Machine Learning Algorithms for Predicting Urinary Tract Infections: Integration of Demographic Data and Dipstick...

Area of Science:

• Urology
• Infectious Diseases
• Medical Informatics

Background:

• Urinary tract infections (UTIs) are common, with traditional culture diagnostics taking 24-48 hours.
• A high percentage (70-80%) of urine cultures are negative, leading to potential overuse of antibiotics.
• Rapid identification of negative UTI cases is crucial for optimizing antibiotic stewardship.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for the rapid prediction of UTIs.
• To compare the performance of ML models against traditional dipstick urinalysis parameters.

Main Methods:

• Analysis of urine samples from 22,961 patients.
• Development and assessment of six ML models using 17 predictive parameters, including dipstick results and demographics.
• Data split into 70% training and 30% independent testing sets.

Main Results:

• The CatBoost Classifier achieved the highest performance, with an area under the ROC curve of 92.0%-94.7%.
• The model demonstrated a negative predictive value consistently above 95% and average precision from 68.2% to 81.6%.
• Traditional dipstick parameters (nitrite and leukocyte esterase) showed significantly lower predictive performance.

Conclusions:

• ML models, especially the CatBoost Classifier, offer a highly accurate and rapid tool for UTI diagnosis, delivering results in under an hour.
• These models can assist clinicians in making timely diagnostic decisions.
• Further validation and studies on antibiotic prescribing impact are recommended.