Machine learning models to predict the zero-fragment rate and lower pole access with FANS during flexible Ureteroscopy-an EAU section of endourology study.

INTRODUCTION

Suction devices such as flexible and navigable suction ureteral access sheath (FANS) are promising tools to reach the zero-fragment rate (ZFR) after flexible ureteroscopy (FURS) and laser lithotripsy. FANS could especially be useful for lower pole stones (LPS), avoiding postoperative retained stone dust. Using Machine Learning (ML) models, we aimed to predict the ZFR after FURS with FANS and secondarily the possible access to the lower pole (LP).

METHODS

Data from patients who underwent FURS in 25 centers worldwide were prospectively collected (Aug 2023-Jan 2024). Exclusion criteria were abnormal renal anatomy and ureteral stones. ZFR and LP access were respectively defined as the total absence of residual fragments on computed tomography at 1-month follow-up and the ability to place the FANS into a LP calyx. After data normalization and splitting (training-test (80-20%)), Eight ML models were evaluated to predict separately ZFR and LPS, using binary classification.

RESULTS

A total of 390 patients were included with a median age of 49 (36-61)years. The median stone volume was 1440 (1006-2219)mm. 70,5% and 59% of patients were first time stone formers and pre-stented. The FANS could access the LP in 75,1% of cases. At 1-month, the ZFR was 56,7% (221/390). Above all ML models, ExtraTreesClassifier presented the highest accuracy (0,83), F1-score(0,85) and AUC (0,83) for ZFR prediction from preoperative data. The LP access was best predicted by the RandomForestClassifier (accuracy (0,86), F1-score (0,91) and AUC (0,76)).

CONCLUSION

Based on preoperative data and ML, we can accurately predict the ZFR and LP access during FURS with FANS. Our models could improve the elective indications for FANS utilization.