Predicting Postoperative Outcomes in Pediatric Ureteroscopy Using Machine Learning and Explainable AI—EAU Endourology Vision-AI Study

Abstract

Aim of the Study:

This study aimed to evaluate the performance of machine learning (ML) algorithms integrated with explainable artificial intelligence (XAI) techniques in predicting outcomes following flexible ureteroscopy (fURSL) in children. By identifying significant preoperative predictors, the goal was to improve individualized surgical risk assessment and planning.

Materials and Methods:

A retrospective analysis was conducted on 391 pediatric patients treated with fURSL for urolithiasis across 8 high-volume centers (2017–2021). Preoperative, intraoperative, and postoperative data were collected. 15 ML models were trained to predict five postoperative outcomes: fever, hematuria, sepsis, residual fragments (RF), and reintervention. A multitask artificial neural network (ANN) was also developed. Performance was evaluated using validation accuracy, confusion matrices, and classification reports. SHapley Additive exPlanations values and decision trees were used for model interpretability.

Results:

Ensemble models outperformed others, with Gradient Boosting achieving 92.4% validation accuracy in predicting postoperative fever, Extra Trees achieving 91.1% for hematuria, and XGBoost reaching 96.0% for sepsis. Predictors included preoperative infections, stone burden, operative duration, and anatomical anomalies. For RF, Gradient Boost and Random Forest yielded strong results with up to 93.7% accuracy. Reintervention was best predicted by Random Forest, with RF as the strongest predictor. XAI techniques provided transparent, clinically interpretable models that aligned with medical reasoning.

Conclusion:

ML models demonstrated high accuracy in predicting adverse postoperative outcomes in pediatric ureteroscopy, with ensemble methods showing the best performance. Integration with XAI enhanced interpretability, supporting clinical decision-making. These findings underscore the potential of ML and XAI to inform personalized treatment strategies, though further prospective validation is needed to develop robust, generalizable predictive tools.