Protein intake during pregnancy and offspring body composition at 6 years: the Generation R Study

Out of 8976 women enrolled in the Generation R Study while pregnant, we restricted our analysis to women of Dutch ancestry (n = 4101). Ancestry was self-reported and defined according to the classification of Statistics Netherlands [18]. We excluded women with missing dietary information (n = 542), women with multiple pregnancies (n = 53) or no live childbirth (n = 24), and women who were lost-to-follow-up (n = 3). In our population of analysis, we included only mother–child pairs with available childhood body composition information at the age of 6 years (n = 2694; Supplemental Fig. 1).

Protein intake during pregnancy (i.e., total, animal, and vegetable protein) was assessed with a 293-item semiquantitative food-frequency questionnaire (FFQ) [19] that women received at enrollment at median 13.4 (IQR 12.2–15.5) weeks of gestation. The FFQ covered the average dietary intake of a Dutch diet over the previous three months. The average daily intake of energy, protein, and other nutrients was calculated using the Dutch food composition table 2006 [20]. Validation of the FFQ against three instances of 24-h dietary recall in 71 pregnant women of Dutch ancestry living in Rotterdam showed an intra-class correlation coefficient of 0.65 for energy-adjusted protein intake. There was no indication for systematic measurement error (Supplemental Fig. 2).

Children visited the research center at a mean (±SD) age of 6.1 ± 0.4 years. We measured height (using a Harpenden stadiometer) and weight [using an electronic personal scale (Seca^®)] to calculate their BMI (kg/m²). This BMI was also used to classify overweight status according to age- and sex-specific cutoffs [21].

During this visit, body composition was measured by dual-energy X-ray absorptiometry (DXA; iDXA; General Electrics-Lunar, 2008, Madison, WI, USA) following standardized procedures [22]. The DXA scanner calculated fat, lean, and bone mass of the total body and of specific body regions, using enCORE software (version 13; GE Healthcare). Fat-free mass index [FFMI (kg/m²); calculated as total fat-free mass divided by height squared] and fat mass index [FMI (kg/m²); total fat mass divided by height squared] were calculated. Additional outcome measurements were lean mass index (kg/m², total fat-free mass minus total bone mass divided by height squared), total fat percentage (total body fat mass divided by total body mass times 100 %), and android/gynoid fat mass ratio (android fat mass divided by gynoid fat mass). All body composition outcomes were standardized for age and sex of the child and analyzed continuously.

At enrollment, we collected information by questionnaire on maternal age, educational level, folic acid supplement use, and parity. Additionally, weight and height were measured at the research center at enrollment to calculate BMI, and a fetal ultrasound was performed to establish gestational age. Energy, fat, and carbohydrate intake during pregnancy were measured using the FFQ described previously. Smoking and alcohol use were assessed during each trimester by questionnaire and categorized into never users, stopped when pregnancy was known, and continued use during pregnancy. Gestational weight gain (g/week) was calculated by subtracting maternal weight at enrollment from the weight in early third trimester and divided by the follow-up duration (weeks).

At birth, we collected information on gestational age at birth, birth weight Z score, sex, and hypertensive pregnancy complications (i.e., preeclampsia and pregnancy-induced hypertension) from delivery reports [23]. Preterm birth was defined as childbirth before 37 weeks of gestation. Breastfeeding practice at 2 months was assessed by a combination of delivery reports and questionnaires. Child protein and energy intake were measured using an FFQ at a median age of 12.9 (IQR 12.6–14.1) months in a subgroup of our population (n = 1591). At the age of 6 years, information on screen time (<2 vs. ≥2 h/day [24]) and participation in sports (yes/no) of the children was collected using a questionnaire.

Maternal protein intake was adjusted for total energy using the nutrient residual method to evaluate the effect of maternal protein intake independent of energy intake and to reduce the magnitude of measurement error [25]. We categorized protein intake into quartiles and used the lowest quartile (Q1) as the reference category. Because of skewed distributions, total body fat percentage and android/gynoid fat mass ratio were natural-log transformed.

We used multivariable linear regression models to assess the associations of maternal total, animal, and vegetable protein intake with childhood body composition measurements. Multivariable logistic regression models were used for childhood overweight.

The analyses were performed with energy-adjusted protein intake during pregnancy. Analyses with animal protein intake were adjusted for vegetable protein intake, and vice versa (model 1). Additionally, for the outcomes, total fat percentage and android/gynoid fat mass, model 1 included also height of the child. The decision to include confounders in the multivariable regression models (model 2) was based on previous literature or a >10 % change in the effect estimate in model 1. The following confounders were considered: maternal age, educational level, parity, smoking and alcohol consumption in pregnancy, folic acid supplementation, maternal BMI at enrollment, energy intake, carbohydrate intake, gestational age at birth, breastfeeding, childhood sedentary time, and childhood physical activity. The confounders included are listed in the footnotes of the figures and tables. Potential intermediate factors were added to a separate multivariable model (model 3), namely gestational weight gain, hypertensive complications during pregnancy, and birth weight Z score. Effect modification was evaluated for gestational weight gain and child sex. In case of significant effect modification (p value for interaction term <0.05), stratified analyses were performed.

