487 Early prediction of pregnancy-associated hypertension : Development of novel strategies with semi-supervised learning

Abstract

To identify high-risk women for pregnancy-associated hypertension (HTN) in early pregnancy, clinical guidelines recommend risk factors, however these result in low accuracy. A few studies tried to predict pregnancy-associated HTN using traditional machine learning methods but failed to show robust performance. Graph-based semi-supervised learning (SSL) is a powerful tool to perform prediction with the propagation of the label information along with the structure of patients’ graph that was constructed according to their characteristics. Here, we selected important clinical variables using feature selection methods and developed a prediction model for pregnancy-associated HTN using graph-based SSL. This is a secondary analysis from a prospective study of healthy pregnant women. Pregnancy-associated HTN was defined as gestational HTN, preeclampsia, eclampsia, and superimposed preeclampsia. For ranking/selecting variables among clinical variables retrieved in the first trimester, four feature selection methods were applied. For developing the prediction model, a patients’ network/graph was constructed with graph-based SSL (label propagation)(Fig1A). We compared the prediction performances with five machine learning methods (graph-based SSL, logistic regression, support vector machine, random forest, and gradient boosting) using three different variable sets; 1) classical risk factors; 2) selected important variables; 3) all 32 variables (Fig1B). Each experiment was carried out with hold-out (train: test=7:3) and was repeated100 times. In the study population, 2.4% of women (33/1347) developed pregnancy-associated HTN. With aggregating ranks, top 9 important variables were selected according to the importance of describing pregnancy-associated HTN (Fig2A). The graph-based SSL using top 9 important variables achieved the best average prediction performance (AUROC 0.80, Fig2B). A more robust and accurate prediction model for pregnancy-associated HTN can be built using selected top 9 variables and graph-based SSL. The proposed model needs to be evaluated in further studies.View Large Image Figure ViewerDownload Hi-res image Download (PPT)