Screening for gestational diabetes mellitus in women with low BMI in an urban, Chinese, population: two retrospective cohorts study

Background We design this study to assess the utility of baseline characteristics in first trimester in the prediction of gestational diabetes mellitus (GDM) in non-obese, women in an urban, Chinese, antenatal population. Methods Multiple, logistic regression analysis was used to develop a predictive model for GDM, in a retrospective cohort of 3956 women (mean age 30.61, years, mean pre-pregnant BMI= 21.45 kg/m2) during 2015 in Shanghai. We refined the predictive model with t-distributed stochastic neighbor embedding (t-SNE) to attempt to distinguish GDM from non-GDM. GDM prevalence in the retrospective cohort was 16.7%. Advanced age, prepregnant obesity, high first-trimester, fasting, plasma glucose, and, a family history of diabetes were positively-related to the development of GDM. The simplified point scoring system yielded an AUC of 0.69. When the model was applied to a prospective cohort of 6572 women recruited in 2016 (mean age 30.80 years, mean prepregnant BMI = 21.18 kg/m2, incidence of GDM 10.8%), the AUC was 0.70. A score of ≥3 points was 91.3% specific and 25.4% sensitive for GDM. In both cohorts, the distinct regions of GDM and non-GDM were ideally separated in the t-SNE, generating transitional boundaries in the image found by different color.


Abstract Background
We design this study to assess the utility of baseline characteristics in first trimester in the prediction of gestational diabetes mellitus (GDM) in non-obese, women in an urban, Chinese, antenatal population.

Methods
Multiple, logistic regression analysis was used to develop a predictive model for GDM, in a retrospective cohort of 3956 women (mean age 30.61, years, mean pre-pregnant BMI= 21.45 kg/m2) during 2015 in Shanghai. We refined the predictive model with t-distributed stochastic neighbor embedding (t-SNE) to attempt to distinguish GDM from non-GDM.

Results
GDM prevalence in the retrospective cohort was 16.7%. Advanced age, prepregnant obesity, high first-trimester, fasting, plasma glucose, and, a family history of diabetes were positively-related to the development of GDM. The simplified point scoring system yielded an AUC of 0.69. When the model was applied to a prospective cohort of 6572 women recruited in 2016 (mean age 30.80 years, mean prepregnant BMI = 21.18 kg/m2, incidence of GDM 10.8%), the AUC was 0.70. A score of ≥3 points was 91.3% specific and 25.4% sensitive for GDM. In both cohorts, the distinct regions of GDM and non-GDM were ideally separated in the t-SNE, generating transitional boundaries in the image found by different color.

Conclusions
Our low BMI, antenatal population now over-consumes a Western, diabetogenic diet during pregnancy though we attribute the declining incidence (2015, 16.7% vs 2016, 10.8%) to an improved antenatal education program. We propose a score system to screen for GDM in first trimester, and also, we should concentrate our efforts on prevention through antenatal education.

Background
Gestational diabetes mellitus (GDM) was perceived as abnormal glucose intolerance that is primarily detected during pregnancy. Rapid societal transition from traditional famine lifestyle famine to an obesogenic environment puts Chinese people at high prevalence of GDM, especially in urban area, ranging from 8.1 to 19.7% 1-3 . GDM results from varying degrees of insulin resistance to placentaderived hormones, and, results in incremental deposition of maternal adipose tissue 4 . Typically, GDM is diagnosed with an abnormal OGTT at 24-28 th week gestation. Recently, it is controversial that fasting glucose greater than 5.1 mmol/l in the first prenatal visit is appropriate for GDM diagnosis 5 .
Clearly, identifying women at risk of GDM as early as possible in pregnancy enables interventions that reduce the risk of adverse fetal outcomes including those associated with fetal macrosomia, and, later maternal consequences including the metabolic syndrome and cardiovascular morbidity 6,7 . Many groups have tried to develop predictive models based on risk factors identified in the first trimester, to identify an abnormal OGTT at 24-28 weeks. Scoring systems, biochemical assays of glucose, glycosylated haemoglobin levels, etc have all been used in different populations with varying degrees of success [8][9][10][11] . Established risk factors for GDM include advanced maternal age, excessive weight gain during pregnancy, overweight or obesity, diabetes in first degree relatives, giving birth to an infant with a macrosomia, etc 12 . Clinical trials have shown that GDM can be prevented to some extent by intensive lifestyle modification implemented before 20 weeks of gestation 13,14 .
Different populations have different risk profiles for GDM. In Western countries GDM tends to occur in overweight women (BMI > 30 kg/m 2 ) or women with increasing gestational weight gain 15,16 whereas in some developing countries it affects populations of low BMI women to varying degrees. The reasons for GDM in these circumstances are not entirely clear. In the present study, our aim is to create a simple, implementable strategy to identify GDM in low BMI women in a Chinese urban population.

