(299) Predicting Success of External Cephalic Version: Machine Learning vs. Traditional Models in a Large Cohort

To develop and compare predictive models for successful external cephalic version (ECV) using traditional statistical methods and modern machine learning (ML) approaches on a large dataset, aiming to improve clinical decision-making through interpretable artificial intelligence.

Study Design:

We conducted a retrospective analysis of all ECV procedures performed at a single tertiary center between 2011-2024. The cohort included 1454 cases - one of the largest datasets in this domain - of which 834 (57.3%) were successful. Four clinically relevant predictors were selected: parity, body mass index (BMI), placental location, and amniotic fluid index (AFI). Two predictive models were developed: logistic regression and a tree-based eXtreme Gradient Boosting (XGBoost) model. The data were randomly divided into training and test sets (80:20). The model performance was assessed using area under the curve (AUC), accuracy, and precision-recall metrics. SHapley Additive exPlanations (SHAP) were used to evaluate interpretability and to visualize the importance of each feature to individual predictions.

Results:

The XGBoost model demonstrated superior performance compared to logistic regression. In the test cohort, AUC was 0.72 for XGBoost and 0.69 for logistic regression. It also achieved an accuracy of 0.66 vs. 0.61 and a precision of 0.73 for both models. SHAP analysis revealed that lower BMI and multiparity were associated with a higher likelihood of ECV success. Posterior placenta and higher AFI also contributed favorably to predictions. These findings were consistent with clinical expectations and improved model transparency.

Conclusion:

Machine learning models, particularly XGBoost, offer improved performance over traditional tools in predicting external cephalic version success. SHAP improved model interpretability and clinicians' understanding and trust. This study, based on one of the largest published ECV cohorts, highlights the value of ML in advancing prediction and personalization in obstetric care.