Hemoglobin A1c as screening for gestational diabetes mellitus in Nordic Caucasian women

We used data from a previously reported randomized controlled trial comparing the effect of a 12-week regular exercise program to standard antenatal care on GDM prevalence [15]. Pregnant women booking an appointment for ultrasound scan in gestational week 18 at St. Olavs Hospital (Trondheim University Hospital) from April 2007 to June 2009 and Stavanger University Hospital from October 2007 to January 2009 were invited to participate. More than 97 % of pregnant Norwegian women attend a free of charge ultrasound scan around week 18. During the study period around 12,000 women had routine ultrasound scans at the two study centers and were eligible for the study. Inclusion criteria were age ≥18 years and a singleton viable fetus. Exclusion criteria were high-risk pregnancies or diseases that could interfere with participation. In addition, women who lived more than 30 min drive from the study center were excluded due to practical reasons. A total of 875 women consented to participate, whereof 20 were excluded due to twin pregnancies (n = 2), miscarriages (n = 5) and not meeting inclusion criteria (n = 13). Data was collected at inclusion (week 18–22) and follow-up (week 32–36). All participants gave written informed consent. The study was approved by the Committee for Medical Research Ethics of Health Region IV in Norway (REK 4.2007.81) and is registered at http://​www.​clinicaltrials.​gov as NCT 00476567. The Declaration of Helsinki was followed throughout the study.

Women in the intervention group received a standardized exercise program of 60 min duration including aerobic activity, strength training and balance exercises instructed by a physiotherapist once a week over a 12-week period. They were encouraged to follow a 45 min home exercise program (30 min endurance training and 15 min strength and balance exercises) at least twice a week. Women in the control group received standard antenatal care and the customary information given by their midwife or general practitioner. They were not discouraged from exercising on their own. Women in both groups received written recommendations on diet, pelvic floor muscle exercises, and pregnancy-related lumbo-pelvic pain.

Women diagnosed with GDM by the World Health Organization (WHO) criteria from 1999 [16] during the study period received standard treatment for GDM, i.e. initially diet and life style advice. Insulin treatment was considered if serum glucose (s-glucose) was persistently elevated, i.e. fasting s-glucose >6.0 mmol/L or >8.0 mmol/L 1–1.5 h after a meal. None of the participants needed insulin treatment.

Age, smoking status and information regarding previous pregnancies and family history of diabetes were collected through questionnaires. Macrosomia was defined as birth weight >4000 g. Weight, height and blood pressure were measured at pregnancy weeks 18–22. After 15 min rest, blood pressure was measured three times with 2 min break between measurements. The average of the two last measurements was used. Data on pregnancy complications and adverse outcomes were obtained from medical records. Preeclampsia was defined as systolic blood pressure over 140 mmHg and/or diastolic blood pressure over 90 mmHg and proteinuria ≥0.3 g/24 h measured more than once 4–6 h apart occurring after gestational week 20.

GDM was diagnosed at pregnancy weeks 18–22 and 32–36 by the WHO criteria from 1999 as fasting s-glucose ≥7.0 mmol/L or s-glucose ≥7.8 mmol/L 2 h after ingesting 75 g glucose orally (OGTT) [16]. After the study we also diagnosed GDM according to modified IADPSG criteria as fasting s-glucose ≥5.1 mmol/L or s-glucose ≥8.5 mmol/L 2 h after the glucose load [3]. We could only use modified IADPSG criteria since we did not have 1-h s-glucose. GDM diagnosed by the WHO criteria will hereafter be named GDM-WHO and by modified IADPSG criteria will be named GDM-IADPSG.

Fasting and 2-h glucose levels were measured in serum by the routine methods used by the hospital laboratory.

Venous blood samples from pregnancy weeks 18–22 and 32–36 were stored at −80 °C. HbA1c is stable at these storage conditions [17, 18]. HbA1c was analyzed over a 3-week period in October 2014 at our hospital laboratory with an immunological method on a Roche Cobas Integra (Roche Diagnostics, Mannheim, Germany) [19]. The method was calibrated against the standard from the International Federation of Clinical Chemistry and Laboratory Medicine [20]. The coefficient of variation for within-laboratory imprecision during the 3-week period was 2.0 % at HbA1c 5.2 % (33 mmol/mol) and 1.4 % at HbA1c 9.6 % (81 mmol/mol) [21].

