Machine Learning-Based Prediction of Stone-Free Rate After Retrograde Intrarenal Surgery for Lower Pole Renal Stones

Background: Lower pole renal stones (LPS) present unique challenges for retrograde intrarenal surgery (RIRS) due to unfavorable anatomical features, often resulting in suboptimal stone-free rates (SFR). Recent advancements in machine learning (ML) offer new opportunities to predict surgical outcomes and guide clinical decision-making. This study aimed to develop and validate ML-based models to predict SFR following RIRS for LPS. Materials and Methods: We retrospectively analyzed data from 327 patients with LPS who underwent RIRS at two academic institutions: Kaohsiung Medical University Hospital (KMUH, n = 193) and Seoul National University Hospital (SNUH, n = 134). Demographic, anatomical, and stone-related variables were collected, including stone burden, Hounsfield unit (HU), pelvic stone angle (PSA), and renal infundibular length (RIL). A Light Gradient Boosting Machine (LightGBM) algorithm was developed using KMUH data and externally validated with SNUH data. SHAP (SHapley Additive exPlanations) analysis was performed to interpret feature importance. Results: The LightGBM model achieved the highest predictive performance. External validation using the SNUH dataset yielded an accuracy of 77.1%, AUC of 0.759, and F1-score of 0.853. SHAP analysis revealed that stone burden, HU, PSA, and RIL were the most influential features. Notably, PSA demonstrated strong predictive relevance, supporting its use as an alternative to the traditional infundibulopelvic angle (IPA) in anatomical assessment. Conclusions: ML-based models, particularly LightGBM, offer robust predictive capability for SFR following RIRS in patients with LPS. These tools may enhance preoperative planning and personalized surgical strategies. Future prospective studies are warranted to further validate their clinical utility and expand on feature integration.