Machine learning for predicting stone-free rate after retrograde intrarenal surgery: Comparison with the T.O.HO....

Urolithiasis remains a significant clinical burden, and accurately predicting stone-free status after retrograde intrarenal surgery (RIRS) is essential for optimizing patient counseling and surgical planning. This study compared the predictive performance of a machine learning-based Random Forest (RF) model with the T.O.HO. scoring system, a traditional prediction tool designed for flexible ureteroscopy. A total of 452 patients who underwent RIRS for renal or proximal ureteral calculi at a single tertiary center were retrospectively analyzed. Preoperative demographic and radiological variables were collected, and postoperative stone-free status was assessed at the third week using KUB radiography or non-contrast CT when clinically indicated. The T.O.HO. score was calculated for each patient. The RF model was trained on 70% of the dataset and tested on the remaining 30%, using only preoperative variables, with model optimization performed within the training set using 5-fold cross-validation, and predictive performance evaluated through accuracy, sensitivity, specificity, F1-score, and receiver operating characteristic (ROC) analysis. The T.O.HO. score showed limited but statistically significant discrimination in this cohort. Higher T.O.HO. values were associated with a lower likelihood of postoperative stone-free status, indicating an inverse relationship between the score and surgical success. When interpreted accordingly for stone-free status prediction, the T.O.HO. score yielded an AUC of 0.615 (95% CI 0.556-0.674). In contrast, the RF model showed higher apparent internally validated discrimination, achieving an AUC of 0.843, an accuracy of 87.5%, a sensitivity of 98.1%, and a specificity of 53.1%. Feature importance analysis indicated that stone thickness, preoperative nephrostomy, stone length, stone width, and age contributed most significantly to the model's predictive ability. Overall, these findings suggest that a machine learning-based approach may provide complementary individualized risk stratification for postoperative stone-free status after RIRS when used alongside established clinical judgment and existing scoring systems. However, this model was developed and evaluated using internal validation only, and external prospective validation is required before its clinical implementation can be considered.