Prediction of cesarean delivery after induction of labor using machine learning (Random Forest)
TL;DR
A developed random forest-based model, which incorporates an intrapartum predictor (time interval to regular contractions), demonstrates high predictive performance, potential for external validation, and clinical utility, thereby supporting timely and informed decision-making regarding labor management.
Abstract
Background. Induction of labor (IOL) is a widely used obstetric intervention. However, the rate of cesarean section (CS) following IOL may exceed 50% in certain subgroups. Existing predictive models predominantly rely on antenatal factors, fail to account for nonlinear relationships, and do not always demonstrate sufficient discriminative ability. Aim. Development, validation, and interpretation of a predictive model for predicting the outcome of IOL in term pregnancy using machine learning methods. Materials and methods. This retrospective cohort study, based on a hospital registry (June 2024 – June 2025), included 1,670 women with singleton pregnancies with gestational age of ≥37 weeks who underwent IOL (amniotomy, oxytocin). Outcomes were defined as successful IOL (vaginal delivery) and unsuccessful IOL (CS). For preliminary feature selection, feature importance was calculated using logistic regression, logistic regression with L1 regularization (Lasso), and random forest to eliminate non-informative variables. The final predictive model was built using the random forest algorithm. SHAP analysis was used to interpret the influence of the selected features on the model predictions. Results. The CS rate was 21.1% (n=352). Seven features were included in the final model (in descending order of importance): the time interval from IOL initiation to the onset of regular uterine contractions, parity, estimated fetal weight by ultrasound, need for pre-induction of labor, maternal age, maternal body mass index, and gestational age. The model achieved an AUC-ROC of 0.904 (95% CI 0.846–0.951) on the test set, with a sensitivity of 75%, specificity of 92%, and net benefit of 0.184. SHAP analysis enabled interpretation of individual predictions. Conclusion. The developed random forest-based model, which incorporates an intrapartum predictor (time interval to regular contractions), demonstrates high predictive performance, potential for external validation, and clinical utility, thereby supporting timely and informed decision-making regarding labor management.