Materials and Methods

Participants

The design and methods of BIGCS have been described previously(13). In brief, eligible women with Chinese nationality, living in Guangzhou who are <20 weeks gestation and who intend to deliver at one of the two Guangzhou Women and Children’s Medical Centre(GWCMC) campuses were recruited into BIGCS. This study was conducted in a planned subgroup of BIGCS in whom maternal and cord blood were analysed for metabolic parameters separately. Pregnant women attending BIGCS with a singleton pregnancy who delivered at GWCMC between Jan 2015 and Jun 2016 and had umbilical cord blood retained are eligible for this study. Women were excluded if: 1) maternal blood samples unavailable at 14-27 gestation week; 2) no records of maternal fasting glucose at 20-28 gestation week; 3) lacking maternal demographic information; 4) diagnosed with health condition prior to pregnancy, including type 1 or type 2 diabetes, thyroid dysfunction, hypertension, virus hepatitis, and renal diseases. The study was powered for the association between maternal triglycerides (the potential weakest risk factors among maternal metabolic traits) with birthweight according to literature(Supplementary file S1). The eligible mother-child pairs were then selected into this study by computer generated randomization. Ethical permission for the study was granted by the GWCMC Ethics Committee.

Study Procedures

Maternal demographic data, including anthropometric measures, socioeconomic status, family and personal medical history, were collected through a semi-structured questionnaire(Q1) at recruitment. Maternal overnight fasting blood samples were collected during second trimester. At 22-28 weeks gestation, women attending their second prenatal visit underwent a standard 2h 75g oral glucose tolerance test(OGTT). Women with OGTT results which met or exceeded at least one threshold of the International Association of Diabetes and Pregnancy Study Groups(IADPSG) criteria(FPG≥5.1 mmol/L, 1h glucose ≥10.0 mmol/l, and 2h glucose≥8.5 mmol/L) were diagnosed as having gestational diabetes mellitus(GDM)(14). For participating children, birth information, including birth characteristics, delivery mode, and perinatal outcomes were obtained from routine medical records. Umbilical cord blood samples were collected by midwives at birth.

Demographic Data

Maternal demographic information(age, height, pre-pregnancy weight, parity, date of last menstrual period, monthly income, education levels, and ethnicity) were collected through Q1 questionnaire. BMI was calculated by dividing weight in kilograms by height in meters squared. Based on the recommendations of the China Obesity Task Force of the Chinese Ministry of Health, maternal pre-pregnancy BMI is classified into two groups: lean group(<24 kg/m2) and overweight group(≥24 kg/m2) (15). Maternal second trimester weight was measured to the nearest 0.1 kg using an electronic scale. Maternal early gestational weight gain(GWG) was calculated by subtracting pre-pregnancy weight from maternal second trimester weight, with documentation of the gestational age at measurement. Maternal fasting glucose concentration was obtained from OGTT test zero-time value in hospital records.

Biochemical Test

Sample collection, delivery, pre-treatment, and measurements were blinded. All blood samples were stored and delivered to pre-treatment laboratory centre. Blood samples were then separated to serum and plasma by immediate centrifugation, and were stored in EDTA tube in the bio-bank at -80℃ until analysis. Plasma lipids(TC, HDL-C, LDL-C, and TG) and insulin levels were measured using commercial kits in fully automated clinical analyser(Roche Diagnostics, Mannheim, Germany). Intra- and inter-day coefficients of variation(CVs) were consistently less than 2 percent for all assays.

Neonatal anthropometry

Gestational age was estimated from ultrasound examination during the first- or second-trimester. Birthweight and other information, including gestational age at delivery, mode of delivery, neonatal sex, and pregnancy complications were obtained from hospital records. Birthweight was measured to the nearest 50g using an electronic scale by midwives immediately after delivery. Birthweight Z-Score and percentile(adjusted for gestational age at delivery and neonatal sex) were calculated using Intergrowth 21st Newborn Size Standard and Tools(16). Large for gestational age(LGA) was defined as a birthweight larger than the 90th percentile for gestational age by sex, while Small for gestational age(SGA) was defined as a birthweight smaller than the 10th percentile based on the same birthweight reference.

Statistical Analysis

Classic statistical methods

For the baseline table data are summarized as mean ± Standard Deviation(SD), median(Inter Quartile Range, IQR), or counts with percentages. Pearson correlation was used to assess the impact of the long-term -80 °C storage on insulin concentrations in EDTA tube. Adjustments were then made to account for any degradation by correcting the initial value using linear regression methods(Supplementary file S2). Similarly, maternal lipid levels were adjusted for gestational age using regression model to account for timing of blood sampling(Supplementary file S3)(17).
Initially, linear and logistic regression were used to estimate the association between maternal metabolic parameters and neonatal continuous and binary outcomes, respectively. Further analyses using linear regression model were performed after all exposures were transformed to Z-Scores. This was to enable comparison of the effect size each maternal metabolic parameter had on birthweight Z-Score and CBI Z-Score. CBI and maternal triglycerides were log-transformed prior to standardization. Multiple imputation was used to handle missing data. Subgroup analyses were conducted in boys and girls respectively. Sensitivity analyses were conducted to compare the estimate differences between GDM and non-GDM participants, fasting blood samples and non-fasting samples, primiparous women and non-primiparous women, lean and overweight group, as well as before and after multiple imputation(Supplementary file S4). All statistical tests were two-tailed and a P <0.05 was considered statistically significant. Statistical analyses were performed in Stata version 14.0(College Station, Texas, USA).

Additive Bayesian Networks (ABN) analysis

To further assess the inter-dependency between maternal metabolic risk factors and their association with birthweight and CBI, Additive Bayesian Network(ABN) model - an unsupervised machine learning method - was conducted. Bayesian network analysis is a form of structure discovery statistical modelling that derives, from empirical data, a graphical network describing the dependency structure between variables, shown as directed acyclic graphs(DAGs)(12). ABNs comprise of DAGs where each node in the graph comprises a generalized linear model(GLM) or a generalized linear mixed model(GLMM). ABN model is suitable for analysing highly complex epidemiological data comprising many inter-dependent variables(11).
Ten variables were chosen for ABN based on prior knowledge gained from literature and findings of the classical statistical analyses. These ten variables were maternal age, maternal pre-pregnancy BMI, maternal fasting glycaemia in OGTT, early GWG, maternal fasting plasma HDL-C and triglycerides in the second trimester, birthweight Z-Score, cord blood insulin, gestational age at delivery, and neonatal sex. GWG was adjusted for gestational age at weight measurement in mid-pregnancy. Cord blood insulin was adjusted for sample storage duration. All continuous variables were standardized to Z-Scores to eliminate the influence of different measurement units. Mother-child pairs with missing data were excluded(n=93/1522, 6%).
Firstly, an optimal DAG with the best goodness of fit(highest log marginal likelihood) was identified. Next, parametric bootstrapping(12800 samples) was performed to address the potential overfitting. Full technical details are provided in the supplementary file(S5). ABN analysis was conducted in R 3.4.4(The R Foundation for Statistical Computing) using ‘abn’ package(11).