From Preliminary Urinalysis to Decision Support: Machine Learning for UTI Prediction in Real-World Laboratory Data

Background/Objectives: Urinary tract infections (UTIs) are frequently diagnosed empirically, often leading to over-treatment and rising antimicrobial resistance. This study aimed to develop and evaluate machine learning (ML) models that predict urine culture outcomes using routine urinaly-sis and demographic data, supporting more targeted empirical antibiotic use. Methods: A real-world dataset comprising 8,065 urinalysis records from a hospital laboratory was used to train five ensemble ML models: Random Forest, XGBoost (eXtreme Gradient Boosting), Extra Trees, Voting Classifier, and Stacking Classifier. Models were developed using 10-fold stratified cross-validation and assessed via clinically relevant metrics in-cluding specificity, sensitivity, likelihood ratios, and diagnostic odds ratio (DOR). To en-hance screening utility, threshold optimization was applied to the best-performing model (XGBoost) using the Youden index. Results: XGBoost and Random Forest demonstrated the most balanced diagnostic profiles, with DORs exceeding 21. The Voting and Stacking Classifiers achieved highest specificity (>95%) and positive likelihood ratios (>10), but exhibited lower sensitivity. Feature im-portance analysis identified positive nitrites, white blood cell count, and specific gravity as key predictors. Threshold tuning of XGBoost improved sensitivity from 70.2% to 87.9% and reduced false negatives by 82%, with an associated NPV of 96.4%. The adjusted mod-el reduced overtreatment by 56% compared to empirical prescribing. Conclusions: ML models based on structured urinalysis and demographic data can support clinical decision-making for UTIs. While high-specificity models may reduce unnecessary antibi-otic use, sensitivity trade-offs must be considered. Threshold-optimized XGBoost offers a clinically adaptable tool for empirical treatment decisions, particularly in settings lacking rapid diagnostics.