Prediction of pregnancy associated hypertension in Asian women using machine learning algorithms

This study investigated capacity of machine learning prediction of pregnancy associated hypertension (PAH), based on early pregnancy data, in Asian women. We used machine learning approach to predict PAH and preterm PAH on retrospective data of 35,004 pregnancies in multicentres of Korea. After randomly dividing the data into training and test sets at a 70:30 ratio, predictive models for PAH were created by three methods; gradient boosting models using variables by 1) prespecified by the American College of Obstetrics and Gynecology (ACOG), 2) the Shapley values (original) and 3) boruta algorithm from the full variables. Performance was evaluated with area under the curve (AUC) and area under the precision-recall curve (auPR), in total and restricted populations after excluding women who received aspirin during pregnancy. The incidence of PAH was 8.11%, with 3.3% of preterm PAH. At baseline, performance of models for PAH with ACOG, original, and boruta variables achieved AUCs of 0.724, 0.8, and 0.782, and auPRs of 0.236, 0.359, and 0.323, respectively. At 11-13 weeks of gestation, performance with original and boruta variables achieved AUCs of 0.821 and 0.816, and auPRs of 0.412 and 0.416. Selected variables by boruta were maternal age, prepregnancy body mass index, MoMs of PAPP-A and hCG, pre-existing hypertension, family history of hypertension, blood pressure at the first trimester, history of PAH, preterm birth, postpartum bleeding, Caesarean section, and the number of live children. At baseline, performance for preterm PAH with ACOG, original, and boruta variables achieved AUCs of 0.727, 0.791, and 0.785, and at 11-13 weeks of gestation, performance with original and boruta variables achieved AUCs of 0.817 and 0.810, respectively. In the restricted population, results were similar. In conclusion, machine learning models in early pregnancy demonstrated moderate to high performances, with the use of routine clinical variables. (Korea Health Industry Development Institute, HI21C1300).