Machine Learning Based Prediction Models for Spontaneous Ureteral Stone Passage

Abstract

A prediction model was developed for spontaneous stone passage (SSP) using decision tree based machine learning (ML) methods which could serve as a promising tool for doctors to decide whether to administer active therapies and interventions to patients with ureteral stones, or wait for the SSP to happen. The work was performed on the medical records of 192 patients suffering from ureteral stones. A total of 20 features were taken into account including the target attribute, laboratory investigation data, radiographic results, stone characteristics, history of stones and interventions. On analysis, the dataset was found to be imbalanced with 65.1% cases of SSP occurred, and 34.9% SSP didn’t occur. The study included three data splitting approaches, namely hold-out method, 10-fold cross-validation, and stratified 10-fold cross-validation. Decision tree (DT), extreme gradient boosting trees (XGBoost), and random forest (RF) were employed as classification techniques in our work. The prediction accuracies on the test set using stratified 10-fold cross validation, 10-fold cross validation, and hold-out method respectively were 90.83%, 81.66%, and 75.00% for DT, 91.66%, 83.33%, and 83.33% for XGBoost, and 87.50%, 87.50%, and 83.33% for RF. The validation accuracies using stratified 10-fold cross validation, 10-fold cross validation, and hold-out method respectively were 85.55%, 81.66%, and 83.33% for DT, 88.89%, 83.33% and 83.33% for XGBoost, and 87.78%, 83.33%, and 77.78% for RF. It was observed that stratified k-fold cross-validation method gave the best results for our imbalanced dataset. The results are encouraging and suggest that the developed approach might be utilized to help urologists with counseling, prognosticating ureteral stone outcomes, and eventually selecting a suitable course of action or monitoring technique to deal with ureteral stones.