The effects of dietary and lifestyle interventions among pregnant women with overweight or obesity on early childhood outcomes: an individual participant data meta-analysis from randomised trials

This IPDMA complied with the PRISMA-IPD guidelines and statement (Additional file 1) [6] and was prospectively registered with PROSPERO (ID number CRD42016047165) [7].

Details have been published previously in our protocol [7]. In brief, individual patient data from RCTs in which women with a singleton, live gestation between 10⁺⁰ and 20⁺⁰ weeks, and of BMI ≥25 kg/m² at the time of the first antenatal visit were randomised to receive a diet and/or lifestyle intervention or continued standard antenatal care and in which longer-term maternal and child follow-up at 3–5 years of age had been undertaken were eligible for inclusion. The included studies were identified by a systematic literature search within the International Weight Management in Pregnancy (i-WIP) Collaborative Group collaboration [4] and last updated March 2019 to ensure no additional studies had been overlooked. In addition, we searched PubMed (MEDLINE including ahead of print citations), PubMed Central (including bio-medical and life sciences journals and manuscripts submitted to comply with NIH open access policy), and Embase (includes MEDLINE and additional sources), through Ovid Medline, using search terms for dietary and lifestyle interventions in pregnancy, overweight, obesity, child cohort studies and childhood obesity (Additional File 2). The last search prior to data acquisition and analysis was undertaken in September 2019 and updated until March 2021.

As outlined in our protocol [7], each trial contributed de-identified participant-level data for each participant randomised, stored in a secure database. Variables included baseline descriptive information, allocated treatment intervention, and maternal pregnancy and birth outcomes, and neonatal outcomes were based on the original i-WIP individual participant data (IPD) [4].

Individual trial data, including missing data and randomisation processes, were checked to ensure consistency internally and with published reports. Initially, data from each trial were analysed separately and verified by the individual investigator before being incorporated into the combined database.

An expanded database was created to include child height, weight, BMI, skinfold thickness measurements, calculated percentage body fat and fat-free mass, dietary and physical activity patterns, blood pressure, neurodevelopmental outcome domains and general health.

The primary childhood outcome was a BMI z-score above the 90th percentile, calculated using the WHO Child Growth Standards [8], determined by our Delphi survey [7]. We utilised a standardised online two-stage Delphi survey (February to April 2016) [9], to prioritise clinically relevant childhood outcomes. The panel involved members of the International Weight Management in Pregnancy Collaborative Group (iWIP) collaborative steering committee, members from the planned IPD investigators and other identified multidisciplinary experts in the field. The members scored each outcome using a Likert scale with a score of 9 considered critical, while a score of 1 was considered of limited importance to patient care. Members could suggest other outcomes which were included in the second round along with the highest scoring outcomes from the first round [7].

A range of secondary maternal and childhood outcomes were assessed 3––5 years following birth, as described in our published protocol [5]. Secondary childhood outcome measures included height, weight, BMI, body circumferences (head, abdominal, and mid-upper arm), skinfold thickness measurements (SFTM) (subscapular, triceps, and biceps), fat mass and blood pressure. A parent-completed questionnaire was used to assess child dietary intake, physical activity, screen time and sleep time, with the parent-completed Ages and Stages Questionnaire used to screen child neurodevelopment [10]. Maternal secondary outcomes included weight, waist circumference, BMI, blood pressure and dietary intake assessed by a self-completed questionnaire. A number of additional measures were outlined in the protocol [7] but could not be reliably incorporated into the planned meta-analysis due to the extent of missing data and variable outcome definitions across the individual follow-up studies.

The primary analysis was based on the raw unimputed data. Our original intention was to base conclusions on results from analyses on multiply imputed data. However, issues with the imputation models necessitated the use of the raw data. Firstly, there was a high proportion (50% or more) of missing data for all outcomes, including some which were systematically missing in individual studies. Secondly, because all 3–5-year missing outcome data tended to be missing together, there were few auxiliary variables available to allow meaningful imputation of these outcomes. Additionally, one-stage (random effects) analyses for imputed data exhibited questionable convergence for almost all outcomes and non-convergence for some imputations. The decision was therefore made to use the raw data analyses as the primary analyses, with the imputed analyses as sensitivity analyses.

