Clinical Chemistry - Journal Club
The Clinical Chemistry Journal Club allows readers to discuss key articles by using focused slides as teaching tools. Each month Clinical Chemistry posts the original article and slides online, and they are then distributed to individuals and university Journal Clubs.
Original Article: https://doi.org/10.1093/clinchem/hvaf088
Slides: Download ppt
Webinar (on demand), available through November 30, 2026: https://myadlm.org/education/all-webinars/webinars/2025/november/machine-learning-algorithms-for-predicting-urinary-tract-infections
Urinary tract infections (UTIs) are among the most common infections encountered in healthcare settings. Current diagnostic practices often require 24–48 h due to the time needed for culture results. Given that 70%–80% of cultures return negative, there is significant interest in rapidly identifying negative samples to reduce unnecessary antibiotic use. This study aimed to develop and evaluate 6 machine learning models to predict UTIs.
Urine samples from 22 961 patients, collected between September 28, 2023 and June 29, 2024, were analyzed. Six machine learning models were assessed for their ability to predict UTIs based on 5 definitions incorporating pyuria and culture outcomes. The dataset was randomly divided into a training set (70%, n = 16 072) and an independent test set (30%, n = 6889). Seventeen predictive parameters, including dipstick reflectance results and demographic variables, were evaluated.
The CatBoost Classifier emerged as the best-performing model, achieving an area under the ROC curve of 92.0%–94.7% depending on the UTI definition, with a negative predictive value consistently exceeding 95%, and an average precision ranging from 68.2% to 81.6%. In comparison, the predictive performance of nitrite and/or leukocyte esterase was significantly lower.
Machine learning models, particularly the CatBoost Classifier, demonstrate high accuracy and offer a promising tool to aid clinicians in UTI diagnosis. Unlike traditional culture methods, these models deliver results within an hour. Further external validation with an independent dataset and prospective studies assessing the impact on antibiotic prescribing practices is recommended.