AI-Driven Prediction of Renal Stone Recurrence Following ECIRS: A Machine Learning Approach to Postoperative Risk...

Area of Science:

• Urology
• Nephrology
• Data Science

Background:

• Kidney stone recurrence prediction is complex due to multifactorial causes.
• Machine learning (ML) offers novel approaches to analyze high-dimensional clinical data for pattern identification.

Purpose of the Study:

• To apply ML models to identify key predictors of kidney stone recurrence after endoscopic combined intrarenal surgery (ECIRS) in patients with calcium stones.
• To evaluate the predictive performance of ML models and identify significant recurrence factors.

Main Methods:

• Retrospective analysis of 72 calcium stone patients achieving stone-free status post-ECIRS.
• Collection of 235 variables including clinical, urine, stone composition, and imaging data.
• Development and evaluation of ML models, with SHapley Additive exPlanations (SHAP) for feature importance.

Main Results:

• Recurrence observed in 40.3% of patients within two years.
• TabNet model achieved highest predictive accuracy (AUC=0.89).
• Independent predictors identified: urinary oxalate ≥ 25.4 mg/day and hemoglobin drop ≥ 0.3 g/dL at 3 months postoperatively.

Conclusions:

• ML accurately predicts kidney stone recurrence post-ECIRS, with 24-hour urine data significantly enhancing performance.
• Identified predictors, including "normal" oxalate levels, suggest current clinical thresholds may need re-evaluation.
• A simplified ML model using oxalate, urine volume, and hemoglobin drop shows clinical utility.