To evaluate if HbA1c could potentially reduce the number of OGTTs we used a diagnostic threshold with high sensitivity to rule-out GDM and a threshold with high specificity to rule-in GDM. Those between the two thresholds would need an OGTT to clarify whether they had GDM or not. We evaluated if HbA1c at weeks 18–22 could predict GDM at 18–22 weeks of pregnancy, and if it could predict GDM throughout pregnancy, i.e. GDM diagnosed at weeks 18–22 and 32–36 combined. We also evaluated if HbA1c at pregnancy weeks 32–36 could predict GDM at weeks 32–36. To evaluate HbA1c as a predictor for GDM and preeclampsia together with other data, we used logistic regression with backwards elimination to find the best combination of variables in predicting the outcome. We used receiver operating characteristic (ROC) curve analysis to evaluate the diagnostic accuracy of HbA1c alone, the models predicting GDM and to find suitable diagnostic thresholds for not performing an OGTT [22]. We also assessed goodness-of-fit for the models by the Hosmer–Lemeshow test. The leverage of individual points was visually judged by inspecting a plot of \(\Delta \hat{\beta }_{i}\) against p, where \(\Delta \hat{\beta }_{i}\) is the amount that the logistic regression model parameters change when the ith observation is omitted from the model, and p is the estimated probability of the outcome.

To find the best combination of variables predicting birth weight, we used linear regression. Model assumptions and fit and identification of observations with potentially high influence on the model were evaluated by inspection of residual plots, R² and the DFBETA statistic which quantifies how much the regression coefficients change when the ith observation is omitted from the model.

For all regression analyses we used the Royston and Altman algorithm to find the simplest (if any) non-linear transformation of the continuous variables [23]. The selection of variables was based on variables that should be easily available to the clinician at the time point of HbA1c testing and known to be associated with or suspected to influence on the dependent variable. Also, we included the intervention as a possible predictor variable in all models for GDM at week 32–36 and throughout pregnancy, preeclampsia and birth weight in order to adjust for possible confounding. The significance level for keeping a variable in the model was p value ≤0.10.

We used bootstrap analysis to assess the stability of the multivariable models where the bootstrap inclusion fraction is an indicator for the importance of each independent variable [24].

The statistical analyses were performed using Stata version 13.1 for Windows (Stata Corp., Texas, USA).

It was impossible to assess HbA1c as a screening test for GDM-WHO around 20 weeks of pregnancy, since only five (0.7 %) women tested positive for GDM-WHO at pregnancy weeks 18–22. Between the 32nd and the 36th weeks of gestation, 37 (5.8 %) women were diagnosed with GDM-WHO and throughout pregnancy (i.e. diagnosed at pregnancy weeks 18–22 or 32–36) 42 (6.7 %) women were diagnosed with GDM-WHO.

The area under the ROC curve for HbA1c in diagnosing GDM-WHO at pregnancy weeks 32–36 was 0.74 (95 % CI 0.64–0.83). HbA1c at 32–36 weeks of pregnancy, age and BMI at inclusion, family history of diabetes mellitus (yes/no), GDM in previous pregnancy (yes/no), previously giving birth to a macrosomic baby (yes/no) and intervention (i.e. being in the exercise group or control group) were potential predictors for the model for GDM-WHO at pregnancy weeks 32–36. Only HbA1c and age were included in the model (Table 2), with an area under the ROC curve of 0.76 (95 % CI 0.66–0.85). Sensitivity and specificity for diagnosing GDM at various levels of HbA1c and the predicted probability of GDM from the model are presented in Table 3.

For HbA1c at week 18–22 in diagnosing GDM-WHO throughout pregnancy, the area under the ROC curve was 0.64 (95 % CI 0.55–0.72). HbA1c at week 18–22, age, BMI, family history of diabetes mellitus, GDM in previous pregnancy, previously giving birth to a macrosomic baby and intervention were potential predictors for the model for GDM-WHO throughout pregnancy. Predictors included in the model were HbA1c, age and a family history of diabetes (Table 2). The area under the ROC curve for the model was 0.67 (95 % CI 0.58–0.76). Table 3 presents sensitivity and specificity for diagnosing GDM at potential cut-offs for HbA1c and the predicted probability of GDM from the model.

None of those with GDM-WHO were diagnosed because of an elevated fasting s-glucose. With only 2-h s-glucose ≥7.8 mmol/L as the outcome, fasting s-glucose had about the same diagnostic accuracy and ability to potentially reduce the number of OGTTs as HbA1c. Including fasting s-glucose in the models predicting a 2-h s-glucose ≥7.8 mmol/L did not improve the screening ability significantly.

At pregnancy weeks 18–22, 32–36 and throughout pregnancy (i.e. at week 18–22 or 32–36) 16 (2.4 %), 29 (4.6 %) and 45 (7.2 %) women were diagnosed with GDM-IADPSG, respectively.

The area under the ROC curve for HbA1c in diagnosing GDM-IADPSG at pregnancy weeks 18–22 was 0.67 (95 % CI 0.54–0.80). HbA1c at 18–22 weeks of pregnancy, age, BMI, family history of diabetes mellitus, GDM in previous pregnancy and previously giving birth to a macrosomic baby were potential predictors for the model for GDM-IADPSG at pregnancy weeks 18–22. HbA1c and BMI were the only predictors included in the model (Table 2), with an area under the ROC curve of 0.70 (95 % CI 0.57–0.84). Potential cut-offs for HbA1c and the predicted probability of GDM with corresponding sensitivity and specificity for diagnosing GDM are presented in Table 3.

