Predictive value of the stone-free rate after percutaneous nephrolithotomy based on multiple machine learning models

Area of Science:

• Urology
• Medical Informatics
• Machine Learning

Background:

• Percutaneous nephrolithotomy (PCNL) is a key procedure for treating kidney stones.
• Predicting the stone-free rate after PCNL is crucial for patient management and treatment planning.
• Machine learning (ML) offers potential for improving predictive accuracy in urological procedures.

Purpose of the Study:

• To develop and compare three ML models (GBDT, RF, XGBoost) for predicting the stone-free rate after PCNL.
• To evaluate the predictive performance and clinical utility of these ML models.
• To identify key clinical factors influencing PCNL outcomes using explainable AI methods.

Main Methods:

• Retrospective analysis of 160 patients undergoing PCNL, divided into training (70%) and testing (30%) sets.
• Development of GBDT, RF, and XGBoost models using clinical data.
• Evaluation of model performance using AUC, accuracy, sensitivity, specificity, F1 score, SHAP, and DCA.

Main Results:

• The overall stone-free rate was 70.6%.
• Significant predictors included stone number, diameter, CT value, prior surgery history, location, and shape.
• The GBDT model exhibited the best performance (AUC: 0.836, accuracy: 0.854), outperforming RF and XGBoost.
• SHAP analysis highlighted stone number and CT value as critical predictive features.

Conclusions:

• ML models, particularly GBDT, can accurately predict stone-free rates post-PCNL.
• The GBDT model aids in identifying patients likely to achieve successful PCNL outcomes.
• These models can assist clinicians in making informed treatment decisions for urinary tract stones.