Social inequalities in pregnancy metabolic profile: findings from the multi-ethnic Born in Bradford cohort study

BiB is a population-based prospective pregnancy cohort that included 12,453 women who experienced 13,776 pregnancies between 2007 and 2011 [19]. Most women were recruited at approximately 26–28 weeks gestation at their oral glucose tolerance test, which is offered to all women booked for delivery at Bradford Royal Infirmary. As blood samples for metabolic profiling were those collected for the oral glucose tolerance test, all participants in this study were recruited at ~ 26–28 weeks. BiB has almost an equal split of White European and South Asian women, all residing in Bradford, UK, a city in the North of England with high levels of socioeconomic deprivation (the BiB study was started due to a high prevalence of poor child health in the city). Mothers, and their partners, recruited into the study provided detailed interview questionnaire data, measurements, and biological samples. The study website gives further information, including protocols, information on data access, and a list of all data (https://​borninbradford.​nhs.​uk/​research/​documents-data/​).

For this study, we included all first enrolled pregnancies to White European and South Asian women (the two main ethnic groups in BiB). After excluding other ethnicities, and those with missing data on SEP, metabolic traits, and gestational age at measurement of metabolic traits, our analysis sample comprised of 3,905 White Europeans and 4,404 South Asians (Fig. 1).

Ethnicity was reported by the mother at the recruitment questionnaire interview or abstracted from medical records (for those missing questionnaire data) and defined according to the UK Office for National Statistics guidelines. For our main analysis, ethnicity groups were defined as White European or South Asian. White European ethnicity included women that indicated they were White British (n = 4,489) or other White European (n = 306). South Asian ethnicity included women that indicated they were Pakistani (n = 5,128), Indian (n = 439), Bangladeshi (n = 263) or other South Asian heritage (n = 63).

A total of 19 individual-, household- and area-based indicators of SEP were used to derive a composite SEP latent variable (Table 1). All individual- and household-based indicators were reported by the mother at the recruitment questionnaire interview (at around 26–28 weeks of gestation) [20], and the area-based indicator was based on linkage to the mother’s residential address at the time of recruitment.

Area SEP was based on the English Index of Multiple Deprivation (IMD) Score in 2007 (i.e., at the time of pregnancy). IMD is a relative composite measure of multiple deprivation at small area level across England (mean population size in each area is 1500 residents). The domains used to derive IMD in 2007 were income deprivation; employment deprivation; health deprivation and disability; education deprivation; crime deprivation; barriers to housing and services deprivation; and living environment deprivation.

Women were coded as being in receipt of means tested benefits (i.e., benefits awarded to those that can demonstrate that their income and savings are below a certain level) if they reported receiving any of income support, income tested jobs seekers allowance, working families tax credit, or housing benefit. Women were asked how well they are managing financially with responses being either living comfortably, doing alright, just about getting by, or quite difficult or very difficult. Women were also asked how they are doing financially compared to a year ago, with responses coded as better off, worse off, about the same, or does not wish to answer. Those coded as does not wish to answer were excluded from this variable. Women also reported if they were able to have two pairs of all-weather shoes, money to make regular savings of £10 a month, and a small amount of money to spend each week on themselves.

Household-based indicators reflected questions about housing tenure, overcrowding, and ownership of material items and goods based on questions from the Households Below Average Income Survey. Housing tenure was reported as one of seven groups; owns outright, owns with a mortgage, lives rent free, owned by a private landlord, living in social housing, other and don’t know. Those coded as other, don’t know, or living rent free were excluded from this variable. Those coded as living rent free (n = 181 (4.7%) White European; n = 437 (10.2%) South Asian) were excluded because they likely comprise a heterogeneous group of women living in multigenerational homes related to cultural factors, and those living with parents because of financial or other difficulties. Responses to questions on numbers of household members and bedrooms were used to derive an indicator of overcrowding based on the person per room approach [21] by dividing the number of persons by the number of bedrooms in this household. Women were also asked whether they were up to date with household bills, if they had contents insurance, enough money to keep the home in a decent state of repair, money to replace any worn out furniture, money to replace or repair major electrical goods, and if they were able to keep their home warm enough in winter. For each of these variables, women that responded as don't want/need, doesn't wish to answer, or don't know were excluded.

