Main Findings
Our predictive model for use around the time of delivery demonstrated
good discriminatory ability (AUC 0.81; 95% CI 0.74, 0.87) and improved
prediction over reported OGTT values at 4-12 weeks postpartum (AUC 0.60;
95% CI 0.53, 0.67)14, the current standard of care,
to identify women at high risk for postpartum impaired glucose
tolerance. Compared to reported values for a postpartum OGTT, our model
also demonstrated improved sensitivity, specificity, PPV and NPV at
several cut-points to identify women with impaired glucose tolerance by
1 year postpartum.14 Our results provide proof of
concept that a risk score may be able to accurately identify women with
recent GDM around the time of delivery who are likely to progress to
impaired glucose tolerance postpartum. With less than 50% of women with
recent GDM receiving any postpartum glucose testing or
intervention,10,13 identifying high risk women around
the time of delivery could help to target follow-up testing and
intervention efforts in this high risk population.
Adults with impaired glucose metabolism are at high risk of progressing
to type 2 diabetes, making them an important population for early
intervention and prevention.26 To develop a clinically
useful model, we prioritized potential predictors that can be readily
collected in the prenatal and early postpartum period, and used a Lasso
regression, which is a data-driven predictive modeling algorithm, to
select covariates for inclusion in a final model. This algorithm has
been used to identify clusters of clinical predictors of progression to
type 2 diabetes in non-pregnant populations,27-29 but
have not been employed in populations with recent GDM. Our final model
included eight covariates that are known to correlate with the risk of
progression to impaired glucose tolerance in women with GDM including,
pre-pregnancy weight and BMI, previous GDM, early GDM diagnosis,
Hispanic ethnicity, and fasting and 2-hour plasma glucose at 2 days
postpartum.19,25,30-34
The predictive ability of the final model did not change meaningfully in
several sensitivity analyses, including excluding plasma glucose values
at 2 days postpartum, which is not currently standard of care. Model
findings were also robust when evaluating the exclusion of collinear
weight and BMI covariates within the model, suggesting that while these
variables are closely correlated, including both in the model overall
improves prediction, rather than adversely affecting the model’s
performance. Model performance did not change meaningfully when Hispanic
ethnicity was removed from the model or replaced with family history of
type 2 diabetes.25