A machine learning algorithm for predicting maternal readmission for hypertensive disorders of pregnancy

Abstract

Maternal postpartum hypertensive emergency is a major cause of maternal mortality and maternal readmission, yet prediction of women who require readmission is limited with false negatives and false positives. This study aimed to develop and validate a predictive algorithm for maternal postpartum readmission from complications of hypertensive disorders of pregnancy using machine learning. We performed a cohort study of pregnant women delivering at a single institution using prospectively collected clinical information available from the electronic medical record at the time of discharge. Our primary outcome was readmission within 42 days of delivery for complications of hypertensive disorders of pregnancy. The data set was divided into a derivation and a separate validation cohort. In the derivation cohort, 10 independent data sets were created by randomly suppressing 10% of the population, and then clinical features predictive of complications of hypertensive disorders of pregnancy were analyzed using machine learning to optimize the area under the curve. Once an optimal model was determined, this model was then validated using a second independent validation cohort. A total of 20,032 delivering women with 238 readmissions for complications of hypertensive disorders of pregnancy (1.2%) were included in the derivation cohort. The validation cohort consisted of 5823 women with 82 readmissions for complications of hypertensive disorders of pregnancy (1.4%). The demographics were similar between the 2 populations as was the test performance characteristics (area under the curve, 0.85 in the derivation cohort vs 0.81 in the validation cohort). Both the derivation and validation algorithms used 31 clinical features that were found to be comparably predictive in both models. In this evaluation of a machine learning algorithm, readmission for complications of hypertensive disorders of pregnancy can be predicted with reasonable accuracy using clinical data at the time of discharge. Validation of this model in other care settings is necessary to validate its utility.