Full details of all metabolomic measurements undertaken in BiB have been published [22]. In this study we focus on maternal pregnancy nuclear magnetic resonance (NMR) metabolic traits. Women had a fasting mid-pregnancy serum sample taken by trained phlebotomists working in the antenatal clinic of Bradford Royal Infirmary (92% were obtained between 26–28 weeks gestation). Samples were processed within 2.5 h and placed in -80° freezers. There were no sample freeze–thaw events prior to their use for metabolomic profiling. In total, 227 metabolic traits were measured using a high-throughput targeted NMR platform (Nightingale Health©, Helsinki, Finland). The metabolic traits were quantified in absolute concentration units or ratios and included circulating lipoprotein lipids and subclasses, fatty acids, fatty acid compositions, amino acids, traits related to glycolysis, ketone bodies, fluid balance, and an inflammatory marker [23, 24]. In this study, derived measures and ratios were excluded, leaving 148 metabolic traits for analysis (Additional File 1: Data Set 1). Gestational age at serum sample collection was recorded.

Latent class analysis (LCA) was applied to all 19 SEP indicators to derive a composite SEP latent variable consisting of SEP sub-groups (latent classes). LCA is a finite mixture model that classifies individuals into unobserved sub-groups (called latent classes) based on their responses to two or more indicator variables, with the aim of identifying subgroups where individuals are more similar within groups than between groups [25]. LCA was done in White European and South Asian women (combined) with data on pregnancy NMR metabolic traits, gestational age at the metabolic traits’ sample collection, and at least one SEP indicator, with estimation by full information maximum likelihood (FIML). To avoid local maxima solutions (i.e., convergence to a likelihood value that is not the global (true) likelihood), we used 1500 random sets of starting values for the initial stage, 150 final stage optimizations, and 15 initial stage iterations. Models with two to six latent classes were compared and the optimal number of classes was identified based on a combination of BIC, entropy statistic, and Lo-Mendell-Rubin adjusted likelihood ratio test (Additional File 2: Table S1). Where these indicators disagreed, the more interpretable model was selected.

Linear regression models with robust standard errors were then used to examine associations between SEP latent class sub-groups (versus a reference SEP sub-group) and each metabolic trait. Models were weighted by the sum of latent class probabilities to allow for uncertainty in SEP latent class membership assignments and were adjusted for gestational age to control for gestational age-related differences in metabolic traits. Ethnic differences in the associations between SEP and metabolic traits were examined by including an interaction term between SEP and ethnicity (i.e., White European, or South Asian) in all models. All metabolic traits were standardised (to mean = 0, SD = 1) to aid comparison of results for different metabolic traits [26]. Standardisation was done separately by ethnicity because we have previously found differences in pregnancy metabolic traits between White European and South Asians in this cohort [14]. In a sensitivity analysis, we repeated LCA and regression modelling separately in White European and South Asian women and separately in the White British and Pakistani women (the two biggest groups within the White European and South Asian ethnic groups). This was done because of previously reported SEP differences between South Asian groups and between White European groups in this cohort [15].

To avoid overloading the main paper, here we present results for major groups of metabolic trait and sub-particles regardless of P-values, and provide all results with exact P-values and false discovery rate (FDR) corrected P-values [27] in additional files. All the results can also be viewed on the accompanying interactive app (https://​aelhak.​shinyapps.​io/​SEP_​NMR_​BiB/​). LCA was done in Mplus version 6, and all other analyses were done in R version 4.2.2.

All women with ≥ 1 SEP indicator were included in the LCA using FIML under the missing at random assumption (i.e., that the probability of a missing SEP indicator value can be entirely explained by other observed SEP indicators and so is not related to its value). For the regression analysis, women with missing data on metabolic traits and gestational age were excluded. To explore the potential impact of missing data, we compared characteristics of included women with those excluded due to missing data (Additional File 3: Table S2).

Following feedback on previous versions of this work presented at scientific conferences and scientific meetings, we decided to make the combined LCA analysis our focus instead of the ethnicity-specific analyses. We decided to analyse traits in SD units instead of perfuming log transformation because most pregnancy metabolic traits were normally distributed (Additional File 4: Fig. S1). No other changes were made to the analysis plan [18].

A total of 3,905 White European and 4,404 South Asian pregnant women with at least one SEP indicator and data on metabolic traits and gestational age at collection of serum samples for the assessment of metabolic traits were included in the study (Fig. 1). Mean gestational age was 26.6 weeks (SD = 1.8) in White Europeans and 26.7 weeks (SD = 1.9) in South Asians, and mean age was 26.6 (SD = 6.0) and 27.9 (SD = 5.2) years respectively. When compared with included women, those excluded due to missing data on metabolic traits and gestational age (n = 1,407) had higher proportion of South Asian ethnicity (63% versus 53%) and broadly similar socioeconomic circumstances as indicated by similar levels across most SEP indicators (Additional File 3: Table S2).

