Development and external validation of machine learning models for predicting hysterectomy in patients with postpartum hemorrhage

Abstract Postpartum hemorrhage (PPH) is the leading cause of preventable maternal mortality worldwide. Emergency postpartum hysterectomy represents the ultimate life-saving intervention for refractory PPH but results in irreversible infertility and severe long-term maternal complications. Current clinical risk assessment relies primarily on empirical judgment and conventional scoring tools, lacking precise, interpretable, and externally validated models for predicting PPH-related hysterectomy. This study aimed to develop and validate a robust machine learning (ML) model with explainable artificial intelligence (XAI) for individualized preoperative risk stratification of postpartum hysterectomy among PPH patients. This retrospective observational study was conducted based on the Medical Information Mart for Intensive Care IV (MIMIC-IV v3.1) database. A total of 1669 eligible PPH patients were enrolled and randomly divided into a training cohort (n = 1169) and an internal validation cohort (n = 500) at a 7:3 stratified ratio. Four rigorous algorithms (LASSO, Boruta, random forest, and XGBoost) were integrated to screen core predictive variables. Eight ML models were constructed and optimized using 10-fold cross-validation. Model performance was comprehensively evaluated via ROC curves, calibration curves, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) analysis was performed to enhance model interpretability. Independent external validation was conducted using 488 clinical PPH patients from the Department of Obstetrics and Gynecology, Chinese PLA General Hospital. The optimal model was further deployed as a web-based bedside prediction tool. Six core predictive features were finally identified, including maternal age, platelet count, magnesium sulfate administration, placental abnormality, uterine leiomyoma, and antihypertensive medication use. Logistic regression (LR) yielded the optimal comprehensive performance and was selected as the final model. In internal validation, the LR model achieved an AUC of 0.917 and an accuracy of 0.962 with excellent calibration and favorable clinical net benefit. In external validation, the model maintained outstanding generalization capability, with an AUC of 0.9656, high specificity (0.9938), and satisfactory sensitivity (0.8). SHAP analysis demonstrated that placental abnormality constituted the dominant risk factor, whereas elevated platelet count served as a protective indicator. A visualized online prediction tool was successfully established to support real-time bedside risk assessment. This study developed and externally validated an interpretable machine learning model for predicting hysterectomy risk in PPH patients. The model exhibits high discriminative accuracy, robust calibration, and favorable clinical applicability across both internal and external validation settings. The model may support early risk stratification and clinical preparedness, and the web-based tool offers a practical means of delivering individualized risk estimates at the bedside. Prospective impact studies are required to determine whether use of this model can improve clinical decision-making or reduce unnecessary emergency hysterectomies and improve maternal outcomes in practice.