Development of a prediction model for postoperative urinary tract infection in ureteral stone patients based on automated machine learning models.

Objective: This study aimed to develop a predictive model for postoperative urinary tract infection (PO-UTI) in ureteral stone patients, addressing limitations of traditional research methods and advancing perioperative infection management from experience-driven to data-driven transformation. Methods: A retrospective cohort design was employed, enrolling 826 ureteral stone surgery patients (January 2020 to January 2024) with data on demographics, disease characteristics, and hematological indicators collected via structured electronic medical records. Feature selection was optimized using an improved ISequoiaOA meta-heuristic algorithm to enhance model optimization stability; the SMOTE-ENN hybrid sampling technique was applied to balance class distribution; an AutoML framework integrating SHAP interpretability analysis was constructed to quantify feature contribution and visualize interaction effects; and a clinical decision support system was developed. Results: (1) optimal performance of the AutoML model on the test set (ROC-AUC=0.9251, PR-AUC=0.8712), significantly outperforming traditional algorithms such as XGBoost and LightGBM; (2) key predictors identified via SHAP analysis included preoperative urinary retention, low serum albumin (ALB), diabetes, double-J stent indwelling time, postoperative catheter indwelling time, and age, with interaction effects revealing a nonlinear synergistic surge in infection risk when stone size exceeded 6 mm and catheter indwelling time exceeded 5 days. Conclusion: This study integrated AutoML and explainable AI technologies to construct an accurate PO-UTI prediction model, groundbreakingly quantifying biological thresholds for synergistic effects and providing a "data-mechanism dual-driven" new paradigm for perioperative infection prevention and control.