Machine learning to predict bacteriuria in the emergency department.

Area of Science:

• Medical Informatics
• Computational Biology
• Infectious Diseases

Background:

• Urinary tract infections (UTIs) are common but often misdiagnosed and mistreated.
• Accurate and timely diagnosis of UTIs is crucial for effective patient management.

Purpose of the Study:

• To evaluate machine learning models for predicting bacteriuria using readily available emergency department (ED) data.
• To assess the predictive performance of various machine learning algorithms for different thresholds of bacterial growth in urine cultures.

Main Methods:

• Retrospective analysis of 62,963 ED encounters with urinalysis and urine culture data (2017-2021).
• Comparison of logistic regression, k-nearest neighbors, random forest, extreme gradient boosting (XGBoost), and deep neural networks.
• Prediction of three urine culture outcomes: any microbial growth, ≥10,000 CFU/mL, and ≥100,000 CFU/mL.

Main Results:

• XGBoost demonstrated the highest predictive accuracy with AUROCs of 86.1%, 89.1%, and 93.1% for the respective outcomes.
• For cases diagnosed with UTI pre-culture, XGBoost achieved an AUROC of 91% in predicting no growth or ≥100,000 CFU/mL.
• The model accurately predicted bacteriuria using only data available during the ED encounter.

Conclusions:

• Machine learning, particularly XGBoost, can accurately predict bacteriuria using routinely collected ED data.
• These algorithms offer a valuable tool for clinicians to predict culture results and inform empiric antibiotic treatment decisions.
• Integrating machine learning into clinical workflows can improve the management of suspected UTIs.