Analyses for all outcomes were performed using a one-stage (random effects) approach [11]. Mixed-effects models were fitted with fixed study-specific intercepts to allow for different baseline levels of outcome between studies, and a random intervention effect to allow for heterogeneity of intervention effect between studies. Covariates in adjusted models were fitted as fixed effects to avoid over-parameterisation and because there was no reason to expect these effects to differ between studies. Binary outcomes were analysed using mixed-effects log Poisson regression with robust variance, as the originally planned log binomial models did not converge. The effect of the intervention was estimated as a relative risk (RR) of the outcome (intervention vs control) and 95% confidence interval (CI). Continuous outcomes were analysed using mixed-effects linear regression models. The effect of the intervention was estimated as a difference in means (intervention − control) and 95% CI. Following recently published recommendations [12], restricted maximum likelihood (REML) estimation was used for linear regression models, with degrees of freedom calculated according to the Satterthwaite approximation. Analyses were undertaken using Stata v16 (StataCorp, College Station, TX).

To evaluate the robustness of the results of these analyses, a wide range of sensitivity analyses were performed. Firstly, alternative one-stage models were fitted in which study-specific intercepts were specified as random rather than fixed effects. Secondly, multiple imputation of child and maternal anthropometric outcomes was undertaken using two methods currently available for IPD of this nature: the two-stage fully conditional specification method [13], using the mice [14] and micemd [15] packages, and the joint multivariate normal method [16], using the jomo package [17], in R version 3.5 (R Foundation for Statistical Computing). For each imputation method, 100 complete datasets were created. Multiple imputation was performed separately by treatment group, with the imputation models including baseline variables (maternal age, BMI and parity at trial entry), pregnancy and birth variables (total GWG, gestational diabetes mellitus (GDM), gestational age (GA) at delivery, birth weight, length and head circumference (HC)) and child sex and age at follow-up. One-stage analyses of imputed data were performed in R v3.5, using the lme4 package [18] to fit models and the mitml package [19] to extract estimates.

Thirdly, two-stage analyses were performed for both raw and imputed data, in which estimates were first obtained separately for each study then combined using standard random-effects meta-analysis [11]. Between-study heterogeneity was estimated using the DerSimonian-Laird estimator; REML methods were also explored but led to convergence issues for some outcomes. Finally, analyses were performed in which the second intervention groups from the TOP [20] and Bogaerts [21] studies were excluded. The Bogaerts study included a second group receiving brochures and was included in the control group in the main analysis. The TOP study included a second group receiving a physical activity-only intervention and was included in the intervention group in the main analysis.

Subgroup analyses were planned to investigate the possibility of the differential effect of the intervention by maternal early pregnancy BMI category (25.0–29.9 vs ≥30.0kg/m²), parity (0 vs 1+) and ethnicity (Caucasian vs non-Caucasian). It was not possible to perform subgroup analyses by maternal ethnicity, as there were too few participants classed as non-Caucasian in all studies. Additionally, the analysis for maternal early pregnancy BMI had to be modified, as four of the six included studies only recruited women with BMI ≥30.0kg/m², making the planned analysis impractical. The analyses were therefore carried out using BMI as a continuous variable. For parity, one study (ROLO) [22] could not be included for the estimation of the interaction effect, as only women with parity 1 were eligible for this study; however, data from ROLO were used to estimate the effect of the intervention in women with parity 1+.

Subgroup analyses were performed using a 2-stage approach only, due to convergence and collinearity issues when one-step models were correctly specified to separate across-study from within-study interaction effects [23, 24]. Within each study, a regression model was fitted including an interaction term between the subgroup (maternal early pregnancy BMI or parity) and intervention. The interaction effect was estimated, as well as the estimated effect of intervention at each level of the subgroup: for parity, the effect of the intervention on nulliparous and multiparous women and, for maternal BMI, the effect of the intervention at the mean BMI of 33.75kg/m² and for an increase of 5kg/m².

Power calculations were undertaken for the expected sample size and demonstrated acceptable power and coverage even for high levels of between-study heterogeneity. Full details can be found in the published protocol [7].