Study design
A retrospective cohort study was conducted from January 2015 to December 2015 (N=3956). Eligible subjects were recruited who were undergoing their first antenatal visit at the International Peace Maternity and Child Care Health Hospital (IPMCH). The validation cohort was recruited in 2016 (N=6572). Demographic data of the subjects were obtained by a face-to-face interview questionnaire at the first prenatal visit, which included last menstrual period, maternal age, gestational age, prepregnant weight, current height, personal history, educational levels, parity, diabetes in first degree relatives. BMI was calculated by dividing the weight (kg) by the height squared (m 2 ). Eligible subjects gave blood samples for analysis of fasting plasma glucose and glycated hemoglobin (HbA1c) at the first visit to antenatal clinic.
GDM was diagnosed at 24-28 weeks of gestation according to the American Diabetes Association (ADA) criteria using abnormal plasma glucose values during the 2 hours, 75-gOGTT. Abnormal values were defined according to thresholds established by the ADA: a fasting level greater than 5.1 mmol/l, a 1-hour value greater than 10.0 mmol/l, and, a 2-hour value greater than 8.5 mmol/l.

Exploratory cohort
This cohort included 4205 women who were recruited at the IPMCH in the first trimester of pregnancy.

Validation cohort
This cohort included 7022 women during 2016, recruited at the IPMCH during the first trimester pregnancy. Women with pre-existing diabetes (FPG ≥ 7mmol/L and (or) HbA1c ≥ 6.5% checked in the first antenatal visit) (n=292) and multiple pregnancies (n=158) were excluded, leaving 6572 pregnancies available for analysis.
The key difference between the two cohorts was the incidence of GDM (16.7% v 10.8%). Many factors may contribute to variations in incidence though we believe that this was largely attributable to an antenatal education program to teach antenatal women to manage their diets and weight during pregnancy. In China it is traditional for women to increase their calorie consumption during pregnancy though this is clearly not helpful if they are consuming a diabetogenic "Western" diet.

Risk scoring system
Risk factors for GDM include advanced age, high pre-pregnancy BMI, diabetes in first degree relatives and high, FPG in the first trimester. In the retrospective cohort, multiple logistic regression analysis was used to estimate the coefficients of each risk factor and mutually-adjusted odds ratios (ORs) assigned for GDM. The continuous predictor variables age, BMI and FPG were found to be linear with log odds of the outcomes. To translate the regression coefficients into scores, we used the increasing risk of every 0.5mmol/L increase in FPG as 1 standard unit and gave it an incremental score of 1.
Receiver operating characteristic (ROC) of continuous variables such as maternal age were identified to calculate the appropriate cut-off value for diagnosing GDM. The following risk factors were stratified: maternal age (<30 years and ≥30 years), diabetes in first degree relatives (no versus yes), the cut-off of FPG in the first trimester referred to FIGO guidelines (<5.1 mmol/L and ≥ 5.1 mmol/L), prepregnancy BMI was classified based on the Chinese criteria (normal, <24.0 kg/m 2 ; overweight, 24.0 kg/m 2 ≤ BMI <28.0 kg/m 2 ; obese, BMI ≥ 28.0 kg/m 2 ). The lowest category of each variable was given a score of 0. The combination of all risk factors for a specific total score was computed to obtain GDM risk values. To quantify the accuracy of the prediction models to discriminate subjects with GDM from subjects without, a ROC curve was plotted, and the AUC was calculated. The best fit of the model was evaluated with the Hosmer and Lemeshow test statistics. Then, sensitivity and specificity were calculated at different risk thresholds. A scoring system was used to assign each woman to one of two risk subgroups in the prospective cohort. Spatially mapped t-distributed stochastic neighbor embedding (t-SNE) was used to distinguish non-GDM from GDM, that is, the ability to screen for a true negative group.

Statistical analysis
Analysis were performed using SPSS version 23.0 (SPSS, Inc., Chicago, IL) and R statistical software version 3.6.0 (packages rms and Rtsne). Continuous variables were presented as the mean value with standard deviation and categorical data was expressed as numbers and percentages. Levene's test was used to determine the homogeneity of the variances; Kolmogorov-Smirnov was used to assess the normal distribution. Summary statistics between both groups were compared using either unpaired Student's t-test or Mann-Whitney tests for continuous data, and chi-squared tests for categorical data. A p-value of <0.05 was considered to indicate statistical significance.