LCA (in the combined sample of White Europeans and South Asians) identified five SEP sub-groups which we have labelled ‘Highest SEP’, ‘High-Medium SEP’, ‘Medium SEP’, ‘Low-Medium SEP’, and ‘Lowest SEP’. The proportions of White Europeans and South Asians in the Lowest SEP, Medium SEP, and the Highest SEP groups were broadly similar, but there were fewer White Europeans than South Asians in the Low-Medium SEP group (29% vs. 79%) and more in the High-Medium SEP group (77% vs. 23%) (Fig. 2). The differentiation into SEP sub-groups was driven largely by five SEP indicators (mother’s educational level and employment status, partner’s educational level and occupational class, and means tested benefits). There was little difference between SEP sub-groups in whether women reported being able to keep the home warm enough in winter and being able to afford to afford two pairs of all-weather shoes. The remaining 12 indicators each contributed with modest differences. When compared with High-Medium SEP subgroup, the Low-Medium SEP sub-group was more likely to own a house outright without a mortgage (Fig. 2).

Lower SEP was associated with (mostly) less favourable levels of 113 metabolic traits at the FDR corrected P < 0.05 threshold (Additional File 5: Data Set 2, Additional File 6: Data Set 3). This included associations between lower SEP and higher VLDL cholesterol, total triglycerides (Fig. 3), glycoprotein acetyls, and VLDL concentration, lower levels of cholines (Fig. 4), and higher VLDL and lower HDL in cholesterol and phospholipids (Fig. 5). Conversely, there was less evidence of associations with LDL particles and no differences in albumin, glycine, histidine, or lactate (Additional File 6: Data Set 3).

There was a statistical interaction between SEP and ethnicity (at the FDR corrected P < 0.1 threshold) for 31 metabolic traits (Additional File 5: Data Set 2). For most of these traits, differences by SEP group were larger and showed a clearer gradient (across SEP categories) in White European than South Asian women (Additional File 7: Data Set 4). This included stronger association with omega − 3 fatty acids, cholesterol and triglycerides in large HDL, docosahexaenoic acid, and degree of unsaturation in White Europeans (Fig. 6). Differences in HDL and VLDL particle size were larger in White Europeans but the difference in LDL particle size was larger in South Asians (Additional File 7: Data Set 4).

Ethnicity-specific LCA identified five sub-groups in White Europeans and three in South Asians. SEP sub-groups in White Europeans were labelled ‘Highest SEP’, ‘High-Medium SEP’, ‘Medium SEP’, ‘Low-Medium SEP’, and ‘Lowest SEP’, and SEP sub-groups in South Asians were labelled ‘Highest SEP’, ‘Medium SEP’, ‘and ‘Lowest SEP’ (Fig. 7). As seen for the combined SEP sub-groups, differentiation into sub-groups was driven by a few SEP indicators and there was little difference between sub-groups in whether being able to keep the home warm enough in winter or able to afford two pairs of all-weather shoes (Fig. 7).

Ethnicity-specific SEP was associated with 115 and 98 metabolic traits at the FDR corrected P < 0.05 threshold in White Europeans and South Asians, respectively (Additional File 8: Data Set 5). Results were consistent with those in the combined SEP analysis and included associations in both White Europeans and South Asians between lower SEP and lower HDL-cholesterol and cholines, and higher triglycerides (Fig. 8). Analyses in White British and Pakistani women identified similar SEP groups, and similar differences in metabolic traits to those found in White Europeans and South Asians, respectively (Additional File 8: Data Set 5, Additional File 9: Fig. S2).

Our study only included White Europeans and South Asians and therefore findings may not generalise to other ethnic groups. The participants were from a high-income country and so findings might not generalise to Whites and South Asians in low-income countries. Women with incomplete data on all SEP indicators were included in LCA using FIML which gives unbiased results under the missing at random assumption. However, if this assumption does not hold, this can produce bias and make the model selection criteria less reliable. We found that only a few SEP indicators explained most of the variation between SEP latent classes therefore, future studies might want to compare LCA to the conventional approach of using one SEP indicator. The Nightingale NMR platform used here primarily covers lipoproteins and therefore we have not assessed other classes of metabolites in detail.