LAUSR

PURPOSE The decision-making of how to treat urinary infection stones was complicated by the difficulty in preoperative diagnosis of these stones. Hence, we developed machine learning (ML) models that can be leveraged to discriminate between infection and non-infection stones in urolithiasis patients before treatment. METHODS We enrolled 462 patients with urinary stones and randomly stratified them into training (80%) and testing sets (20%). ML models were constructed using 5 algorithms (decision tree [DT], random forest classifier [RFC], extreme gradient boosting [XGBoost], categorical boosting [CatBoost], and adaptive boosting [AdaBoost]) and 15 preoperative variables and were compared with conventional logistic regression (LR) analysis. Performance measurement was the area under receiver operating characteristic curve (AUC) in the testing set. We also analyzed the importance of 15 features on the prediction of infection stones in each ML model. RESULTS Sixty-two (13.4%) patients with infection stones were included in the study. On the testing set, all the 5 ML models demonstrated strong discrimination (AUC: 0.892 - 0.951). The RFC model was chosen as the final model (AUC: 0.951 [95% CI, 0.934 - 0.968]; sensitivity: 0.906; specificity: 0.924), significantly outperforming the traditional LR model (AUC: 0.873 [95% CI, 0.843 - 0.904]). Gender, urine white blood cell (WBC) counts, and urine pH level were the top three important features. CONCLUSION Our RFC model was the first model for the preoperative identification of infection stones with superior predictive performance. This novel model could be useful for the risk assessment and decision support for infection stones.