Baseline characteristics for two cohorts
In the retrospective cohort of 3956 women during 2015, 662 developed GDM, giving an incidence of 16.7%, mean age was 30.61 years, mean prepregnancy BMI was 21.45 kg/m 2 , mean FPG was 4.45 mmol/L, positive finding of diabetes in first degree relatives was 27.4%. In the validation cohort of 6572 women in 2016, 739 developed GDM, giving an incidence of 11.24%, mean age was 30.84 years, mean prepregnancy BMI was 21.20 kg/m 2 , mean FPG was 4.40 mmol/L, positive finding of diabetes in first degree relatives was 16.5%. Comparison of baseline characteristics between GDM and non-GDM in two cohorts is shown in Table 1.

ROC results
We depicted ROC curves with 3956 subjects from the retrospective cohort, showing a good discrimination for the GDM prediction model with AUC of 0.69 [95%CI0.67-0.72] (Fig. 1a). Applying the retrospective set estimates to the prospective set yielded a similar of AUC 0.70 [95% CI 0.68-0.72] (Fig. 1b). Figure 2 shows the P-value of the Hosmer and Lemeshow tests for comparing the mean probability estimates in each decile of probability in both cohorts. The observed disease incidence were 0.51 (Fig.   2a) and 0.98 (Fig. 2b), respectively.  (Fig. 3).

Screening ability
To assess whether the structure revealed by t-SNE could be linked to clinical outcome, and thereby discriminate subpopulations, we converted the t-SNE space to a*b color space. In the resulting t-SNE image obtained by density-based analysis, each pixel was colored according to its property (orange for GDM and blue for non-GDM). In both cohorts ( Fig. 4a; Fig. 4b), the distinct regions of GDM and non-GDM found by t-SNE were separated, generating transitional boundaries that could be highlighted by two different colors, indicating the ability to distinguish GDM from non-GDM women.

Discussion
Our aim was to create a simple and easily accessible strategy based on clinical history to identify GDM in low BMI (ca.21 kg/m 2 ) women in a Chinese urban population. A 7-point risk-scoring system was developed and validated with these clinical factors for GDM. Although our study demonstrates a reasonable discrimination and calibration, its sensitivity as a first trimester screening test is not high enough. The reasons for the low sensitivity may lie in recent dietary changes in Shanghai where the antenatal population now consumes a high sugar, diabetogenic, Western diet combined with a traditional propensity to increase calorie consumption during pregnancy. The effects of this combination of risks are not detectable in early pregnancy though may become apparent as pregnancy advances. It is clearly important that our antenatal population avoids the risk of overeating a diabetogenic diet, and, we believe that changes to our antenatal education program in 2015- Our study tested the model both internally and externally to assess its predictive performance.
Including both first and second degree relatives with diabetes in the retrospective cohort may have over-estimated the contribution of family history with some measurement bias. So, in the prospective cohort we only considered the first degree relatives with diabetes. There are some limitations to be addressed. Firstly, our study was based on the hospital cohort, and the homogeneity of the cohort increased the internal validity but weakened the confidence of the conclusion. Secondly, we did not incorporate excessive gestational weight gain (GWG) though it was a key factor for developing GDM.
Finally, the sample size of the retrospective cohort was smaller than that in the prospective cohort, because some women returned to their home towns to deliver, and, others refused to participate with the questionnaire or OGTT test in 2015. The final sample size (n=3956) was sufficient for establishing the prediction model, suggesting that these attritional features did not result in a type I error.
In conclusion, we hypothesize that the recent change to a Western, diabetogenic diet together with a traditional propensity to over-consume during pregnancy provides an explanation for the high incidence of GDM in our low BMI, population, and, why our validated, first trimester model has such low sensitivity for GDM in the first trimester of pregnancy. Low sensitivity of a prediction model will cause high false negative rates, which will delay the patient's visit, affect the course of disease and prognosis. Similarly, low specificity of a prediction model will result in high false positive rate, which will waste medical resources and increase patients' anxiety. Our simple model with high specificity can be used as a screening method to identify those who won't develop GDM. Continued monitoring after the first trimester with timely intervention are necessary for those women who are at continuing risk of GDM. Our results suggest that first trimester modeling of GDM risk may not be a practical proposition because the condition largely develops after the first trimester of pregnancy, and, our efforts may be better directed at antenatal, dietary education in the first trimester to attempt to prevent GDM in the first instance.

Declarations
Ethics approval and consent to participate

Consent for publication
Not applicable.

Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no conflict of interest.  Calibration results. Comparison of the mean probability estimated in each decile of probability with the observed disease incidence in both cohorts. The models predicted risk very well in both retrospective cohort (p=0.51) (a) and prospective cohort (p=0.98) (b).

Figure 3
Nomogram to estimate the risk of GDM. Each predictor is assigned a score on each axis.
Compute the sum of points for all predictors and denote this value as the total points. The corresponding "risk of GDM" of "total point" was converted to a predicted probability of GDM. GDM: gestational diabetes mellitus; BMI: body mass index; FPG: fasting plasma glucose.