Each participant in the individual trials and follow-up studies comprising the i-WIP-3 collaboration provided written informed consent to participate, with the data being used for the purposes for which the individual studies had approval. De-identified data were made available by the lead investigators of each trial.

No research participants, patients or members of the public were involved in the conceptualisation of this research study or setting the research question or outcome measures. They were not involved in the planning or implementation of this work, nor were they asked for advice or interpretation of the results.

There were no statistically significant or clinically meaningful differences between the intervention and control groups in relation to any of the maternal anthropometric outcomes (Table 4). Estimates of between-study heterogeneity of intervention effect were small but demonstrated some instability.

These results were not altered for any of the sensitivity analyses. Specifically, there were no significant differences where the second intervention groups from TOP [20] and Bogaerts [21] studies were removed, where random study intercepts were substituted for fixed intercepts, in the 2-step analyses or in analyses of the imputed data (data not shown).

There were no statistically significant or clinically meaningful differences between the intervention and control groups in relation to any of the child diet, activity or development outcomes (Table 5). Estimates of heterogeneity of intervention effect were small for most outcomes, but with some substantial heterogeneity observed for screen time, and for some Ages and Stages scores.

The sensitivity analyses undertaken for these outcomes comprised 2-step analyses, analyses substituting random intercepts for fixed study intercepts and analyses excluding the second intervention groups from TOP [20] and Bogaerts [21] (only for those outcomes which were collected in those studies). The results were not affected by any of these sensitivity analyses.

There was no evidence of an effect of intervention on any of the maternal diet outcomes (Table 6). In general, there was evidence of substantial between-study heterogeneity, reflecting the differences in the questionnaires used to collect these outcomes. Sensitivity analyses—2-stage analyses and analyses substituting random intercepts for fixed study intercepts—did not affect the results.

There was no evidence of a differential effect of intervention by maternal early pregnancy BMI for any of the child anthropometric outcomes. However, several of the interaction terms for maternal anthropometric outcomes at 3–5-year follow-up were statistically significant, including maternal BMI (interaction effect −0.07 (−0.14, −0.00), p=0.044), maternal waist circumference (interaction effect −0.20 (−0.40, −0.01), p=0.044), maternal diastolic blood pressure (interaction effect −0.20 (−0.39, −0.00), p=0.045) and maternal weight change from trial entry to follow-up (interaction effect −0.19 (−0.38, −0.01), p=0.038).

There is some evidence to suggest that as maternal BMI increases, maternal 3–5-year follow-up measures in the intervention group decrease relative to those in the control group. However, overall, the estimates of the difference between the intervention and control groups at the overall mean BMI (33.8kg/m²) were not statistically significant. These results should be interpreted with a high degree of caution as the p values have not been adjusted for multiple comparisons; this is a secondary, exploratory analysis and the effect size is modest.

In sensitivity analyses on imputed data, these interaction effects were observed in data imputed using the multivariate normal (MVN) method, but not in data imputed using the 2-stage FCS method.

There was no evidence of a differential effect of the intervention by parity on any child or maternal anthropometric outcome, in either the main analyses or the sensitivity analyses.

Our findings demonstrate that pre-school aged children born to women with overweight or obesity during pregnancy are themselves at risk of high BMI, with approximately 30% of the cohort having a BMI z-score above the 90th percentile. While many of the individual trials identified pregnancy intervention to be associated with improvements in maternal diet [22, 25, 26, 33, 34], and a reduction in risk of high infant birth weight [25, 35], there was no evidence of an effect on childhood weight, adiposity, or dietary and physical activity patterns at 3–5 years of age. Furthermore, there was no evidence of a persistent difference in maternal weight 3–5 years after pregnancy, despite modest differences in GWG evident from some studies [21, 22, 25, 26, 33]. These findings are robust, with the original trials being conducted in different countries across the globe, and despite considerable variation in terms of the intensity of the intervention ranging from three [22] up to weekly sessions [27] across pregnancy.

By combining and analysing the extensive volume of RCT data available, we have been able to evaluate the longer-term maternal and childhood health outcomes with sufficient statistical power, while avoiding the expense, duplication of effort and inevitable time delays which would have occurred by undertaking another large-scale RCT with pre-specified primary outcomes relating to longer-term maternal or child health. The sample size of 2529 child participants represents the largest prospectively collected data set available from participants of randomised trials during pregnancy, with a standardised assessment of anthropometric measures, and consistent evaluation of dietary, physical activity, sedentary behaviour and sleep patterns, all of which are well-recognised early life factors contributing to child overweight and obesity [36].

