# Forecasting Stone-Free Status Following Percutaneous Nephrolithotomy Utilizing Explainable Machine Learning

## TL;DR

This study uses explainable machine learning to predict whether patients will be stone-free after a kidney stone surgery called PNL.

## Contribution

The study introduces an explainable machine learning model, particularly XGBoost, for predicting post-PNL stone-free outcomes with clinical transparency.

## Key findings

- XGBoost achieved the highest accuracy (0.916) and ROC–AUC (0.975) among tested models.
- SHAP analysis revealed anatomical anomalies as the strongest predictor of stone-free outcomes.
- Model features like stone placement and operation duration showed clinical relevance.

## Abstract

Background: This study aimed to create and evaluate explainable machine learning models for forecasting postoperative stone-free status following percutaneous nephrolithotomy (PNL) utilizing a substantial clinical cohort. Methods: This retrospective single-center analysis encompassed 2144 adult patients who received PNL from 2010 to 2024. We employed clinical, radiographic, stone-related, and surgical data to train four supervised machine learning models: Extreme Gradient Boosting (XGBoost), Random Forest, Light Gradient Boosting Machine (LightGBM), and Adaptive Boosting (AdaBoost). We used the Synthetic Minority Oversampling Technique exclusively on the training set to fix the class imbalance. We assessed the model’s accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (ROC–AUC) to see how well it worked. SHapley Additive exPlanations (SHAP) were used to measure explainability. Results: The total stone-free rate was 84.8%. XGBoost had the best predictive performance of the models tested, with an accuracy of 0.916 and a ROC–AUC of 0.975. LightGBM was close behind. Random Forest and AdaBoost had relatively inferior performance. SHAP analysis identified anatomical anomalies as demonstrated the strongest association with stone-free outcomes. The size of the access sheath and the number of stones were next. Other parameters that were identified by SHAP as important contributors to model predictions were the placement of the stone, Guy’s Stone Score, the length of the operation, and the density of the stone. These feature associations demonstrated clinical coherence with established knowledge in surgical practice. Conclusions: Explainable machine learning algorithms, especially XGBoost, can accurately predict stone-free outcomes following PNL in a way that makes sense to doctors. The incorporation of SHAP improves transparency and facilitates the prospective application of these models as decision-support instruments in personalized surgical planning.