Dense phenotyping from electronic health records enables machine-learning-based prediction of preterm birth

Abstract: Identifying pregnancies at risk for preterm birth, one of the leading causes of worldwide infant mortality, has the potential to improve prenatal care. However, we lack broadly applicable methods to accurately predict preterm birth risk. The dense longitudinal information present in electronic health records (EHRs) is enabling scalable and cost-efficient risk modeling of many diseases, but EHR resources have been largely untapped in the study of pregnancy. Here, we apply machine learning to diverse data from EHRs to predict singleton preterm birth. Leveraging a large cohort of 35,282 deliveries, we find that a prediction model based on billing codes alone can predict preterm birth at 28 weeks of gestation (ROC-AUC=0.75, PR-AUC=0.40) and outperforms a comparable model trained using known risk factors (ROC-AUC=0.59, PR-AUC=0.21). Our machine learning approach is also able to accurately predict preterm birth sub-types (spontaneous vs. indicated), mode of delivery, and recurrent preterm birth. We demonstrate the portability of our approach by showing that the prediction models maintain their accuracy on a large, independent cohort (5,978 deliveries) with only a modest decrease in performance. Interpreting the features identified by the model as most informative for risk stratification demonstrates that they capture non-linear combinations of known risk factors and patterns of care. The strong performance of our approach across multiple clinical contexts and an independent cohort highlights the potential of machine learning algorithms to improve medical care during pregnancy.