Prediction of gestational diabetes mellitus by multiple biomarkers at early gestation

It remains unclear which early gestational biomarkers can be used in predicting later development of gestational diabetes mellitus (GDM). We sought to identify the optimal combination of early gestational biomarkers in predicting GDM in machine learning (ML) models.

This was a nested case-control study including 100 pairs of GDM and euglycemic (control) pregnancies in the Early Life Plan cohort in Shanghai, China. High sensitivity C reactive protein, sex hormone binding globulin, insulin-like growth factor I, IGF binding protein 2 (IGFBP-2), total and high molecular weight adiponectin and glycosylated fibronectin concentrations were measured in serum samples at 11-14 weeks of gestation. Routine first-trimester blood test biomarkers included fasting plasma glucose (FPG), serum lipids and thyroid hormones. Five ML models [stepwise logistic regression, least absolute shrinkage and selection operator (LASSO), random forest, support vector machine and k-nearest neighbor] were employed to predict GDM. The study subjects were randomly split into two sets for model development (training set, n = 70 GDM/control pairs) and validation (testing set: n = 30 GDM/control pairs). Model performance was evaluated by the area under the curve (AUC) in receiver operating characteristics.

FPG and IGFBP-2 were consistently selected as predictors of GDM in all ML models. The random forest model including FPG and IGFBP-2 performed the best (AUC 0.80, accuracy 0.72, sensitivity 0.87, specificity 0.57). Adding more predictors did not improve the discriminant power.

The combination of FPG and IGFBP-2 at early gestation (11-14 weeks) could predict later development of GDM with moderate discriminant power. Further validation studies are warranted to assess the utility of this simple combination model in other independent cohorts.