Endocrine Abstracts

Background and aim: Gestational Diabetes Mellitus (GDM) is often diagnosed at 24-28 weeks of pregnancy when the fetal phenotype is already altered. We aimed to develop a machine learning model based on clinical variables, lifestyle features and genetic markers for GDM prediction in the first trimester of pregnancy based on the 2013 World Health Organization (WHO) criteria.

Methods: Using multivariable logistic regression analysis, different models to predict GDM were developed based on clinical variables from early pregnancy (age, pre-pregnancy body mass index (BMI), arterial hypertension, a history of GDM, impaired glucose tolerance, polycystic ovary syndrome, family history of type 2 diabetes, and parity), lifestyle features (food consumption, physical activity and smoking habits assessed through questionnaires), genetic markers (number of risk alleles in rs10830963 of the MTNR1B gene) and their combination. The input data were obtained from 1050 pregnant women participating in prospective studies performed in the Almazov National Medical Research Centre (655 GDM cases and 395 controls). Receiver operating characteristic (AUC) analysis assessed the model’s performance with eight-fold cross-validation.

Results: C-statistics for logistic regression models were as follows: clinical covariates alone: 0.690 (95% CI: 0.658 to 0.722) and 0.688 (95% CI: 0.647 to 0.729) for an eight-fold cross-validated assessment of the score; rs10830963 alone: 0.597 (95% CI: 0.565 to 0.629) and 0.546 (95% CI: 0.375 to 0.717) cross-validated; combination of clinical covariates and rs10830963: 0.721 (95% CI 0.690 to 0.752) and 0.715 (95% CI 0.672 to 0.759); combination of clinical covariates and lifestyle features: 0.806 (95% CI 0.779 to 0.833) and 0.802 (95% CI 0.724 to 0.881) cross-validated; combination of clinical covariates (age, pre-pregnancy BMI, a history of GDM), rs10830963 and lifestyle features (the frequency of pre-pregnancy consumption of meat, bread and alcohol): 0.823 (95% CI: 0.797-0.849) and 0.813 (95% CI: 0.734-0.892) cross-validated.

Conclusions: A first trimester machine learning-based model, which incorporates classical risk factors and novel biomarkers, has a high accuracy to predict GDM based on the 2013 WHO criteria.