Leveraging explainable AI to predict surgical success and safety in ureteroscopy for Horseshoe Kidneys. An EAU section of endourology study

Abstract

Introduction: Flexible ureteroscopy (fURS) is a well-established approach for managing urolithiasis in patients with congenital renal anomalies like horseshoe kidney (HSK). However, these anatomical variations pose distinct challenges that can complicate both stone removal and procedural planning. This study seeks to utilize machine learning (ML) and explainable artificial intelligence (XAI) techniques to identify factors that predict stone-free status (SFS) after fURS in individuals with HSK. Material and methods: We retrospectively analysed 288 adult patients with HK who underwent fURS and laser lithotripsy for renal stones at tertiary referral centres. A ML model incorporating clinical and intraoperative variables was developed to predict SFS. SHAP (SHapley Additive exPlanations) values and decision-tree analysis were used to interpret feature importance and model behaviour. Results: Ensemble models, particularly CatBoost, XGBoost, and Random Forest, achieved the highest predictive performance for sepsis, reintervention, and surgical abandonment. SFS was harder to predict, with ensemble methods offering moderate accuracy. On-table RF and reintervention were consistently associated with poor outcomes, including lower SFS. Explainable AI methods such as decision trees and SHAP plots enhanced model transparency, identifying stone diameter, number of stones, and comorbidities as key predictors. Conclusions: ML ensemble models can successfully predict perioperative outcomes in HSK patients undergoing fURS. In this study, residual fragments and prior reintervention resulted as primary predictors of surgical success and safety. By integrating XAI into clinical workflows, risk stratification can be more accurate, allowing for personalised management strategies in patients with complex renal anatomy.