Background
The prevalence of diabetes in pregnancy, the majority of which is gestational diabetes mellitus (GDM), has been increasing in many jurisdictions [
1]. GDM is defined as glucose intolerance that first occurs or is first diagnosed during pregnancy and usually resolves soon after delivery [
2]. It raises substantial health concerns both for its short-term adverse effects on pregnancy outcomes and for potential serious long-term consequences for both mothers and their offsprings [
2,
3,
4]. For instance, GDM is associated with large for gestational age (LGA), high infant adiposity, shoulder dystocia, caesarean section and preeclampsia [
5,
6,
7]. The prevalence of GDM in Chinese women ranged between 13.0 and 20.9%, and the variation was partly due to different criteria [
8,
9,
10,
11]. It is important to predict the risk of GDM early in pregnancy to enable early interventions to prevent GDM. Unfortunately, pregnancy is a complex and dynamic process, involving profound changes in energy and nutrient metabolism to sustain fetal development and growth, and to meet the requirements of labour and lactation.
The changes of energy and nutrient metabolism during pregnancy have been postulated to follow a biphasic pattern, predominantly anabolic in early pregnancy and mainly catabolic in later pregnancy [
12]. These dynamic changes have not been completely illustrated. For example, in a large cross-sectional study in Israel, triglyceride (TG), total cholesterol (TC) and low density lipoprotein (LDL) levels decreased between the time of conception and 8 weeks of gestation and then gradually increased and peaked just before delivery [
13]. However, in a prospective cohort study in Brazil, levels of these lipids increased from the first trimester to the third trimester [
14]. Despite the difficulties in risk prediction of GDM due to by these dynamic changes, efforts have been made to develop risk prediction models for GDM incorporating factors such as maternal age, ethnicity, BMI [
15], family history of diabetes, personal history of GDM, fasting plasma glucose (FPG), vitamin D3, macrosomia and chronic hypertension, high-sensitive C-reactive protein, placental protein 13, pentraxin 3, myostatin, follistatin, and soluble fms-like tyrosine kinase-1 [
15,
16,
17,
18,
19,
20,
21].
It has been reported that the model combining maternal factors including history of GDM, family history of diabetes, ethnicity, parity, BMI, mean arterial pressure (MAP), uterine artery pulsatility index (UtA PI) and pregnancy associated plasma protein A (PAPP-A) in the first trimester predicted GDM with an area under the receiver operating characteristic curve (AUC) of 0.90 (95% CI 0.87–0.92) [
22]. Another model developed by applying a machine learning algorithm to data extracted from health records for the first trimester to predict risk GDM at 24–28 weeks of gestation achieved an AUC of 0.86 and accuracy of 62.2%. This algorithm also included maternal factors such as age, parity, BMI, education, FPG and other hematological and biochemical test results [
23]. In addition, a model has been developed in obesity pregnant women with age, previous GDM, family history of type 2 diabetes, systolic blood pressure, sum of skinfold thicknesses, waist-to-height ratio, and neck-to-thigh ratio provided an area under the curve of 0.71 for GDM prediction (95% CI 0.68–0.74) [
24]. However, it is uncertain whether these models are applicable to Chinese women.
Furthermore, a growing body of evidence suggests that hypertriglyceridaemia is associated with GDM [
25,
26]. Hypertriglyceridaemia itself is a well known risk factor for metabolic syndrome. Moreover, it is independently associated with pregnancy outcomes such as LGA, macrosomia [
24,
27,
28,
29] and preterm delivery [
2,
24]. Whether and to what extent TG level predicts GDM is unknown. Therefore, the purpose of the present study was to develop a simple model incorporating maternal age, BMI, and FPG and TG levels in early pregnancy to predict the risk of GDM.
Discussion
The association of unusual rise of FPG and TG with GDM and/or pregnancy outcomes had been explored [
5,
10,
11,
13,
15,
26]. FPG ≥5.1 mmol/L predicted LGA risk [
34] and anti-diabetic treatment (e.g., insulin, metformin) might reduce the risk of adverse perinatal outcomes [
35]. Zhao et al. went even further to try to quantitatively explore these associations by logistic regression [
36]. If the predictors are correlated as in this case, it is improper to use multivariable logistic regression by frequentist’s method. Liu et al. might have realized this pitfall of using logistic regression, but still they failed to show that FPG correlated with TG or other lipids in their analysis [
28]. There are other GDM predictive tools incorporating more biochemical or ultrasonic maternal factors. To our knowledge, the model developed in this study using information on maternal factors such as age, prepregnancy BMI, FPG and TG and using robust modeling methods is the first model applicable to Chinese women for whom no ethnicity-specific GDM risk prediction model has been available. Further validation of the model is necessary before generalizing it to the vast rural population of China.
