Machine learning-based prediction of Clavien-Dindo ≥ II complications after ureteroscopy

Clavien-Dindo (CD) grade II complications after ureteroscopy occur in approximately 13-17% of cases, typically involving infections or bleeding requiring medical intervention. Thus, the purpose of this study was to create and internally validate a lean prediction model for CD ≥ II complications after ureteroscopy using routinely available perioperative predictors.

Retrospective study of ureteroscopic lithotripsy patients (2010-2014) evaluated outcomes and risk factors. Covariates included age, stone size, operative time, status of impacted stones, and the presence of a percutaneous nephrostomy (PN). The performance of the discriminative models was evaluated using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals. Overall model performance was further assessed in terms of discrimination, calibration, and accuracy, while clinical utility was evaluated using decision curve analysis. Model explainability was addressed through variable importance measures and complementary interpretability techniques.

A total of 389 patients were included, of whom 8.5% experienced clinically significant complications. The binomial generalized linear model (GLM) with Elastic Net regularization and the gradient boosting model (GBM) were developed using an 80:20 train-test split. In the testing dataset, both models demonstrated excellent discrimination (GLM AUC 0.893 vs. GBM AUC 0.857; p = 0.672). Calibration demonstrated strong performance, and decision curve analysis confirmed that both models outperformed treat-all and treat-none strategies, with GBM showing a modest advantage. Operative time and impacted stones were key predictors.

By integrating routinely available perioperative variables, the model enables individualized risk stratification and may support preoperative counseling and clinical decision-making.