Prediction of spontaneous preterm birth among twin gestations using machine learning and texture analysis of cervical ultrasound images

Background/Introduction: Preterm birth (PTB) is the main cause of neonatal mortality and morbidity in twin pregnancies. There is a need for new biomarkers to improve the predictive accuracy for PTB. The aim of this study was to use quantitative texture analysis of ultrasound images of the cervix combined with machine learning algorithms to detect patterns that are predictive of PTB in twins. Methodology: This proof-of-concept study involved a retrospective cohort of women with twin gestation who underwent serial monitoring of cervical length in a single referral center between 2016–2017. Ultrasound images of the cervix at 22+0–26+6 weeks of gestation were extracted and analyzed. An automated software module was developed to extract texture features from four regions of interest. Texture features were applied to a random forest classifier machine learning algorithm to predict PTB. Results: A total of 170 images from 98 women with twins (61/98 PTB) were analyzed. The predictive value for PTB, expressed as the area under the receiver operating characteristic curve (AUC), was higher for samples obtained from the external os versus internal os, and was highest for the anterior aspect compared with the posterior aspect of the external os (0.75, 95% CI 0.61–0.78). The sensitivity and specificity for PTB for samples obtained from the anterior external os were 69% and 70%, respectively. Conclusion: In this preliminary proof-of-concept study we have found a novel biomarker that may improve the prediction of PTB.