Predicting the stone-free status post percutaneous nephrolithotomy: a meta-analysis of machine learning models.

Percutaneous nephrolithotomy (PCNL) is the primary treatment for large or complex renal stones, yet postoperative stone-free status (SFS) remains difficult to predict, contributing to variation in clinical outcomes. Machine learning (ML) models have recently been applied in urology and show promise for predicting treatment success by integrating clinical and radiologic features. This study aims to evaluate the diagnostic performance of ML models in predicting the stone-free status following PCNL. We followed PRISMA-DTA guidelines throughout the entire process. A systematic search of PubMed, Scopus, Web of Science, and Embase was conducted from inception to January 2025. Studies assessing the sensitivity and specificity of ML models for predicting SFS after PCNL were screened and retrieved according to predefined criteria. Primary outcomes included pooled sensitivity, specificity, positive and negative likelihood ratios, and diagnostic odds ratio. All analyses were performed using MetaDisc software, applying a bivariate random-effects model and quality assessment was performed using QUADAS-2 tool. Six studies including 17,977 patients were analyzed. ML models showed a pooled sensitivity of 0.714 (95% CI 0.062–0.99) and specificity of 0.767 (95% CI 0.175–0.981). The diagnostic odds ratio was 8.218, indicating moderate accuracy with high heterogeneity. Subgroup analyses of XGBoost, logistic regression, and random forest models showed similarly variable performance. Our results indicate that current ML models demonstrate moderate ability to predict SFS after PCNL but lack consistency for independent clinical use. Stronger validation and standardized methodology are needed before these tools can be reliably implemented.