Development of an Interpretable Machine Learning Model to Predict Pregnancy Outcomes Following Cervical Cerclage

Cervical insufficiency is a major cause of spontaneous preterm birth. Although McDonald cerclage improves outcomes, adverse events remain frequent. Accurate prediction of post-cerclage outcomes is essential for individualized management. Machine learning (ML) may enhance risk stratification, but clinical evidence remains limited.

We retrospectively analyzed 462 pregnant women who underwent McDonald cerclage at the Department of Obstetrics and Gynecology, General Hospital of Northern Theater Command, from June 2021 to June 2024. Clinical, obstetric, and laboratory parameters were incorporated into multiple ML models, including logistic regression, random forest (RF),support vector machines (SVM), decision trees (DT), and extreme gradient boosting (XGBoost). Model performance was evaluated using discrimination, calibration, and clinical utility, with SHAP analysis applied to interpret predictor contributions.

Logistic regression achieved the highest discrimination (AUC = 0.796), while XGBoost provided the best precision-recall balance (F1 = 0.712). RF demonstrated the most balanced performance, combining robust accuracy, interpretability, and reliability. SHAP analysis identified elevated C-reactive protein, increased white blood cell count, and amniotic fluid sludge as the strongest predictors. Conception method, maternal weight, and cerclage subtype also contributed to risk.

The RF model provided a clinically useful and interpretable framework for predicting outcomes after cerclage, emphasizing inflammatory status, maternal characteristics, and cerclage indication as key determinants of preterm birth. An online prediction tool was developed to facilitate individualized risk assessment. Despite the retrospective, single-center design and lack of external validation, these findings support the integration of ML into clinical decision-making, and warrant multicenter prospective validation.