Prediction of preterm birth from cervical length measurements in twin pregnancies using machine learning

Multiple Cervical Length (CL) measurements are typically acquired throughout the course of twin pregnancy to detect the early stages of labour and identify pregnancies at a high risk of preterm delivery. This study uses Machine-Learning (ML) approaches to determine the optimal timing of repeated CL measurements when used for predicting spontaneous preterm birth (sPTB) in twin pregnancies. Serial CL measurements from ultrasounds performed between 16 and 28 weeks of gestation were retrospectively acquired from 2,095 patients carrying twin pregnancies. These measurements were used for creating several CL feature sets, which were subsequently evaluated for their utility in predicting PTB < 37, sPTB < 37, sPTB < 34, and sPTB < 32 weeks. The highest accuracies for predicting sPTB < 37, sPTB < 34, and sPTB < 32 were found for the Logistic Regression model, which performed at 58%, 63%, and 73%, respectively. Post-hoc analysis showed that using multiple CL measurements did not significantly improve the sPTB prediction accuracy, irrespective of the ML model. Specifically, a single CL measurement at 18-20 weeks of gestation was sufficient for predicting sPTB < 32 weeks with the same accuracy. Future work should expand patient cohorts by including early CL measurements and investigating the time between a CL exam and sPTB from a regression standpoint.