Machine learning-based prediction of stone-free status following extracorporeal shock wave lithotripsy

To develop a machine learning model for predicting stone-free (SF) outcomes following extracorporeal shock wave lithotripsy (SWL) and to identify key clinical and stone-related predictors using interpretable machine learning techniques.

A total of 369 patients who underwent SWL between 2013 and 2022 at a single institution were retrospectively analyzed. Clinical and stone-related features were used to train a Light Gradient Boosting Machine (LightGBM) classification model. The dataset was split into training (70%) and testing (30%) sets, and model performance was evaluated using accuracy, area under curve (AUC), recall, precision, and F1-score. SHapley Additive exPlanations (SHAP) values were used to assess feature importance. SMOTETomek was performed to address class imbalance.

The LightGBM model achieved an overall accuracy of 82.88% and an AUC of 0.651 in the testing set. However, recall and precision for the SF group were low (18.2% and 16.7%, respectively), reflecting model bias due to class imbalance. SHAP analysis identified mean stone size, ureter stone, and stone burden as the top predictors of SWL success. After class balancing, hypertension and age were ranked among the most influential features in the model. Logistic regression confirmed the significance of mean stone size and stone location, while SHAP revealed additional nonlinear feature contributions not captured by traditional methods.

The LightGBM model combined with SHAP interpretation effectively identified important predictors of SWL outcomes. Despite good overall performance, limited sensitivity for SF cases highlights the impact of class imbalance. Addressing this through balancing techniques and incorporating additional clinical variables may improve prediction and support personalized treatment planning in SWL.