Development of an interpretable machine learning model for predicting stress urinary incontinence following the delivery of a first child.

INTRODUCTION

The goal of this research is to use machine learning (ML) techniques to create a risk prediction model for postpartum stress urinary incontinence (PSUI). To improve screening accuracy and optimize clinical care techniques, the goal is to determine the best model for clinical screening.

METHODS

Telephone interviews and computerized medical records were used in this study to gather data, which was then combined into a follow-up data form. We recruited women who underwent surface electromyography (sEMG) at Xinjiang Production and Construction Corps Hospital's postpartum rehabilitation clinic between January 2020 and January 2025. A total of 449 primiparous women were included in the study after the inclusion and exclusion criteria were applied. A training set and a validation set were randomly selected from the dataset in a 7:3 ratio. Predictive models for persistent postpartum stress urinary incontinence (PSUI) were created using five machine learning algorithms: logistic regression, decision tree, random forest, support vector machine (SVM), and extreme gradient boosting (XGBoost). Model performance was evaluated using metrics including the area under the receiver operating characteristic curve (AUC). Model interpretability was achieved through SHapley Additive exPlanations (SHAP) analysis.

RESULTS

Eight features were kept in the final model: neonatal birth weight, pre-pregnancy BMI, age at delivery, perineal laceration, mean pre-resting phase, mean post-resting phase, maximum fast-twitch phase, and mean slow-twitch phase. Out of the five algorithms, the Random Forest model performed the best, with a test set AUC of 0.920 (95% CI 0.8704-0.9688, P < 0.05), accuracy of 0.8444, precision of 0.6341, recall of 0.8125, specificity of 0.8544, and an F1 score of 0.7123. Ten-fold cross-test and the test set were used to confirm the model's strong generalizability.

CONCLUSION

Strong clinical potential for PSUI risk prediction and screening is demonstrated by the random forest model. To further improve the model's clinical usefulness, future research should increase the sample size and include multicenter data.