Application of Machine-Learning Models in Predicting Transfusion Among Complex Renal Stones Patients Receiving Percutaneous Nephrolithotomy：A Retrospective Study

Abstract To predict and compare machine-learning models with logistic regression (LR) for transfusion requirement among complex renal stones patients receiving PCNL. The data of 513 patients with complex renal stones receiving PCNL between January 2016 and December 2021 from 2 centers were analyzed retrospectively and reviewed the pre-, intra-, and post-operation factors. We used 80% training set and 20% test set for the predictions of transfusion by LR and machine-learning algorithms, random forest (RF), support vector machine (SVM), and extreme gradient boosting trees (XGBoost), and compared the prediction accuracies, precisions, and area under curve (AUC) values of receiving operating characteristic (ROC) curves in each model. In our study, the prediction accuracies of RF (Accuracy = 0.974), SVM (Accuracy = 0.968) and XGBoost (Accuracy = 0.968) models all outperformed LR (Accuracy = 0.948), and the precision of RF (Precision = 0.818) was higher than XGBoost (Precision = 0.8), SVM (Precision = 0.8) and LR (Precision = 0.8). Moreover, RF (AUC = 0.962), SVM (AUC = 0.786) and XGBoost (AUC = 0.867) offered higher AUC values than LR (AUC = 0.77). The RF confirmed size of tract, number of tracts, and puncture correctness as significant risk factors, which was the same as LR model, while the size of tract was the main factor in XGBoost. All machine-learning algorithms yielded better predictions than LR for transfusion among PCNL patients. The importance of machine learning algorithm can give matched insights to domain knowledge on effective and influential factors for transfusion requirement.