[Predictive models of gestational diabetes, a new prediction mode]

Recognized the value of gestational diabetes (GD) as a health problem, our aim in this work has been to analyze the diagnostic performance of the different today's existing criteria (GEDE, O'Sullivan and Carpenter) after the overload of 100 g of glucose and revise how to increase its efficiency.

We carried out a description of all the variables. In the analytical phase of the work, we used Pearson's Chi square to see if there were differences in the percentage of cases collected in each health center and the proportions contrast test to study the differences between the experimental prevalence. We completed binary logistic regression models using as result variable having or not having gestational diabetes (yes/no) and as predictors the four measurements of the curve with 100 g of glucose overload. To decide which model was better, the stepwise backward-forward analysis and the surface of the ROC curve generated by each of them were considered.

We obtained a sample of 170 pregnant women from six different Primary Care Area of Seville health centers who had shown a positive O'Sullivan test with a median age of 35 years. There were significant differences in the prevalence proportions according to the criteria used: GEDE/O'Sullivan p < 0.001; GEDE/Carpenter p < 0.001. Logistic models with three and four predictor variables were equal in discriminatory diagnostic capacity when the GEDE criteria were used (area under the ROC curve = 0.96, 95% CI: 0.93-0.98). The back-forward stepwise analysis stayed with the three-variable model as the most parsimonious. The same did not occur when applying the other two criteria.

Regarding an observational design, we state that there are significant differences in the prevalence proportions observed according to the criteria applied (p < 0.001) and we can also support that using the GEDE criteria, the taking of the third hour could be dispensed with, based on Bayesian criteria and the application of the ROC curve analysis.