How Well Can Preoperative Parameters Predict Postoperative Bleeding After Mini-PCNL? A Machine Learning Study

Objective: To develop and evaluate machine-learning models for predicting postoperative bleeding after mini-percutaneous nephrolithotomy (mini-PCNL) using preoperative clinical parameters only. Materials and Methods: We retrospectively analyzed 1515 patients who underwent mini-PCNL between January 2014 and January 2026. Postoperative bleeding and a postoperative day-1 hemoglobin decrease ≥2 g/dL occurred in 273 patients (18%). All intra- and postoperative variables were excluded to preserve a genuinely preoperative tool, leaving 15 preoperative predictors. Penalized logistic regression, random forest, Extreme Gradient Boosting (XGBoost), and light gradient boosting machine (LightGBM) were developed under a strictly patient-level scheme with nested cross-validation; all preprocessing and categorical-aware oversampling (SMOTENC) were fitted within training folds only. Discrimination (AUC: with 95% confidence intervals [CIs]), calibration (Brier score, calibration slope), and decision-curve analysis were assessed on an untouched hold-out set, with a sensitivity analysis excluding preoperative hematocrit (HCT). Results: In the patient-isolated hold-out set (303 patients, 52 events), random forest achieved the best discrimination (AUC: 0.856, 95% CI: 0.787–0.925), followed by XGBoost (0.838) and LightGBM (0.825); all three significantly outperformed penalized logistic regression (0.722; DeLong p ≤ 0.013). Random forest was well calibrated (slope 1.099, Brier 0.098) and showed positive net benefit across clinically relevant thresholds. Discrimination remained high after removing preoperative HCT (AUC: 0.828), indicating a multivariable signal. Preoperative HCT, caliceal access site, and stone size were the most influential predictors. Conclusion: Ensemble machine-learning models can predict postoperative bleeding after mini-PCNL with good, well-calibrated discrimination from preoperative variables alone, outperforming conventional logistic regression. External, temporal, and prospective validation are required before routine clinical use.