Real-time data analysis using a machine learning model significantly improves prediction of successful vaginal deliveries

Abstract

Background: The process of childbirth is one of the most crucial events in the future health and development of the offspring. The vulnerability of parturients and fetuses during the delivery process led to the development of intrapartum monitoring methods and to the emergence of alternative methods of delivery. However, current monitoring methods fail to accurately discriminate between cases in which intervention is unnecessary, partly contributing to the high rates of cesarean deliveries worldwide. Machine learning methods are applied in various medical fields to create personalized prediction models. These methods are used to analyze abundant, complex data with intricate associations to aid in decision making. Initial attempts to predict vaginal delivery vs cesarean deliveries using machine learning tools did not utilize the vast amount of data recorded during labor. The data recorded during labor represent the dynamic process of labor and therefore may be invaluable for dynamic prediction of vaginal delivery.

Objective: We aimed to create a personalized machine learning-based prediction model to predict successful vaginal deliveries using real-time data acquired during the first stage of labor.

Study design: Electronic medical records of labor occurring during a 12-year period in a tertiary referral center were explored and labeled. Four different models were created using input from multiple maternal and fetal parameters. Initial risk assessments for vaginal delivery were calculated using data available at the time of admission to the delivery unit, followed by models incorporating cervical examination data and fetal heart rate data, and finally, a model that integrates additional data available during the first stage of labor was created.

Results: A total of 94,480 cases in which a trial of labor was attempted were identified. Based on approximately 180 million data points from the first stage of labor, machine learning models were developed to predict successful vaginal deliveries. A model using data available at the time of admission to the delivery unit yielded an area under the curve of 0.817 (95% confidence interval, 0.811-0.823). Models that used real-time data increased prediction accuracy. A model that includes real-time cervical examination data had an initial area under the curve of 0.819 (95% confidence interval, 0.813-0.825) at first examination, which increased to an area under the curve of 0.917 (95% confidence interval, 0.913-0.921) by the end of the first stage. Adding the real-time fetal heart monitor data provided an area under the curve of 0.824 (95% confidence interval, 0.818-0.830) at first examination, which increased to an area under the curve of 0.928 (95% confidence interval, 0.924-0.932) by the end of the first stage. Finally, adding additional real-time data increased the area under the curve initially to 0.833 (95% confidence interval, 0.827-0.838) at the first cervical examination and up to 0.932 (95% confidence interval, 0.928-0.935) by the end of the first stage.

Conclusion: Real-time data acquired throughout the process of labor significantly increased the prediction accuracy for vaginal delivery using machine learning models. These models enable translation and quantification of the data gathered in the delivery unit into a clinical tool that yields a reliable personalized risk score and helps avoid unnecessary interventions.

Keywords: cesarean delivery; machine learning; obstetrics; personalized medicine; prediction; trial of labor; vaginal delivery.