Study design, randomisation and participants
UPBEAT was a multicentre RCT of a complex behavioural intervention of diet and physical activity advice versus standard antenatal care in obese pregnant women to prevent GDM and delivery of LGA neonates [10]. It involved eight inner-city centres in London (three centres), Bradford, Glasgow, Manchester, Newcastle and Sunderland. Approvals were obtained from the UK research ethics committee (UK integrated research application system, reference 09/H0802/5) and local Research and Development departments in participating centres approved participation of their respective centre; all women provided written informed consent prior to entering the study. UPBEAT is registered with Current Controlled Trials, ISRCTN89971375.
UPBEAT recruited and randomised 1555 obese women (BMI ≥ 30 kg/m2), aged 16 years or older and with a singleton pregnancy between 15+0 and 18+6 weeks of gestation (hereafter weeks). Exclusion criteria were multiple pregnancy, current use of metformin, unwilling or unable to provide written informed consent, or underlying medical disorders [10]. Women were randomised using an internet-based, computer-generated sequence that ensured concealment. In order to reduce differences between groups occurring by chance, the randomisation procedure included minimising by age, ethnic origin, centre, BMI and parity [10]. During pregnancy, women were followed up at 27+0 to 28+6 weeks (when an oral glucose tolerance test (OGTT) was completed) and at 34+0 to 36+0 weeks.
For the purposes of this study, women from two centres (King’s College Hospital, London, and Sunderland) were excluded as no blood samples were taken from participants in these centres for resource and logistical reasons (n = 360 in total; 178 and 182, respectively, from control and intervention arm were excluded). From the remaining six centres, all women with at least one metabolic profile assessment were included. Of the 1194 from the remaining six centres, 1158 (97%) had at least one set of nuclear magnetic resonance (NMR) metabolic measurements together with complete data on all covariables and were included in our analyses; proportions included were similar for the intervention and control groups (Fig. 1).
Metabolic profiling
Venous blood samples were taken on three occasions – at recruitment, prior to randomisation (median (IQR) gestational age, 17.0 (16.1–17.9) weeks), post-randomisation at the time of the OGTT (27.7 (27.3–28.1)), and during the third trimester (34.7 (34.3–35.1)). Samples taken prior to randomisation and in the third trimester were non-fasting; those taken as part of the OGTT were after an overnight fast. All blood samples were initially kept on dry ice, processed within 2-hours and then stored at – 80 °C until metabolic profiling was performed. Metabolic analyses were undertaken on serum in two batches (March 2015 and August 2015), with samples from the three timepoints randomly distributed across these batches. Sample processing was automated with a Gibson 215 liquid processor. The complete process is illustrated in Additional file 1: Figure S1. A total of 158 metabolic features were measured and quantified using an NMR targeted platform (http://www.computationalmedicine.fi/), including 129 lipid measures (lipoprotein particle subclasses, particle size, cholesterols, fatty acids and apolipoproteins), 9 glycerides and phospholipids, and 20 low-molecular weight metabolites including branched-chain and aromatic amino acids, glycolysis-related markers, and ketone bodies [12]. This platform has been used in several large-scale epidemiological studies [4, 13,14,15,16] and further details and a full list of the 158 metabolites are presented in Additional file 1: Figure S1, Box S1 and Table S1. All blood samples were processed by laboratory technicians blinded to participant data, including the allocated randomisation arm.
Statistical analyses
Full details of the statistical modelling are provided in Additional file 1: Box S1. Multilevel (random intercept and random slope) models were used to analyse the repeatedly assessed metabolic traits [17, 18]. We restricted the timeframe of these models to 16–36 weeks so that we were not predicting beyond available data. These models provide an individual (predicted) level of each metabolite at 16 weeks (the intercept) and an individual slope, which we present as the change in each metabolite per 4-week increase in gestational age between 16 and 36 weeks. In all analyses, we controlled for the minimising variables used in randomisation (BMI, ethnicity, parity, age and clinic centre) [10]. An interaction term between time (gestational weeks) and randomised arm (control or intervention) was also included.
We focus our discussion and interpretation of all results on the magnitude of the point estimates (i.e. pregnancy change in metabolites or effect of the intervention) and their precision (i.e. 95% confidence intervals) as recommended by the American Statistics Society and others [19,20,21]. We explore the role of chance by providing p values after controlling for multiple testing using the false discovery rate using the method of Benjamini and Hochberg [22].
