From conventional scores to explainable AI: a six-method comparative framework for failure prediction in percutaneous...

Area of Science:

• Urology
• Medical Informatics
• Artificial Intelligence

Background:

• Percutaneous nephrolithotomy (PCNL) is standard for large kidney stones.
• Existing predictive models have limitations due to linear assumptions.
• Accurate prediction of PCNL failure is crucial for patient outcomes.

Purpose of the Study:

• Develop and validate a machine learning (ML) model with explainable AI (XAI) to predict PCNL failure.
• Compare the ML/XAI model's performance against traditional logistic regression and scoring systems.
• Improve pre-operative risk stratification for PCNL.

Main Methods:

• Retrospective analysis of 287 PCNL patients.
• Utilized demographic, laboratory, and imaging data.
• Employed multiple feature selection techniques and ML algorithms (Voting Classifier).
• Applied SHAP and LIME for model interpretability.
• Compared performance with logistic regression, GSS, and CROES scoring systems.

Main Results:

• The ML Voting Classifier achieved superior performance (AUC=0.839, accuracy=84.5%) compared to logistic regression (AUC=0.812) and scoring systems (AUC=0.615-0.653).
• Key predictors identified by SHAP analysis include stone-to-skin distance and stone length.
• Logistic regression identified multiple stones as a risk factor and stone-to-skin distance as protective.

Conclusions:

• The developed ML/XAI model offers higher accuracy and interpretability for predicting PCNL failure.
• This clinically applicable model enhances pre-operative risk stratification in urology.
• Further validation in multicenter prospective studies is recommended for generalizability.