Regression discontinuity studies use variations in assignment methods to emulate random treatment assignment, analogous to RCTs16. For instance, when diagnostic tests are used to assign subjects, random analytical/pre-analytical variations of test results around the threshold acts to quasi-randomise subjects to an intervention16. This quasi-random allocation of subjects based on OGTT is pertinent given the poor reproducibility of OGTT for individuals17; nonetheless, there are robust correlations of GDM outcomes to OGTT plasma glucose for a population13.
Regression discontinuity studies assumes that there is a continuous relationship between outcome variables16 such as rates of LGA, caesarean sections and the dependent variable, the HAPO glucose composite scores. We demonstrate a linear continuous relationship between LGA and caesarean sections and HAPO glucose composite scores in non-GDM subjects (Fig. 2).
Regression discontinuity studies typically estimates treatment effects proximate to the diagnostic cut-off16. However, the diagnosis of GDM is unusual in having multiple diagnostic cut-offs at OGTT fasting, 1 hour, and 2 hour plasma glucose. Furthermore, the heterogeneity of OGTT plasma glucose trajectories with multiple cut-offs resulted in overlaps in the HAPO glucose composite scores, between GDM and non-GDM subjects. This allowed inference of treatment effects to wider bandwidths than other forms of regression discontinuity study designs. Nevertheless, the counterfactuals represent those with milder GDM.
The overall GDM prevalence and OGTT glucose levels were similar to HAPO centres18. GDM subjects compared to the counterfactual group had higher BMI. Expectedly, as OGTT were used to determine diagnosis, GDM subjects had higher OGTT plasma glucose levels. As covariates associated with adverse outcomes were higher for GDM subjects, the interventions had to overcome these selection biases to show improved outcomes.
Our centre’s overall caesarean section rate of 27.8% is similar to the Organisation for Economic Co-operation and Development rate of 27.9%19; but below the Australian rate of 34%19.
Capillary blood glucose targets and treatment intensification thresholds
Presumably, higher maternal glucose levels are the driver of adverse clinical outcomes in GDM13 and that moderating elevated maternal glucose levels improves outcomes. Therefore it is important to discuss clinical outcomes with the understanding of treatment targets and thresholds for treatment intensification. ADIPS9 suggested treatment targets were fasting CBG ≤ 5.0 mmol/L and the post-prandial ≤ 6.7 mmol/L. Ostensibly the aim was to have maternal CBG levels within the reference range of a healthy population9. This is tighter than RCT CBG targets in the Australian Carbohydrate Intolerance Study (ACHOIS) 20: fasting ≤ 5.5 mmol/L, and 2 hour post-prandial of ≤ 7.0 mmol/; and the Maternal-Fetal Medicine Unit study (MFMU) 21: fasting ≤ 5.3 mmol/L, and 2 hour post-prandial of ≤ 6.7 mmol/L.
The threshold for intensifying treatment will also impact medication usage. ADIPS9 suggested intensification if ≥ 2 out 7 results (29%) were above target; in comparison, the MFMU study 21 intensified treatment if more than 50% of CBG were above target. Glucometer imprecision will affect treatment intensification; the median coefficient of variation (CV) among of 18 commercial glucometers in recent survey was 9.3%22. To illustrate this point, take a theoretical woman with ‘true’ consistent fasting CBG levels within recommended levels at 5.0 mmol/L, but due to glucometer imprecision 95% (2 sd or 2CV) of results would be between 4.1 to 5.9 mmol/L, of which half would be > 5.0 mmol/L. Therefore it is probable that more than 29% of fasting CBG will be above 5.0 mmol/L, necessitating treatment intensification.
Notwithstanding glucometer imprecision issues, our population’s upper reference interval (97.5th percentile) of OGTT fasting plasma glucose was 5.3 mmol/L. Our OGTTs were collected using sodium fluoride collected at room temperature and batched centrifuged10. This pre-analytical method has been recently documented, via the process of glycolysis, to lower reported OGTT fasting glucose levels by approximately 10%10,23. Therefore the adjusted lower and upper reference level of fasting glucose should be 10% higher — approximately 4.1 to 5.8 mmol/L. A fasting glucose target of 5.0 mmol/L is well within our population’s reference intervals.
Treated GDM subjects had a relative risk of LGA of 0.37, reducing the absolute risk from 12.6 to 4.6%. The RCTs, ACHOIS20 and MFMU 21, reported a decrease in LGA rates from 22 to 13%, 13–7.1%, respectively. As our CBG targets were lower it was not surprising that our rates of foetal overgrowth were also lower.
The relative risk for caesarean sections for GDM as compared to the counterfactual group was 0.75, the absolute risk reduced from 43.0 to 32.3%. Subgroup analysis suggests the reduction of caesarean section rates were driven by women with BMI ≥ 30, where the relative risk was 0.58 compared to counterfactual group with BMI ≥ 30. The absolute risk decreased from 55.9–32.9%. In contrast, GDM women with BMI < 30 did not demonstrate a reduction in caesarean section rates. Our study suggests the treatment response upon caesarean section rates is modified by obesity.