Change in metabolic profiles across pregnancy in obese women in the control arm
We present the change in metabolic profiles between 16 and 36 weeks in standard deviation (SD) units in women who were randomised to the control arm of the trial. SD units were used to aid comparison of results between metabolites and with those from other studies. We also present the mean absolute differences in the original units of measure for each metabolite (mostly mmol/L) in the control arm. The full model results (mean intercept and mean slope per 4 weeks) for each metabolite in their original units are also presented for all 1158 women included in the analyses. As the model includes a term for the randomised arm for each woman, these can be interpreted as the mean level of each metabolite at 16 weeks, and its change per 4 weeks of gestation between 16 and 36 weeks having adjusted for any effect of the intervention. The slope is therefore an indication of mean rate of change in metabolites in obese women in general (i.e. without any intervention effect).
Effect of the intervention on change in metabolic profiles across pregnancy
In the main analyses, we present the effect of the intervention on rate of change in metabolites in SD units per 4 weeks (using the SD value of the control group), for ease of interpretation and to enable the effects of the intervention to be compared across different metabolic measures. We also present the difference in mean rate of change in original units (mostly mmol/L per 4 weeks). We present exact p values for all results and focus our discussion of the effect of the intervention on point estimates and confidence intervals as recommended by the American Statistics Society [19]. We also indicate the role of chance by indicating, in figures, which results reached conventional p < 0.05 levels of statistical significance after controlling the false discovery rate using the method of Benjamini and Hochberg to deal with multiple testing [22].
Analysis assumptions and sensitivity analyses
Repeat metabolite assessments occurred at three time-points within a narrow range of gestational ages, such that there are gaps of up to 10 weeks with no (or very little) data between each of the measurements (Additional file 1: Figure S2). Therefore, we had to use linear spline methods and could not explore smoothing methods or use fractional polynomials to determine the exact shape of metabolic trait change over pregnancy [18]. Furthermore, our main analyses assume that the effect of the intervention is consistent between the first two measurements (~16–28 weeks) and the second two (~28–36 weeks). To test this assumption, we modified the multilevel model to include the possibility that the magnitude of metabolite change might alter at 28 weeks, and visually compared the effect of the intervention for each trait between 16–28 and 28–36 weeks. The linear spline method we have used assumes the model residuals to be approximately normally distributed, which may not be the case. There is evidence that estimates of population average change, such as those we present here, are robust to non-normality in the residuals (for example, see [23]); however, these methods may not deal with skewness. Therefore, to explore this further, we have repeated our analyses of the effect of the intervention on differences in mean change of the metabolites using generalised estimating equations with robust standard errors and unstructured correlation matrices, which should be robust to non-normality and (some) misspecification of the working correlation matrices. The robust standard errors were calculated using the White–Huber sandwich estimator and are robust to heteroskedasticity. We undertook a set of 14 further sensitivity analyses using different approaches to examine the sensitivity of our conclusions to possible heteroskedasticity, skewness and data outliers. Additional details of the linear spline model and its assumptions, including assumptions related to missing repeat measurements, and these additional 14 sensitivity analyses are provided in Additional file 1: Box S1.
Indirect comparison with metabolic profiles in unselected ‘healthy’ pregnancy
We were keen to compare our findings in obese pregnant women to those in women not selected for being obese. As all of the participants in our study were selected for being obese, we were only able to do this indirectly by searching the literature for other studies of similar metabolite profiles in general populations of pregnant women. We identified one previous study, by Wang et al. [4], that examined cross-sectional differences using the same NMR metabolic profiles between women of reproductive age who were pregnant and those who were not (n = 4260 women; 322 of whom were pregnant). In addition to those cross-sectional analyses, longitudinal change in the metabolites was assessed in a subgroup of women (n = 583) who were either pregnant at baseline and not at a follow-up assessment 6 years later, or were not pregnant at baseline and were so 6 years later; the study also compared results separately by trimester of pregnancy. We compared the magnitude of longitudinal change and differences by trimester using summary data from Wang et al. (specifically the results shown in Figs. 1, 2, 3 and 4 of that paper) [4] with our results to obtain some insight into whether pregnancy-related metabolic change differed between obese and non-obese women.