Deep Learning to Improve Interpretation of Intrapartum Fetal Monitoring - PROJECT SUMMARY Electronic fetal monitoring (EFM) is used in greater than 85% of births in the United States, with the goal of allowing clinicians to detect changes in the fetal heart rate that may indicate fetal academia, enabling them to intervene prior to irreversible fetal injury. EFM was disseminated into practice prior to robust assessment of its efficacy, and decades of work since have shown that EFM is limited in achieving its intended goal of preventing intrapartum hypoxic-ischemic injury, and is associated with a significant increase in obstetric intervention, especially cesarean delivery. Much work has been done to develop guidelines and standardized frameworks for EFM interpretation in the hopes of improving both neonatal outcomes and the precision of obstetric interventions, but nearly 175,000 cesarean deliveries continue to be performed unnecessarily in the US annually due to false-positive interpretations of EFM. In response to many of these challenges, computerized interpretation of EFM has been explored since the 1980s. Unfortunately, the existing software programs, which are largely designed to detect the same EFM features that clinicians do, have not demonstrated clinical benefit in randomized controlled trials. In answer, our group hypothesized that a novel, data-driven deep learning approach could potentially detect meaningful data patterns in EFM, beyond those features that clinicians or other software programs recognize, that could help improve the predictive accuracy of EFM. Accordingly, we developed a deep-learning model that uses EFM data to predict fetal acidemia. The best performing model achieved an AUROC of 0.85 at an umbilical cord gas pH threshold of <7.05. These initial results provide proof-of-concept that a purely data-driven model can achieve promising predictive performance. Given this exciting start, we are now seeking to further validate and refine this model, and assess its potential role in clinical care. Accordingly, we propose to 1) Improve and refine our preliminary AI-EFM model by leveraging large, diverse datasets 2) compare the performance of the model to clinical interpretation (gold standard) and 3) assess the acceptability and perceived barriers to use of such a model in intrapartum patient care by conducting qualitative interviews with clinicians. The completion of this proposal and each of these aims are essential next steps in moving this innovative technology towards impactful clinical application and future evaluation in a randomized controlled trial. This work has the potential to enhance the accurate detection of true fetal distress warranting obstetric intervention, while also reducing the subjectivity and cognitive biases in interpretation of EFM. If rigorously tested and prudently deployed, this technology has the potential to improve outcomes for millions of birthing patients and neonates.