The two randomised led trials of GDM had shown contrasting caesarean section rates. ACHOIS20, showed no effect of treatment on caesarean section rates, but in the MFMU study 21, treatment resulted in a relative risk of 0.79 (p = 0.02). Although these studies differed in treatment targets and selection of subjects, our BMI subgroup analysis may partly explain the differences in caesarean section outcomes. The baseline BMI of subjects differed between these two RCTs, the trial not showing caesarean rate reduction had lower median BMI of 26.820, the other had a mean BMI of 30.1 kgm2 21.
Apart from tighter treatment targets, another factor that may have contributed to the larger absolute and relative caesarean risk reduction seen in our study compared to the RCTs is the marked differences in the baseline risks of caesarean sections. Our counterfactual group had caesarean section rate of 43.0%, driven primarily by women with BMI ≥ 30; this is in the context of the overall study centre rate of 27.8%. In contrast, the RCTs had rather muted untreated GDM (control) caesarean section rates of 32% and 33.8%. This is in the context of national caesarean section rates in Australia (2004) 24 of 29.4% and the United States (2005)25 of 30.3%. We note the established linear correlation of caesarean section rates with increasing maternal plasma glucose13. Therefore in women with very high maternal plasma glucose levels, as in untreated GDM, we would have expected much higher caesarean section rates than the centre or national levels. In our study, obesity was a strong driver of higher caesarean section rates; and our rates for the counterfactual group are consistent with the odds ratios seen in a meta-analysis investigating obesity’s impact on caesarean sections: 2.05 and 2.89 for obese and severely obese women, respectively26.
In RCTs, the act of randomisation6 and subject selection processes may have recruited women with better baseline prognoses relative to the target population27. RCTs selects for particular types of willing participants27. Moreover, inclusion in a study may have affected the behaviour of subjects, especially as the primary interventions in GDM are behavioural change. This may have resulted in RCTs lowering adverse outcomes in comparator controls thereby attenuating the treatment response.
As our counterfactual group represents those borderline for GDM, we likely understated the overall treatment response. If the counterfactual group were to represent untreated GDM, mild and severe, they would have been higher baseline adverse outcomes. Linear regression (Fig. 2) visualises treatment effects beyond mild OGTT maternal plasma glucose levels.
GDM women had the relative risk of neonatal hypoglycaemia of 2.66. This may be the result of detection bias. The local protocol mandates screening for neonatal hypoglycaemia of all GDM women and neonates with a birth weight > 4000 g. In a study of neonates with no risk factors of hypoglycaemia, close serial monitoring within the 24 hours showed a neonatal hypoglycaemia rate, as defined by < 2.6 mmol/L, of 14%, which is similar to our rate of 15.8%28.
Increased rates of induced deliveries and consequently lower birth gestational age of GDM subjects is perhaps due to clinicians’ perception of higher risk for GDM women.
Of GDM women, 61.1% were treated with insulin (Table 4). Nearly all of those on insulin were on Isophane insulin, aimed to treat elevated fasting CBG. Our usage of insulin is higher than reported in randomised control studies: ACHOIS20 of 20%, and MFMU 21 of 7.6%. ADIPS treatment targets were tighter and so were the thresholds for the intensification of therapy. Moreover, as the effective treatment target for the fasting CBG level was well within our population’s reference intervals the extent of basal insulin usage was unsurprising.
Limitations and strengths
Our method may not be suitable for centres using the currently recommended citrate tubes for OGTT plasma glucose. Samples using citrate tubes would report fasting glucose levels of approximately 10% higher23. The HAPO glucose categories will no longer represent equivalent risks of adverse outcomes, as the HAPO study did not use citrate tubes10. Furthermore, quasi-randomisation partially relies on variations in pre-analytical handling, which would be reduced with citrate tubes.
Our study has limitations. As a retrospective design we did not have complete data on GDM risk factors such as family history of diabetes mellitus, smoking; in addition to data on maternal weight gain, and maternal hypertension. We did not record adherence to therapy with records of CBG but the high medication usage rate indicate that our GDM subjects were treated to ADIPS suggestions. We did not search for maternal hypoglycaemia in the database, but as the author was also the clinician involved there was no recollection of severe hypoglycaemia.
As a single centre study design with a predominately Caucasian population our results may not be applicable to all centres. Obstetric practices such as rates of caesarean section may vary between and within countries8 and therefore our risk reduction of caesarean sections may not be able to be extrapolated.
LGA/SGA outcomes relied on age standardised growth charts not adjusted for ethnic composition. However, as our population was predominately Caucasian there would have been marginal impact.
Our study may have been underpowered to determine differences in incidence rates for some less common end points such as shoulder dystocia.
We expect the loss of subjects by birthing at other centres to be rare, as our hospital is a referral centre for high-risk pregnancies for smaller regional hospitals. The closest hospital able to deliver high-risk pregnancies is more than 90 km from BHS. We did not see women expected to deliver at the local private hospital.
Subjects was analysed on an intention-to-treat basis irrespective of the adherence to therapy. Since our subjects were treated contemporaneously and by the same group of health professionals, our study is unlikely to be confounded by evolving medical practises.
It is our view that our study design is particularly suited to the study of GDM. This method, which can be performed in retrospective and prospective cohorts, will allow other centres with differing populations and obstetric practices to assess the local effectiveness of GDM treatments. Furthermore, this quasi-experimental design may be suitable to evaluate the effects GDM interventions on childhood adiposity and glucose tolerance.