Development and validation of an explainable machine learning-based risk prediction model for macrosomia in han Chinese women with singleton cephalic pregnancies

To develop and internally validate an explainable machine learning model to predict macrosomia risk among Han Chinese women with singleton cephalic pregnancies during late gestation or at admission. Data from 3,233 women who delivered at a tertiary hospital from April 2024 to March 2025 were included (158 macrosomia cases). All candidate variables were derived from objective clinical data, including electronic medical records, admission measurements, and prenatal ultrasound examinations. To avoid data leakage, the dataset was initially partitioned using an 8:2 stratified split into training and held-out test sets. Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively within the training set and cross-validation training folds to address class imbalance. A complete analytical pipeline incorporating median imputation, appropriate feature scaling, and SMOTE was implemented. Multiple models were constructed and compared, including logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), CatBoost, k-nearest neighbors (KNN), and multi-layer perceptron (MLP). Mean stratified cross-validation ROC-AUC on the training set served as the primary criterion for model selection, supplemented by the Brier score, calibration curves, and decision curve analysis (DCA) to evaluate model reliability and clinical net benefit. LR achieved the highest cross-validated ROC-AUC on the training set, yielding an AUC of 0.934 (95% CI: 0.906–0.957), sensitivity of 0.833 (95% CI: 0.727–0.923), specificity of 0.876 (95% CI: 0.856–0.907), precision of 0.248 (95% CI: 0.184–0.322), and F1 score of 0.382 (95% CI: 0.295–0.458) on the held-out test set. Among the machine learning models, CatBoost exhibited relatively balanced performance regarding AUC, calibration, and clinical interpretability. Although XGBoost demonstrated the highest accuracy, its sensitivity was low, and it was thus not selected as the best model based solely on mean accuracy, precision, and recall. SHapley Additive exPlanations (SHAP) analysis indicated that fetal abdominal circumference (AC), fundal height, maternal height, gestational weeks at delivery, and maternal AC at admission contributed most significantly to prediction. This study established a reproducible and interpretable macrosomia risk prediction pipeline tailored for late-pregnancy/admission decision support. The model provides supplementary information for managing delivery and care in high-risk pregnancies; however, its widespread clinical application requires large-scale, multicenter, and prospective external validation.