Comprehensive Summary
The purpose of this study was to develop a machine learning-based preoperative prediction model to predict urinary calculi composition in order to determine what targeted personalized surgical strategies should be employed in order to treat patients with urinary stone disease. This study drew data from 708 urolithiasis patients being treated for Second Affiliated Hospital of Zhengzhou University from January 2019 to November 2024. 9 different machine learning models were trained on data from 70% of patients based on 20 optimal predictive features identified by LASSO regression analysis, which were then evaluated using the test set(remaining 30% of patients). The study found that the Binary Logistic Regression(BLR) model achieved the highest overall predictive performance and diagnostic accuracy out of the machine learning models. The study also identified the strongest predictors for 4 different stone subtypes: calcium oxalate stones, infection stones, uric acid stones, and calcium-containing stones. While the prediction model developed by the study does not reach 100% diagnostic accuracy, it still provides a reliable method for predicting urinary stone composition, drawing data from only a few available clinical parameters.
Outcomes and Implications
This study of the binary classification framework demonstrated superior predictive accuracy and enhanced clinical utility compared to previous studies involving machine learning and urinary stones. One important aspect of this study is the use of the SHAP interpretability framework which was used to interpret the predictive outcomes and mitigate opacity of the models, which would usually cause similar studies to fall victim to the “black box effect”.Urolithiasis is a very common, costly, recurrent, and most importantly painful disease for patients. The treatment of this disease requires information about the stone composition for personalized surgical interventions as each stone subtype has a different pathology, but current preoperative techniques for predicting stone composition are not accurate enough. This study changes that; machine learning models can draw from many different sets of data including serological data and 24-hour metabolic urine tests in order to make accurate predictions of stone composition, optimizing the urolithiasis treatment process for both the patients and the hospitals.