Our study is not without limitations. Despite the agreement by investigators of pre-specified outcomes, and the a priori generation of our protocol [7], there were a number of measures that could not be reliably incorporated into the meta-analysis due to the extent of missing data and variable outcome definitions across the individual follow-up studies. Furthermore, there is a potential risk of selection bias. Of the total eligible randomised cohort, there was a considerable variation in the proportion of children assessed and who contributed data, ranging from approximately 34% [20, 26] up to 67% [25]. However, baseline and clinical characteristics of women and children for whom data were available and who participated in the follow-up studies were similar between the randomised intervention and control groups and also similar to the full randomised cohort. Sensitivity analyses were conducted under a wide range of different scenarios, with the findings consistent under a variety of plausible assumptions. On balance, therefore, we do not consider the risk of bias to be significant, and any potential impact on the validity of our findings is low.

A further potential limitation of our trial is the generalisability and external validity of our findings. Across all trials, the population was approximately 90% Caucasian, precluding our ability to evaluate the role of maternal ethnicity as we had originally proposed [7]. This continues to be a limitation, with the available randomised trial literature to date predominantly recruiting women who are Caucasian [4].

While appropriate statistical methodology to evaluate IPDMA continues to evolve, we utilised approaches as recommended in the most recent literature and conducted sensitivity analyses where there was any question about the most appropriate methods (for example, in the imputation of missing data). Nevertheless, we encountered numerous challenges in implementing these methods in practice, from imputation for a small number of trials where a very large proportion of data are missing and auxiliary data are not consistently available, to convergence issues with one-stage meta-analysis models.

We are aware of a number of trials [22, 26, 27] that have conducted and reported findings of childhood follow-up at 6 months [37‐39], 18 months [31, 40‐43] and 3–5 years [30, 44] after birth. Together, findings from the individual studies alone and when incorporated into the IPDMA suggest little longer-term effect on child BMI and adiposity measures.

However, we have demonstrated that children born to women with overweight or obesity during pregnancy themselves remain at risk of early childhood overweight and obesity. This cohort of pre-school aged children has a prevalence of BMI z-score above the 90th percentile of approximately 30%. This is in contrast to data reported from the broader childhood population in Australia [45] and Europe [46] where a combined total of 20 [45] to 30% [46] of pre-school aged children are overweight or obese.

Our study also demonstrates the frequent occurrence of obesogenic behaviours, even at age 3–5 years, with the majority of children not meeting the recommended number of daily servings of vegetables, while exceeding both fruit intake and discretionary food intake [47]. Furthermore, the majority of children who contributed data to this IPD-MA did not meet physical activity recommendations of at least 3 h per day and were at the upper range of the 1 h screen time per day [48].

These findings are broadly consistent with dietary intake data from Australian children aged 4–8 years who have similarly poor consumption of vegetables in particular, while far exceeding intake of calorie-dense discretionary foods [49]. The health benefits of fruit and vegetable consumption, even from an early age, are well recognised [50], with their consumption from infancy [51] contributing to persistent sub-optimal eating habits in later childhood and adolescence [51].

Current clinical recommendations internationally advocate intervention in pregnancy [52] through improved diet and limiting gestational weight gain. While healthy diet and physical activity in pregnancy are prudent [53], a significant paradigm shift is required if maternal and child health is to be improved, particularly in relation to child obesity. A continued focus on intervening in pregnancy and a relentless search for the illusory effective pregnancy dietary and/or lifestyle intervention is unlikely to be successful in light of the amassed randomised trial evidence both during pregnancy [4] and now extending into childhood.

A timely opportunity exists in which to refocus research efforts towards ensuring optimal maternal health and weight prior to conceiving, as well as ongoing evaluation of the role of early childhood interventions. This will undoubtedly be difficult to enact and requires a truly multi-disciplinary life-course approach involving systems spanning childhood and adolescence, and commencing at a time well before pregnancy is contemplated [54].