Efficacy and safety of ureteroscopy in children with lower pole renal stones – a machine learning predictive model from the EAU section of Endourology

Introduction The rising incidence of kidney stone disease in children presents growing clinical challenges, particularly in managing lower pole (LP) calculi, which are anatomically difficult to treat. Flexible ureteroscopy with laser lithotripsy (fURSL) has emerged as a preferred minimally invasive treatment. However, surgical outcomes remain variable, especially in the paediatric LP stone cohort. This study aimed to apply machine learning (ML) techniques to predict surgical outcomes based on preoperative characteristics and identify key predictors of incomplete stone clearance. Materials and Methods A retrospective analysis was conducted on paediatric patients (< 18 years) who underwent fURSL between January 2017 and December 2021 across eight tertiary centres. From a multicentre database of 280 patients, 91 with isolated LP stones were selected. Preoperative, intraoperative, and postoperative variables were analysed. Fifteen ML models—including ensemble algorithms and a multitask neural network—were developed to predict LP stone presence and postoperative outcomes. Model performance was evaluated using accuracy, precision, recall, F1-score, and SHAP (SHapley Additive exPlanations) values for interpretability. Results LP stones were present in 32.5% of cases and were associated with older age, solitary stones, and higher stone burden. Random Forest outperformed all other models (validation accuracy: 80.95%; F1-score: 76.67%), followed by Gradient Boosting. SHAP analysis identified stone number, total stone burden, age, and operative time as top predictors. LP stones were associated with a higher rate of residual fragments (RF) and lower need for preoperative stenting or ureteral access sheath use. Infectious and bleeding complications were less frequent in the LP group. Conclusion fURSL is safe and effective in children with LP stones, though incomplete stone clearance remains a challenge. ML models demonstrated strong predictive performance and could support preoperative risk stratification. Further external validation and prospective studies are warranted to refine predictive tools for clinical use.