A prediction model for late-onset urotoxicity following radiotherapy for cervical cancer based on bladder dosimetric parameters and clinical features: integrating machine learning interpretability and double machine learning analysis

Objective To establish a prediction model for late-onset urotoxicity after radiotherapy for cervical cancer based on bladder dosimetric parameters and clinical features, and to evaluate the predictive contribution and adjusted associations of dosimetric parameters using SHapley Additive exPlanations (SHAP) and double machine learning (DML). Methods This single-center retrospective study included 258 patients treated from June 2014 to December 2015. Late-onset urotoxicity was defined as urinary adverse reactions occurring ≥ 3 months after radiotherapy. Clinical data and bladder dose-volume histogram (DVH) parameters, including Bladder_V40, Bladder_V45, Bladder_V50, Bladder_Dmean, and Bladder_D2cc (expressed as EQD2), were collected. Variables were selected using univariate Logistic regression and Least Absolute Shrinkage and Selection Operator (LASSO) regression, followed by multivariate Logistic regression. Model performance was evaluated using receiver operating characteristic (ROC) curve, Bootstrap validation, calibration, Brier score, and decision curve analysis. Results Late-onset urotoxicity occurred in 65 patients (25.2%). Patients with urotoxicity had higher bladder dosimetric parameters. Multivariate Logistic regression identified Bladder_D2cc as an independent predictor [odds ratio= 1.61, 95% confidence interval (95% CI): 1.42–1.83, P < 0.001]. The model showed good discrimination, with an area under the ROC curve (AUC) of 0.943 and a Bootstrap mean AUC of 0.953 (95% CI: 0.928–0.974). Calibration and Brier score (0.083) indicated good agreement and clinical benefit. SHAP showed that Bladder_D2cc contributed most to prediction, while DML suggested positive adjusted associations between all bladder dosimetric parameters and late-onset urotoxicity, especially Bladder_V50. Conclusion This model effectively identified high-risk patients after cervical cancer radiotherapy. Bladder_D2cc was retained as a representative high-dose predictor and showed robust predictive performance, while Bladder_V50 yielded the largest standardized DML association estimate.