Antepartum prediction of shoulder dystocia using machine learning

Shoulder dystocia (SD) is a serious delivery complication, often occurring without identifiable risk factors. Accurate risk assessment is essential for educated decisions regarding the mode of delivery. This study aimed to develop and validate a machine learning-based model for SD prediction.

We conducted a retrospective analysis of term singleton vaginal deliveries at a single academic hospital over 10 years. Exclusion criteria included cesarean deliveries and multiple gestations. Maternal and fetal characteristics were compared between SD and non-SD cases. Eleven features were selected for model development. Missing values in normally distributed variables were imputed using mean values. To address class imbalance, repeated random sampling of non-SD cases was performed. Data were split into training and test sets (70:30) and standardized. Multiple machine learning models-including logistic regression, decision tree, random forest, support vector machine, XGBoost, and CatBoost-were evaluated using cross-validation and area under the ROC curve (AUC).

Among 51,628 deliveries, 94 (0.18%) involved SD. SD was associated with higher BMI, shorter stature, and increased diabetes rates (all p < 0.05). Mean birthweight was significantly higher in the SD group (3751 g vs 3287 g, p < 0.01). The CatBoost model achieved the highest performance (AUC = 0.83, 95% CI 0.77-0.89). Key predictive features were sonographic estimated fetal weight (EFW, 55.6% of the model's decision-making weight), maternal BMI (20.1%), and clinical EFW (7.9%).

Our machine learning-based model predicted SD with an AUC of 0.83 and may support clinicians in delivery planning.