"Preoperative Prediction of Uterine Scar Tenderness Using Machine Learning: A Decision-Support Tool for Cesarean Delivery Planning"

Abstract

When planning a repeat caesarean section (CS), evaluation of the integrity of the uterine scar is crucial. Conventional methods, like lower uterine segment (LUS) ultrasonography, are frequently operator-dependent and restricted to late gestation. This study aimed to develop and validate a machine learning (ML) model using routinely collected clinical data for preoperative prediction of intraoperative uterine scar tenderness—a surrogate marker of scar integrity.

A retrospective analysis was conducted on 353 women undergoing repeat CS. Intraoperative scar tenderness was graded on a 3-point scale: Grade 0 (none), Grade 1 (mild/moderate), and Grade 2 (severe). An XGBoost classifier was trained using preoperative maternal, fetal, and obstetric features. Data preprocessing included imputation, encoding, and feature selection. Model performance was evaluated using accuracy, macro-average F1-score, and area under the ROC curve (AUC). SHAP values were used to assess feature importance, and Youden’s Index optimized probability thresholds.

The model achieved an overall accuracy of 86.4%, a macro-average F1-score of 0.83, and AUCs of 0.93, 0.87, and 0.91 for Grades 0, 1, and 2, respectively. For Grade 2, a probability threshold of ≥0.65 yielded a positive predictive value (PPV) of 87.5% and sensitivity of 82%. Key predictors included number of previous CS, short inter-delivery interval, emergency category, high BMI, and placenta previa. Pilot clinical integration demonstrated high usability and clinician satisfaction.

By predicting uterine scar tenderness early and non-invasively, this ML model supports individualized surgical planning and aligns with Saudi Vision 2030’s digital transformation goals in maternal healthcare.