Odds, risks and appropriate diagnosis of gestational diabetes

The International Association of the Diabetes and Pregnancy Study Groups (IADPSG) diagnostic process and criteria for gestational diabetes mellitus (GDM) are designed to identify women at increased risk of a range of adverse pregnancy outcomes related to maternal hyperglycaemia; in particular, excessive fetal growth and fetal hyperinsulinaemia.1 The relationship between maternal hyperglycaemia and adverse outcomes is continuous, and more than one elevated glucose level from oral glucose tolerance testing equates to higher glucose exposure; however, one elevated glucose value is sufficient to impart a higher risk of pregnancy complications.2

We note with concern the recent article in the Journal by d'Emden, which proposes “a more statistically valid basis for diagnosing GDM”.3 It suggests that the criteria for GDM diagnosis proposed in 2010 by the IADPSG1 would result in up to 50% of women with a single elevated oral glucose tolerance test result being “inappropriately diagnosed with GDM as they do not meet the agreed risk threshold”. We consider these statements to be incorrect and offer the following arguments in rebuttal.

Interpreting odds ratios and confidence intervals

No distinction appears to have been made between “odds” and “risks” in the statistical analysis. This is incorrect4,5 and leads to a variety of erroneous conclusions.

Underlying the arguments in the article is an incorrect interpretation of odds ratios and their 95% confidence intervals (CIs). Odds ratios vary on a non-linear scale from zero to infinity, with 1.0 being the central value (point of no difference). If the 95% CI of an odds ratio does not cross unity, the association described is considered significant at the 5% level.4,5

It is wrong to suggest that 50% of subjects actually carry a risk lower than the reported point estimate. Rather, the confidence limits mean that, were a study to be repeated 1000 times, the actual odds ratio obtained would be expected to be less than the reported lower 95% confidence limit on 25 occasions.

D'Emden proposed that the lower 95% confidence limit should be used in place of the point estimate of the odds ratio in the determination of diagnostic thresholds for GDM. We strongly contend that the lower 95% CI is intrinsically less likely to be a true reflection of the prevailing odds. No statistical support was quoted by d'Emden, and such a method does not appear to be accepted in the literature. In practical terms, had the IADPSG taken this approach, it would have given a similar result to selecting a higher odds ratio threshold, based on the conventional point estimate. The IADPSG specifically chose a point estimate of 1.75 for the determination of thresholds. The consensus process specifically considered and eventually rejected alternative prespecified odds ratio thresholds of 1.5 and 2.0.

Understanding the comparisons made using data

We believe that d'Emden misinterpreted the data that he presented in Box 2.3 The odds ratios shown there, extracted from an article by two of us (A R D and B E M),2 relate to a comparison of women diagnosed post hoc as having GDM by IADPSG criteria according to various combinations of oral glucose tolerance test results ≥ threshold (1, 2 or 3 abnormal values) versus women classified post hoc as not having GDM. They bear no relation to the continuous logistic regression model that was used to determine the IADPSG-recommended diagnostic thresholds. The comparison made is entirely different. For the reasons stated above, it is incorrect to say, for example, that for women with only one elevated glucose value, “50% ... did not reach the agreed risk threshold”.

As previously outlined,1 the diagnostic thresholds were defined as the average glucose values at which the odds for the three selected outcomes reached 1.75 times the estimated odds of these outcomes at mean glucose values, based on fully adjusted logistic regression models. Further, an individual woman does not have her “own” odds ratio. She, or her baby, either experience a particular outcome or they do not. As presented previously,2 the comparison after assignment of a diagnostic label at this point should be between groups of women with and without GDM as defined by the recommended criteria.

Changing the criteria may not increase frequency of GDM diagnosis

Outside these statistical concerns, we note that many objections to the IADPSG criteria appear to be founded on a presumption of more frequent GDM diagnoses with a change to the IADPSG approach from historical criteria (in Australia, predominantly the 1991 ad hoc criteria).6 Although an increase in the projected frequency of GDM (from 9.6% to 13.0%) in a cohort of women from Wollongong, New South Wales, was demonstrated with this change in criteria,7 this was predominantly due to increased GDM frequency in women treated in the private sector, suggesting possible sociodemographic influences. Another Australian study, among Indigenous women from Cape York, Queensland, reported a slightly lower GDM frequency (decreased from 14.2% to 13.4%) after a change to IADPSG criteria, but noted a threefold overall increase in GDM due to improved testing practices.8 A cohort study from Vietnam that compared the same criteria also showed a marginally lower GDM frequency (from 20.8% to 20.4%) with a change from the ad hoc criteria to the IADPSG approach.9

The IADPSG criteria include a lower threshold for fasting glucose and a higher 2-hour glucose threshold than the previous Australian ad hoc criteria, and introduce a new 1-hour threshold value. The frequency of GDM detected with IADPSG criteria will inevitably vary between population groups, likely due to different underlying prevalences of fasting hyperglycaemia as opposed to post-glucose load hyperglycaemia across varying ethnicities.

