Predicting the occurrence of stress urinary incontinence after prolapse surgery: a machine learning-based model

BACKGROUND: Previous prediction models for postoperative stress urinary incontinence (SUI) cannot be applied to patients receiving transvaginal mesh (TVM) surgery and colpocleisis or those with preoperative subject urinary incontinence. This study aimed to develop and validate a new machine learning model and compare it to previous models. METHODS: Female patients who underwent prolapse surgeries for stage 2–4 anterior or apical prolapse between January 1, 2015, and December 31, 2019, at Peking Union Medical College Hospital were enrolled. Prolapse surgeries included native tissue repair, LeFort/colpocleisis, sacrocolpopexy, and TVM surgery. The existing models to predict postoperative SUI were externally validated. Subsequently, the dataset was randomly divided into 2 sets in a 4:1 ratio. The larger group was used to construct and internally validate models of logistic regression, random forest, and extreme gradient boosting (XGBoost), which were then externally validated. The discrimination of the prediction models was evaluated using the area under the curve, while the calibration of the models was measured using the Spiegelhalter z test, mean absolute error (MSE), and calibration curves. RESULTS: Overall, 555 patients were enrolled, and 116 experienced SUI 1 year postoperatively. Previous logistic models had poor performance, with areas under the curve of 0.544 and 0.586. In the model construction, the areas under the curve were 0.595, 0.842, and 0.714 for the logistic, random forest, and XGBoost models, respectively. However, only the XGBoost model exhibited good discrimination and calibration for both internal and external validations. Body mass index (BMI), C point of pelvic organ prolapse (POP) quantification stage, age, Aa point of POP quantification stage, and TVM surgery were the 5 most important predictors of postoperative SUI in the XGBoost model. CONCLUSIONS: Previous models had poor discrimination and calibration among a Chinese population. Hence, we developed and validated an XGBoost model, which performed well irrespective of the preoperative subjective urinary incontinence (preUI) and surgical methods. Further validation is still required.