Random Forests for Accurate Prediction of the Risk of Hypertensive Disorders of Pregnancy at Term [A208

Abstract

INTRODUCTION: Hypertensive disorders of pregnancy (HDP) complicate 10% of all pregnancies and confer elevated risks of subsequent conditions, such as chronic hypertension and diabetes. Risk models that accurately predict HDP can improve clinical management by risk stratification, allowing earlier interventions and closer interval follow-up. This study develops and internally validates a risk prediction model for the development of HDP in the third trimester. METHODS: This analysis was conducted using the Eunice Kennedy Shriver Institute of Child Health and Human Development / Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be (nuMoM2b) data for approximately 9,000 low-risk nulliparous patients. HDP was the primary outcome for this analysis, defined as gestational hypertension, mild/severe preeclampsia, eclampsia, or HELLP syndrome (hemolysis, elevated liver enzymes, and low platelets) as an indication for admission. We included 518 potential covariates of interest, including outpatient vital signs, medication usage, patient and family medical history, and laboratory information. We trained a random forest predictor model on 80% of this dataset, and further refined it using recursive feature elimination to restrict the number of input variables to 20 for a reduced model. RESULTS: Our full and reduced models predicted HDP with area under the receiver operating characteristic (AUROC) scores of 0.82 and 0.81 respectively. Performance evaluation of the reduced model by race showed poorer performance for non-Hispanic Black (0.76), Hispanic (0.79), and non-Hispanic White (0.80) patients than overall. CONCLUSION: The reduced model has been deployed for demonstration at hypertension.delfina.com. Admission for HDP can be predicted with excellent discriminative ability using random forest prediction modeling. External validation and integration with electronic health records is needed to assist providers in triaging and informing at-risk patients.