To evaluate whether the observed associations were due to a higher protein intake rather than a lower intake of another macronutrient, we assessed whether substituting protein with other macronutrients (e.g., carbohydrates and fat) had any effect on our results [26]. For example, the substitution model for replacing protein by carbohydrates included the macronutrients [in energy percent (E%)] protein, fat, and alcohol, but not the macronutrient carbohydrate. As a result, the regression coefficients for protein from these models reflect the effect of replacing 1 E% from carbohydrates with 1 E% of protein.

To evaluate the robustness of our findings, several secondary analyses were performed. First, we further adjusted our models for protein intake of the children (n = 1591). Second, we restricted analyses to women with a child born after 37 weeks of gestation, those without hypertensive complications in pregnancy, and to children with a normal birth weight (which we defined as a gestational age- and sex-adjusted birth weight between ±2 SD). Also, we excluded siblings (n = 185), and finally, we did not include the covariate child height in the multivariable model 2 since height might also be associated with obesity [27].

To reduce bias due to missing data, missing covariates (0–17.7 %) were imputed using multiple imputation which includes fully conditioned specification of the imputation. Ten imputed datasets were created, and the analyses were performed in each dataset before the results were pooled by Rubin’s rules [28] taking into account uncertainty with the prediction of missing data. Details on the imputation procedure are described in Supplemental Table 1. All statistical analyses were performed in SPSS version 21.0 (IBM Corp., Armonk, NY, USA).

Maternal and child characteristics are presented in Table 1. The main sources of protein in our study population were dairy products, meat and meat products, and nuts and seeds (together explaining 60 % of the variance in total protein intake). Mothers with a higher protein intake were, on average, older, had greater levels of education, more often non-smokers, and used more frequently folic acid supplementation than those with a lower protein intake (Supplemental Table 2).

Children of mothers in the highest quartile (Q4) of protein intake did not have a statistically significant higher BMI than children of mothers in the lowest quartile (Q1; Table 2). Both animal protein and vegetable protein intake were not associated with a higher childhood BMI in model 1, whereas higher animal protein as well as higher vegetable protein intake was associated with higher childhood BMI after adjustment for confounders (model 2, Table 2).

Higher total protein intake during pregnancy was associated with a higher FFMI in children aged 6 years in the unadjusted model (model 1) as well as in the multivariable adjusted model (model 2, Table 2). The effect estimates were comparable for animal and vegetable protein (Table 2). Maternal total protein intake remained significantly associated with childhood FFMI [difference 0.15 SD (95 % CI 0.04, 0.25) for Q4 vs. Q1, p value for trend = 0.01] after additional adjustment of the potential intermediate factors gestational weight gain, hypertensive complications, and birth weight (model 3). Total maternal protein intake was not associated with childhood FMI and neither was the intake of animal protein during pregnancy (Table 2). Vegetable protein intake was associated with lower FMI in model 1, but this association did not remain after adjustment for lifestyle factors and sociodemographic background (model 2, Table 2).

In Supplemental Table 3, we added the results from the main analyses that were performed in a non-imputed dataset. The magnitude of the effects was stronger in the imputed data, which was mainly seen in the analysis for vegetable protein intake. Nonetheless, the 95 % confidence intervals from the imputed and non-imputed data largely overlapped and were not statistically significantly different.

Total protein and vegetable protein intake, but not intake of animal protein, were associated with lower childhood total fat percentage in the unadjusted analysis (model 1), but after adjustment for confounders (model 2), these associations did not remain (Supplemental Table 4). Vegetable protein intake was associated with a lower android/gynoid fat mass ratio in the unadjusted model only, and no association was found with total protein or animal protein intake and android/gynoid fat mass ratio (Supplemental Table 4).

We did not observe specific substitution effects when protein (E%) was exchanged for different types of macronutrients in the association with FFMI (Table 3). Additional adjustment for protein intake of the child at 14 months of age (n = 1558; 59 %) slightly attenuated the results with FFMI (Supplemental Table 5).

Strengths of this study are the prospective population-based design, the large sample size, the postnatal follow-up of the offspring through 6 years of age, and the collection of numerous confounding factors. A further strength is the detailed information we collected with regard to body composition measurements, since DXA has a high accuracy of measuring fat mass and other soft-tissue body composition components [32, 33].

However, some limitations should be considered when interpreting our results. A limitation of our study is the measurement of protein intake using an FFQ, which is not very precise. However, FFQs have been shown to be accurate in ranking participants according to their intake [34], and energy adjustment may have reduced the magnitude of measurement error [25]. Also, there might be measurement error of the anthropometric measurements (i.e. BMI during pregnancy) and other covariates. However, since these are measured before the objective outcome measurement (i.e., body composition of the child), this measurement error is most likely non-differential and not lead to differential associations between dietary protein intake during pregnancy and body composition of the child. Since we had no data on maternal physical activity during pregnancy, residual confounding due to maternal physical activity levels could influence our results. A third limitation is the restriction to women of Dutch ancestry in our analyses within this multi-ethnic prospective cohort study. While the inclusion of other ethnicities could have led to differential misclassification of dietary intake [35], the restriction with regard to ethnicity may reduce the external validity of our results. Also, the Generation R Study consisted of a higher percentage of women with higher socioeconomic status than those that were eligible to participate [17]. However, such a selection bias has not been found to influence exposure–outcome associations [36]. Finally, we did not have complete data on all included covariates; the percentage of missing covariates ranged between 0.0 and 17.7 %. We used multiple imputation procedure to impute these missing covariates since complete case analysis would result in a considerable loss of information and may lead to biased estimates [37].