The Hyperglycemia and Adverse Pregnancy Outcome (HAPO) study

Considering the HAPO study data in particular, post hoc risk ratios for a wide range of outcomes, associated with varying diagnostic thresholds based on prespecified values (1.5, 1.75 and 2.0) of adjusted odds ratios, have been published in detail.10 From review of these published data (summarised in the Box) it can be seen that the use of the IADPSG-recommended odds ratio threshold of 1.75 to determine GDM diagnostic thresholds leads to actual risk ratios approximating or greater than 2.0 when comparing women classified post hoc as GDM versus non-GDM (normal).

Measures selected by the IADPSG

D'Emden stated that neonatal adiposity (defined as per cent body fat > 90th centile) was not a predefined outcome of the HAPO study. Although not mentioned in the primary HAPO results article,11 it was clearly identified as a primary outcome from the earliest descriptions of HAPO study methods.12 Full data regarding per cent body fat and cord C-peptide were available for 73% of the total HAPO cohort of 23 316 neonates, representing the largest dataset of this type ever collected.

The inclusion of cord C-peptide > 90th centile as one criterion used for the determination of diagnostic thresholds was criticised on the basis that “it is not a routine test performed in clinical practice”. We would strongly dispute that this should be a primary reason for inclusion or exclusion of a particular measure in the definition of GDM. In fact, all the selected measures (birthweight > 90th centile, per cent body fat > 90th centile and cord C-peptide > 90th centile) are based on understanding of the pathophysiology connecting maternal hyperglycaemia to pregnancy complications. They are measures of key features of diabetic fetopathy (excess growth, excess adiposity and hyperinsulinism, respectively). Their inclusion relates directly to this nexus and we strongly contend that this is well grounded, irrespective of current clinical practice, which arguably may fall short of qualifying as a gold standard.

D'Emden also suggested that these measures were selected by the IADPSG “because they had had the greatest difference between the at-risk group and the normal group”. As participants in the IADPSG consensus process, we categorically deny this assertion. By definition, any consensus process is post hoc, as it must consider information that has already been collected. Although HAPO study investigators were involved in the consensus process, they comprised a minority of the IADPSG consensus panel, which was clearly distinct from the HAPO group. Therefore, they were not constrained to the use of primary HAPO outcome variables in selection of diagnostic thresholds. We can affirm that the measures selected and the nature of further statistical calculations were decided a priori. Further, until the thresholds were selected, there were no at-risk and normal groups to compare, so this assertion is incorrect.

Conclusion

The thresholds recommended by the IADPSG, although greatly influenced by the HAPO study, also took into consideration other available epidemiological data13-16 as well as the inclusion criteria and results of the two major randomised clinical trials in this area.17,18 In common with any dichotomous classification of essentially continuous variables, the thresholds for GDM diagnosis are, by nature, arbitrary. Further, pregnancy complications have multiple potential underlying causes. No set of criteria for diagnosing GDM could ever fully separate women at risk and not at risk of pregnancy complications. However, the IADPSG criteria do represent a considered consensus view from an expert group and are underpinned by high-level statistical expertise.

The World Health Organization agreed to adopt the IADPSG criteria,19 which have been accepted for local implementation in Australia by a consensus group including the Royal Australian and New Zealand College of Obstetricians and Gynaecologists, the Australian Diabetes Society, Royal College of Pathologists of Australia and the Australasian Diabetes in Pregnancy Society. From January 2015, the criteria were introduced into clinical practice in Queensland, the Australian Capital Territory and variably across other states. Other local groups such as the Royal Australian College of General Practitioners have dissented, making complete local agreement unlikely. However, we continue to support the underlying methods used to develop the IADPSG consensus criteria and contend that they represent a well reasoned approach to the diagnosis of GDM.