Predicting Neonatal Pathologic Disposition From Cardiotocography Using Machine Learning [25A]

Abstract

INTRODUCTION: Continuous cardiotocography (CTG) is widely used to monitor fetal heart rate and uterine contractions as a means of detecting intrapartum fetal distress. CTG use is associated with reductions in neonatal seizures and with increases in rates of cesarean sections (C-sections) and instrumental vaginal births. Given that unnecessary C-sections are associated with higher rates of morbidity and mortality, correct identification of C-section candidates is critical. However, agreement in CTG interpretation remains elusive, with high inter-and intra-observer variability in interpretation of tracings, which may lead to inappropriate interventions. Further, machine learning (ML) advances in other vital monitoring analytics have not been matched in CTG analysis. The objective of this study was to develop a cardiotocographic ML algorithm, a random forest (RF) model, which accurately predicts neonatal pathologic disposition to inform obstetric intervention. METHODS: We trained RF models to classify neonatal condition as healthy or pathologic on 13 summary statistics (features) from 1,831 continuous CTGs from the University of Porto SisPorto dataset. All measures of performance were calculated as out-of-bag (OOB), a statistical proxy for cross-validation, over 128 weak learner models. RESULTS: Our model identified healthy CTGs as not requiring obstetric intervention with an OOB sensitivity of 82.4%, a specificity of 99.4%, an accuracy of 97.9%, and a C-statistic of 0.908. CONCLUSION: Our model's OOB performance exceeds that of current methods and could dramatically reduce the rate of cesarean sections if widely implemented. The model is accurate, rapid, readily deployable, and extensible. Further research should test our algorithm on a larger population and in real-time.