170-LB: Predictors of Diabetes First Diagnosed in Pregnancy: A Machine-Learning Model

Abstract

Introduction: Pregestacional (PG) DM first diagnosed in pregnancy (diabetes in pregnancy, DIP) carries poor outcomes, which may be prevented by DM diagnosis before pregnancy. Aim: To identify risk factors (RF) for DIP that may help to diagnose DM before pregnancy. Methods: Retrospective cohort, type 2 DM women; specialized prenatal care facility; public hospitals; Brazil. Data - age, ethnicity, schooling, DM family history (FH), previous obstetric history (number of pregnancies, GDM, preeclampsia (PE), macrosomia) - and PG BMI were collected from May 2005 to July 2019. Univariable analyses (student T and chi-square tests) with SPSS vs.18; two machine learning (ML) models (model 1: all variables; model 2: age, GDM, FH, number of pregnancies and PG BMI) with software Orange Data Mining vs. 3; random sampling statistics. Results: Of the 466 women, 157 (33.7%, 95% CI 29-38%) had DIP. Baseline characteristics/main outcomes were similar between DIP and PG DM women. For model 1, AUC was 0.526 and precision, 0.566. For model 2 (Figure), AUC was 0.525 and precision, 0.556. Age was the most important RF, followed by previous macrosomia in women ≤ 28 years and by PG BMI in those > 28 years. Conclusion: In type 2 women, RF identified by ML with the algorithms used were poor predictors of DIP diagnosis. They were the same as for known pregestational type 2 DM. Age was the most important determinant, followed by previous macrosomia (for DIP) and BMI (for PG DM). Disclosure M.A. de Campos: None. M.R. Oppermann: None. V.K. Genro: None. C.B. Leitao: None. V.N. Hirakata: None. A.J. Reichelt: None.