Machine learning prediction of positive urine cultures in an academic pediatric emergency department (ED).

BACKGROUND

Urinary tract infections (UTIs) are among the most common pediatric infections, but urine culture, the diagnostic gold standard, requires 1-3 days for results. This delay may lead to unnecessary antibiotic use or treatment delays.

OBJECTIVE

To develop and validate a machine learning (ML) model that predicts urine culture positivity in pediatric emergency department (ED) encounters using electronic health record (EHR) data.

METHODS

We conducted a cross-sectional study of patients aged 3 months-18 years who had urine cultures obtained during ED visits at a quaternary academic medical center between January 2020 and December 2023. Demographics, arrival mode, chief complaint, labs, vitals, and encounter characteristics were extracted from the EHR. Logistic regression, K-nearest neighbors, neural networks, and gradient-boosted trees (XGBoost) were trained on 80% of encounters and tested on 20%. Model performance was assessed using accuracy, precision, recall, F1-score, and AUC-ROC. Feature importance was evaluated with Shapley Additive Explanations (SHAP).

RESULTS

Among 3788 encounters with urine cultures, 47.5% were positive. The XGBoost model achieved the best performance (accuracy 72.0%, precision 71.3%, recall 69.3%, AUC-ROC 78.1%). The model demonstrated appropriate probability calibration with a Brier score of 0.194. Key predictive features included ED diagnosis of UTI, urine collection method, sex, temperature, urinalysis findings, and select laboratory and vital sign measures.

CONCLUSION

A machine learning model applied to routinely collected EHR data accurately predicted urine culture positivity in pediatric ED patients. Such models may support clinician decision-making, reduce unnecessary antibiotic use, and minimize delays in UTI treatment.