Prediction model for additional procedure requirement in flexible ureterorenoscopy using explainable artificial intelligence

Comprehensive Summary

This study developed machine learning models to predict which patients will need an additional procedure after flexible ureterorenoscopy (f-URS). The analysis included 656 patients treated between 2015 and 2025, with 180 patients (27.4%) requiring further treatment such as repeat f-URS, ureteroscopy, shock wave lithotripsy, or percutaneous nephrolithotomy. The authors tested 14 machine learning models and used feature selection methods including Boruta, LASSO, and ElasticNet. They also used explainability tools such as SHAP and LIME, to show which variables mattered most. The strongest predictor was the ureteropelvic junction–pelvis angle (UPJ-PA). Patients below 110° had a much higher intervention rate than those above it (84.3% vs 2.8%). Explainability analyses consistently ranked UPJ-PA as the most important feature, followed by FANS-UAS use and access sheath size. Robustness testing also showed that the model stayed stable even when noise was added to UPJ-PA measurements.

Outcomes and Implications

The results suggest that machine learning can help identify patients who are more likely to fail initial f-URS and need another procedure. A UPJ-PA below 110° may signal difficult scope movement and higher risk of residual stones. This could help surgeons counsel patients more clearly, plan equipment, or consider alternative approaches in high-risk cases. The use of explainable AI also makes the model more useful because it shows why the prediction is being made instead of giving a black-box output. However, the study has important limits because it was retrospective, single-center, and lacked external validation. The 110° cutoff was also data-driven rather than physiologically proven, so it should not be treated as a universal threshold yet. Overall, the findings are clinically interesting, but they need prospective and multicenter validation before routine use.