Utilizing machine learning to predict unplanned cesarean delivery.

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Unplanned cesarean delivery (uCD) remains a significant concern in modern obstetrics.
• Accurate prediction of uCD risk at labor admission is crucial for optimizing delivery management and patient outcomes.
• Existing prediction methods often lack comprehensive accuracy or real-time applicability.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting unplanned cesarean delivery (uCD).
• To identify key predictive features available at the time of labor admission for singleton pregnancies.
• To assess the clinical utility and accuracy of the ML model in a real-world setting.

Main Methods:

• A retrospective cohort study involving 73,667 women with singleton vertex pregnancies (≥34 weeks gestation) admitted for vaginal delivery.
• Data from March 2011 to May 2019 was used for training (80%) and validation (20%), with a separate cohort (June 2019-April 2021) for testing.
• Feature selection was performed using Random Forest, and an XGBoost model was developed and evaluated.

Main Results:

• The final ML model incorporated 13 key features and demonstrated strong predictive performance.
• The XGBoost model achieved areas under the curve (AUC) of 0.874 (training), 0.839 (validation), and 0.840 (testing).
• The model exhibited a 65% positive predictive value for uCD in the highest risk group and >99% negative predictive value in the lowest risk group.

Conclusions:

• The developed ML model provides clinically useful risk stratification for unplanned cesarean delivery.
• The model's accuracy is maintained across gestational weeks 34-42 and various clinical risk groups.
• This predictive tool can enhance clinical decision-making and support individualized patient counseling during labor.