HbA1c had an area under the ROC curve of 0.76 (95 % CI 0.67–0.85) in diagnosing GDM-IADPSG at 32–36 weeks of pregnancy. HbA1c at weeks 32–36, age, BMI, family history of diabetes mellitus, GDM in previous pregnancy, previously giving birth to a macrosomic baby and intervention were potential predictors for the model for GDM-IADPSG at pregnancy weeks 32–36. HbA1c, BMI and intervention were the predictors included in the model. We assessed possible interaction between HbA1c and intervention, but the interaction term was not statistically significant (p = 0.71), and we excluded the intervention variable from the model. The area under the ROC curve for the model with HbA1c and BMI as predictors (Table 2) was 0.77 (95 % CI 0.68–0.87). Table 3 shows sensitivity and specificity for diagnosing GDM at potential cut-offs for HbA1c and the predicted probability of GDM estimated by the model.

The area under the ROC curve for HbA1c at pregnancy weeks 18–22 in predicting GDM-IADPSG throughout pregnancy was 0.69 (0.60–0.77). HbA1c at 18–22 weeks of pregnancy, age, BMI, family history of diabetes mellitus, GDM in previous pregnancy, previously giving birth to a macrosomic baby and intervention were potential predictors for the model for GDM-IADPSG throughout pregnancy. HbA1c and BMI were included in the model (Table 2) with an area under the ROC curve of 0.72 (0.64–0.80). Sensitivity and specificity for diagnosing GDM at various levels of HbA1c and the predicted probability of GDM from the model are shown in Table 3.

Only four women had history of GDM in a previous pregnancy and none of them had GDM-IADPSG, so this variable was omitted from the analyses of GDM-IADPSG. We performed the analyses for the multivariable models for GDM-IADPSG without the previous GDM variable, and a sensitivity analyses without those with previous GDM. The results did not change after exclusion of those with previous GDM.

The prevalence of preeclampsia was 2.9 %. Predictor variables considered for the preeclampsia model were HbA1c at 18–22 weeks of pregnancy, age, BMI, smoking (yes/no) and systolic blood pressure at inclusion, intervention, nulliparity and GDM-WHO. We also performed the analyses with GDM-IADPSG as an independent variable instead of GDM-WHO, and with HbA1c at weeks 32–36 instead of at weeks 18–22. Variables included in the model for preeclampsia were GDM-WHO (OR 2.97, 95 % CI 0.82–10.7, p = 0.10) and nulliparity (OR 2.80, 95 % CI 0.92–8.5, p = 0.07). The area under the ROC curve for this model was 0.67 (95 % CI 0.58–0.76). Neither HbA1c at pregnancy weeks 18–22 or 32–36 nor GDM-IADPSG were statistically significantly associated with preeclampsia (p > 0.10).

None of the nine women who smoked had preeclampsia, and this variable was omitted from the analyses of the model for preeclampsia. We therefore performed the analyses without the smoking variable, and a sensitivity analyses without those smoking. The results did not change after exclusion of those who smoked.

The following predictor variables were considered for the model predicting birth weight; HbA1c at pregnancy weeks 18–22, age, BMI and smoking at inclusion, intervention, nulliparity, previously giving birth to a macrosomic baby and GDM-WHO. We also performed the analyses with GDM-IADPSG as an independent variable instead of GDM-WHO, and with HbA1c at pregnancy weeks 32–36 instead of at weeks 18–22. Variables included in the model for birth weight were HbA1c at pregnancy weeks 32–36 (β 137, 95 % CI −10 to 283, p = 0.07), BMI (β 35, 95 % CI 23–47, p < 0.0005), nulliparity (β −102, 95 % CI −178 to −28, p = 0.008) and previously giving birth to a macrosomic baby (β 301, 95 % CI 161–440, p < 0.0005). Neither HbA1c at pregnancy weeks 18–22, GDM-WHO nor GDM-IADPSG were statistically significantly associated with birth weight (p > 0.10).

We did not find any non-linear associations between the continuous variables and the outcome for any of the models. We assessed the stability of the multivariable models by bootstrap analysis [24]. The results were in consistence with our final models as HbA1c was selected in 74–100 % of the replicates for all GDM models, except for the model predicting GDM-IADPSG at pregnancy weeks 18–22 where it was chosen in 31 % of the replicates. For the other predictors of GDM, only BMI in the model for GDM-IADPSG throughout pregnancy was chosen in more than 60 % of the replicates (79 %).