Results
We included 203 women with an HbA1c measure at 1 year postpartum; of
whom 71 (35%) had impaired glucose tolerance. Nine women (4%) had
HbA1c ≥ 6.5%, consistent with type 2 diabetes at 1 year postpartum.
Overall, our cohort was diverse (31% Hispanic and 10% non-Hispanic
Black), women were overweight (median BMI 29.6; IQR 25.7, 34.9), over
half required insulin or medication for GDM management during pregnancy,
and 12% were diagnosed with GDM at <24 weeks’ gestation
(Table 1). Hispanic women were over represented among those who
developed impaired glucose tolerance (48% with impaired glucose
tolerance versus 21% without), while white women were under-represented
(32% with impaired glucose tolerance vs 57% without). Women who
developed impaired glucose tolerance were more likely to report a family
history of type 2 diabetes (64% vs 48%), be obese pre-pregnancy (70%
vs 36%), be diagnosed with GDM at <24 weeks’ gestation (23%
vs 6%), have experienced GDM in a previous pregnancy (34% vs 17%),
and require insulin or medication for GDM management in their most
recent pregnancy (67% vs 52%).
Eight candidate predictors were selected for inclusion into the final
multivariable model. They included continuous pre-pregnancy weight,
continuous pre-pregnancy BMI, pre-pregnancy obesity (BMI ≥ 30
kg/m2, versus not), GDM in a previous pregnancy (yes
versus no or first pregnancy), Hispanic ethnicity (yes versus no), GDM
diagnosis before 24 weeks’ gestation (yes versus no), and continuous
measures of fasting and 2-hour plasma glucose at 2 days postpartum. Wea priori specified that candidate predicators would be included
into the final model if they were selected by a Lasso model in
>60% of imputed datasets; however, all but one of the
predictors included in the final model were selected in 100% of imputed
data sets (Supplemental Figure S3). The final multivariable model had an
AUC of 0.81 (95% CI 0.74, 0.87; Table 2).
In sensitivity analyses, we examined the robustness of the final
multivariable model in several ways. First, we excluded 2-day postpartum
glucose data because glucose testing at this time is not standard of
care. Second, we removed each of the weight and BMI variables
individually to evaluate the impact of having highly correlated
variables (e.g. weight and BMI) in the final model. Third, we removed
the indicator for Hispanic ethnicity since ethnicity is not a reliable
indicator of genetic differences and this variable likely captures a
complex mix of ethnicity and social processes which may influence
Hispanic women’s risk of impaired glucose tolerance.25Finally, we replaced Hispanic ethnicity in the final model with family
history of type 2 diabetes to evaluate if a more direct measure of
genetic risk (e.g. family history) influenced model results. All
sensitivity analyses resulted in comparable or slightly reduced AUC
compared to the final multivariable model from the primary analysis
(Supplemental Table S2).
To identify women with impaired glucose metabolism at 1 year postpartum,
we examined several possible cut-points of predicted probabilities from
the final multivariable model. A cut-point of 0.13 resulted in a
sensitivity of 96% and specificity of 35%, 44% PPV and 94% NPV to
identify women with impaired glucose tolerance at 1 year postpartum
(Table 3; Supplemental Figure S4). This cut point would identify 76% of
the population for potential postpartum intervention, miss predicting
impaired glucose tolerance in 2% of women, and incorrectly predict
impaired glucose tolerance in 42% of women. A cut-point of 0.65
resulted in specificity of 96%, sensitivity of 39%, PPV 8% and NPV
75%. This cut-point would identify 17% of the population for
additional intervention, miss predicting impaired glucose tolerance in
21% of women, and incorrectly predict impaired glucose tolerance in 3%
of women.
To assess calibration of the final multivariable model, we compared
observed and predicted probabilities of the outcome. Observed and
predicted probabilities were similar for women in the lower two
quartiles of risk, but were overestimated for women in the highest two
quartile or risk, which may indicate model overfitting because of the
small number of events in this sample (Supplemental Figure
S5).22