Abstract / Summary
Predicting treatment success after shock wave lithotripsy (SWL) and retrograde intrarenal surgery (RIRS) remains challenging. We aim to develop and validate treatment-specific machine learning models for predicting short-term stone-free outcomes following shock wave lithotripsy (SWL) and retrograde intrarenal surgery (RIRS) using routinely available non-contrast CT (NCCT) parameters. A two-stage study was conducted. In the retrospective phase, 324 adult patients treated with SWL or RIRS for upper urinary tract stones were analyzed. Demographic variables and NCCT-derived stone characteristics, including size, volume, density, and heterogeneity, were used to train supervised machine learning models. Fifteen algorithms were evaluated using five-fold cross-validation, and the best-performing model for each treatment modality was selected. In the prospective phase, model performance was tested in an independent cohort of 40 patients (20 SWL, 20 RIRS). Shapley Additive Explanations (SHAP) analysis was applied to provide interpretability by quantifying the direction and relative contribution of individual CT-derived features to treatment-specific predictions. The CatBoost model demonstrated the highest predictive performance for SWL outcomes (accuracy 0.85, AUC 0.89), while Random Forest performed best in the RIRS cohort (accuracy 0.84, AUC 0.90). Mean stone density (Hounsfield units) and stone geometry (long edge for SWL, vertical edge for RIRS) emerged as the most influential predictors. In the prospective cohort, all SWL cases and 95% of RIRS cases were correctly classified. Patients undergoing RIRS exhibited larger stone size, greater volume, and a higher prevalence of challenging stone locations, reflecting confounding by indication in treatment allocation. Treatment-specific machine learning models using routinely available NCCT parameters promising predictive performance in estimating post-procedural stone-free outcomes after SWL and RIRS, though results require validation in larger prospective cohorts before clinical implementation. The integration of SHAP-based interpretability enhances clinical applicability, supporting their role as adjunctive tools for individualized patient counseling and expectation management rather than treatment selection.