Machine learning-based clinical models for prediction of urinary incontinence after robot-assisted laparoscopic radical prostatectomy.
TL;DR
Total PSA level, nerve sparing procedure, IPSS score, length of membranous urethra, underlying disease, preoperative urinary tract infection, and urinary leakage are independent risk factors for early UI post-RARP.
Abstract
Objective
To investigate risk factors for urinary incontinence (UI) after robot-assisted laparoscopic radical prostatectomy (RARP) using interpretable machine learning methods, establish and validate a predictive model.
Methods
Clinicopathological data of 464 localized prostate cancer patients undergoing RARP at our institution (June 2022-June 2024) were retrospectively analyzed. UI status was assessed at 24-48 h, 1, 3, and 6 months post-catheter removal, with early UI defined as requiring ≥ 1 pad/day at 3 months. Patients were randomly split 7:3 into training/validation sets, patients from three external centers were used as the test set. Between-group comparisons used t-tests, Mann-Whitney U tests, or chi-square tests. Multivariate logistic regression identified risk factors in the training set, followed by construction of logistic regression and five machine learning models. Model performance was evaluated via ROC curves. The optimal model was interpreted using SHAP (Shapley additive explanations). Statistical significance was set at P < 0.05.
Results
Early UI occurred in 54 (training set), 28 (validation set) patients and 19 (test set) patients, with no intergroup baseline differences (P > 0.05). Univariate analysis linked total PSA level, PI-RADS score, ISUP grade, IPSS score, underlying disease, length of membranous urethra, prostate volume, nerve sparing procedure and urinary leakage preoperative total PSA, ISUP grade, dysuria, comorbidities, urinary tract infection (UTI), and postoperative urinary leakage to early UI (P < 0.05). Multivariate analysis confirmed total PSA level, nerve sparing procedure, IPSS score, length of membranous urethra, underlying disease, preoperative urinary tract infection, and urinary leakage as independent risk factors. Among models, support vector machine (SVM) achieved the highest validation AUC of 0.882 in the validation set and AUC of 0.858 in the test set. SHAP analysis revealed PSA level as the strongest predictive feature.
Conclusion
Total PSA level, nerve sparing procedure, IPSS score, length of membranous urethra, underlying disease, preoperative urinary tract infection, and urinary leakage are independent risk factors for early UI post-RARP. The SVM-based machine learning model demonstrates superior predictive performance and clinical utility.