fats in an overall diet reflects poor diet healthfulness. Despite the slight differences in dietary composition between the diet from the combined sample versus the Hispanic/Latina-specific diet, these intake variations, particularly for fruit, whole grains, nuts and seeds, and oils were sufficient to affect the direction of the association with HEI-2015, from negative to positive.

Although the Hispanic/Latina-specific diet could be considered “healthy” by our HEI-2015 standards, this racial/ethnic-specific diet was found to be associated with lower BW and BW-for-age z-scores among Hispanic/Latina individuals. Although the Hispanic/Latina-specific diet suggests a protective effect against infants with macrosomia and LGA in our study, future work identifying specific dietary components in the overall diet that may be driving the observed findings is needed before dietary recommendations/guidance can be made.

A major strength of our study is the use of data from sociodemographically distinct pregnancy cohorts, allowing the inclusion of pregnant individuals from underrepresented backgrounds (i.e., Hispanic/Latina individuals) and subsequent assessment of racial/ethnic-specific dietary differences in birth outcomes. We were able to evaluate the “healthfulness” of derived dietary patterns using a widely used diet quality index previously linked to health [37]. Additionally, 24-h diet recalls have less systematic error than food frequency questionnaires in dietary intake estimates associated with participant recall [38]. Finally, comparison of select study participant characteristics by included/excluded individuals suggested that the potential for selection bias in our study was low (Supplementary Table 3).

A key study limitation is the use of data from different studies with somewhat different sampling and data collection approaches that could not be accounted for in the analysis. For example, in Healthy Start, blood biospecimen collection was typically performed on the same day as dietary assessment, while in MADRES, biospecimen collection was, in some cases, performed before dietary assessment, which may undermine the temporality between diet and maternal FBG in our study. Given the observational nature of our study, we are unable to make causal inferences due to potential unmeasured confounding. Another limitation of our study is the use of maternal FBG as the sole intermediate variable on the pathway between maternal diet during pregnancy and birth outcomes when there may be additional key mediating biological mechanisms (e.g., vascular endothelial growth factor) [39].

5. Conclusion

Despite the small proportion of maternal FBG explained by derived dietary patterns, we found that adherence to a diet rich in refined grains and low in whole grains, solid fats, and nuts and seeds was associated with increased risk for multiple adverse birth outcomes, particularly among Hispanic/Latina pregnant individuals. Despite similar glucose distributions among Hispanic/Latina and non-Hispanic White pregnant individuals, we observed slight differences in dietary components between derived racial/ethnic-specific diets. Racial/ethnic-specific FBG-based diets predicted adverse birth outcomes better than a diet derived in racially/ethnically mixed population. Our findings, therefore, should encourage data-driven dietary pattern studies, particularly in racially/ethnically mixed populations, that derive racial/ethnic-specific dietary patterns. This is important because effect measure modification of dietary associations with birth outcomes by race/ethnicity for a diet derived in a racially/ethnically mixed population may not detect racial/ethnic-specific dietary effects.

Highlights.

• Overall, we found a diet high in refined grains and low in whole grains, solid fats, and nuts/seeds.

• Diet patterns significantly correlated with the 2015 Healthy Eating Index.

• Racial/ethnic-derived diets were compared with diets derived from overall sample.

• Racial/ethnic-derived diets better predicted birth outcomes.

Data Availability Statement:

Select de-identified data from the ECHO Program are available through NICHD’s Data and Specimen Hub (DASH). Information on study data not available on DASH, such as some Indigenous datasets, can be found on the ECHO study DASH webpage.