Machine learning model integrating radiomics and clinical features for predicting postoperative bleeding after percutaneous nephrolithotomy

Comprehensive Summary

This study developed a machine learning model that combines radiomic features from CT scans with clinical variables to predict the risk of postoperative bleeding in patients undergoing percutaneous nephrolithotomy (PCNL). The retrospective analysis included 151 patients, with 31% experiencing postoperative bleeding. Radiomic features were selected using Lasso regression, and clinical variables such as stone number, diameter, shape, location, operation time, surgical channel, and staged surgery were identified through univariate analysis. Three machine learning models were tested: logistic regression, random forest, and support vector machine, with logistic regression and random forest performing best (accuracy 75.6%, AUC 0.760 and 0.740, respectively). Key predictors included stone shape, operation time, stone diameter, and specific radiomic texture features. Decision curve analysis and SHAP interpretation confirmed the model’s clinical usefulness and explained individual variable contributions.

Outcomes and Implications

The study shows that integrating radiomics and clinical data can improve preoperative risk stratification and guide personalized postoperative monitoring, although limitations such as single-center design, small sample size, and lack of external validation highlight the need for future multicenter studies.