The incidence of GDM in this study (12.8%) was lower than the overall frequency (17.8%) reported in HAPO study [
37]. The prevalence of GDM was much lower in studies that used either the WHO 1980–2013 or ADA 2002–2014 criteria (13.0 to 13.9%) than in studies that used IADPSG and China MOH criteria which gave a prevalence of 19.9 and 20.9%, respectively [
8]. The prevalence of GDM also differed by screening methods, with the one-step screening method yielding a prevalence of GDM of 14.7%, and the two-step screening method a prevalence of 7.2%, about half that of the one-step method [
8,
38,
39].The prevalence of GDM would have been higher if the lower fasting and post-load glucose thresholds for GDM diagnosis which was previously applied in south Asian women (from 17.4 to 24.2%) had been used [
5]. In a systematic review, the GDM was postulated to affect 14% of pregnant women annually and accounted for 90% of hyperglycemia pregnancies [
40].
Women with GDM also had elevated FPG and TG levels in the early stage of pregnancy, and as a consequence had an increased risk of adverse perinatal outcomes such as high birth weight (e.g., macrosomia or LGA) and caesarean section [
28,
33]. In a consensus guideline the Australasian Diabetes in Pregnancy Society (ADIPS) has urged more research to identify the most appropriate tests for detecting GDM in early pregnancy. Only following early detection can risk stratification and early intervention for GDM become possible, but still it is difficult. The available models suffer from one or more of the limitations including applying only to sub-populations [
24], involving complicated algorithms [
23], requiring specific laboratory tests [
21,
22,
24] or applying in specific periods before GDM diagnosis [
21,
22,
41]. The model derived in the present study did not have AUC, sensitivity or specificity as high as some previous models but it was simple, applicable to the Chinese population and covered a broad period before GDM diagnostic tests, and therefore has the potential to be widely used in less developed jurisdictions.
Risk factors for GDM included a history of hyperglycemia, having a first degree relative with diabetes mellitus, a history of macrosomia, polycystic ovarian syndrome, age > 40 years, pre-pregnancy BMI > 30 kg/m [
2], and taking medications such as corticosteroids, and antipsychotics [
15,
17,
40,
41]. This list is increasing. Physically active Asian women have a reduced risk of GDM (OR = 0.56, 95% CI 0.32–0.98), especially those who were overweight (OR = 0.52, 95% CI 0.29–0.93) or obese (OR = 0.34, 95% CI 0.15–0.77) [
42]. Weight gain has been reported to be an independent risk factor of GDM [
43], especially in the first three months of pregnancy. Levels of C-reactive protein (hs-CRP) and sex hormone binding globulin (SHBG) in the first trimester may predict GDM with high sensitivity [
44]. Blood count had been incorporated in a GDM prediction model [
45]. Hemoglobin A1C (HbA1c) level may be helpful in GDM diagnosis, especially in the third trimester [
46,
47]. Correlations between glucose and lipid levels, together with interactions among potential risk factors, further complicate GDM prediction. How to develop a reliable GDM risk prediction model by using the above-mentioned complicated risk factors poses a challenge. Regression analysis using Bayesian inference is applicable to predictors that are independent or correlative. However, principal component analysis and other methods are also approaches to combine correlated or interacting predictors. We used Bayesian inference in regression analysis because it has a similar interpretative framework to frequentist logistic regression but is more robust when correlated predictors are present. In the sensitivity analysis using frequentist logistic regression, the ORs and AUC, sensitivity and specificity were very similar to those obtained by Bayesian inference (Additional file
3: Table S2, Additional file
2: Table
2, Additional file
1: Figure S1).
We also classified the continuous variables such as maternal age, pre-pregnancy BMI, FPG, TG, TC, HDL, LDL etc. into categories according to clinical reference intervals for healthy adults. In terms of predicted risk probability and AUC, the model based on categorical variables did not show higher predictive performance than the model based on continuous variables (AUC 0.748 vs 0.770, p = 0.316). The categorical variables did not show significant interaction effects. Fasting glucose and triglycerides may increase with gestation weeks. The original full model also included a history of abnormal gravidity (preterm birth, miscarriage and induced abortion). Following variable selection the best model incorporated four factors including maternal age, prepregnancy BMI, FPG and TG. The pseudo-R-squared was 0.165 for the best fit model, suggesting that the potential risk factors included could accounted only for a small fraction of the variance of GDM. Further studies on GDM risk factors and prediction models should pay more attention to understanding the interactions among the risk factors.
The strengths of this study include a moderate sample size, a robust modeling strategy for correlated predictors, and its development of a simple formula for prediction based on only four maternal factors (age, prepregnancy BMI, FPG and TG). This formula can be applied even if only a calculator is available, such as in less-developed parts of China. The limitations are that we were not able to incorporate all known potential risk factors in modeling, that the participants came from a single medical center, and that we have not yet conducted external validation. Due to intrinsic nature of observational data in this cohort study, the internal validity was low to moderate.
Conclusions
From 8th to 20th gestational week, the average FPG level indicated a slight decreasing trend while, in line with previous studies, TG levels displayed a slight increasing trend. The metabolites were intercorrelated during this period of pregnancy. We developed a prediction model in Chinese women, which addressed the correlation of predictors and incorporated maternal age, prepregnancy BMI, FPG and TG, with a predictive accuracy of 0.64 and an AUC of 0.766 (95% CI 0.731, 0.801). Thus, a simple model, which can be applied using a hand calculator, may predict the risk of GDM and be used in less economically developed parts of China.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.