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

PURPOSE

This study aimed to develop three types of machine learning (ML) models based on gradient boosting decision tree (GBDT), random forest (RF), and extreme gradient boosting (XGBoost) to explore their predictive value for the stone-free rate after percutaneous nephrolithotomy (PCNL).

PATIENTS AND METHODS

A retrospective analysis was conducted on 160 patients who underwent PCNL. The patients were randomly divided into a training set and a test set in a 7:3 ratio. Clinical data were collected, and univariate analysis was performed to identify important data significantly associated with the stone-free rate after PCNL. Three ML models (GBDT, RF, and XGBoost) were developed using the training set. The predictive performance of these models was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) on the test set, confusion matrix, specificity, sensitivity, accuracy, and F1 score. For the top-performing prediction model, the study further employed the SHapley Additive exPlanations (SHAP) method to enhance model interpretability by elucidating the contribution of individual features to the prediction outcomes and ranking the relative importance of the predictive data. Finally, the clinical utility of the model was assessed through decision curve analysis (DCA), which quantified the net clinical benefit of applying the model across various risk thresholds.

RESULTS

Postoperative statistics indicated a stone-free rate of 70.6% ( = 113) among the patients. The data significantly associated with the absence of residual stones included the number of stones, stone diameter, stone CT value, history of previous stone surgery, stone location, and stone shape ( < 0.05). All three models demonstrated strong predictive effects in the validation set, with the GBDT model showing superior performance [AUC: 0.836 (95% CI: 0.785-0.873); accuracy: 0.854; sensitivity: 0.853; specificity: 0.857] compared to the XGBoost [AUC: 0.830 (95% CI: 0.792-0.868); accuracy: 0.771; sensitivity: 0.824; specificity: 0.643] and RF models [AUC: 0.803 (95% CI: 0.763-0.837); accuracy: 0.792; sensitivity: 0.824; specificity: 0.714]. The F1 scores for the GBDT, RF, and XGBoost models were 0.892, 0.836, and 0.849, respectively. The DCA decision curve analysis confirmed that the GBDT model offers a favorable net clinical benefit. In addition, the SHAP analysis identified the number of stones and the stone CT value as the most critical features influencing the model's predictions, contributing significantly to its overall predictive performance.

CONCLUSION

The prediction models developed based on three machine learning algorithms can accurately predict the stone-free rate after PCNL in patients with urinary tract stones. Among these, the GBDT model can effectively identify patients who are most likely to achieve successful outcomes from PCNL based on demographic and stone characteristics, thereby assisting in clinical treatment decision-making.