Using routine laboratory tests to perform early prediction of urine culture results

BackgroundUrinary tract infections (UTIs) are among the most common bacterial infections worldwide, typically diagnosed using a urine culture. However, urine cultures can take up to 72hrs to result, delaying and inhibiting treatment decisions. Here we aim to investigate whether machine learning models can enable earlier prediction of urine culture results. MethodsWe analyzed 30,369 urine cultures from 10,761 patients in the University of Washington Medicine Network. Random forest models were developed to predict overall culture positivity, as well as presence of each of the 10 most common infectious agents, using patient demographics and routine laboratory tests (blood counts, metabolic panels, etc.). Age- and sex-adjusted univariate associations of lab markers with culture positivity were also analyzed via logistic regression. ResultsML models predicted culture positivity with moderate accuracy (AUC: 0.76), and high precision (80-90%) at clinically relevant recall rates, with higher performance in cases where urinalysis was available (AUC: 0.82). Individual pathogen prediction was somewhat lower (AUCs: 0.64-0.72) though, particularly for less common pathogens. At least one marker from all common test panels showed a significant univariate association with culture results. ConclusionsML models using routine laboratory tests can predict overall urine culture positivity with clinically useful accuracy, though individual pathogen prediction remains challenging. Given limited urinalysis availability, these models may be best suited for triaging presumptively negative cultures to improve laboratory efficiency rather than directly informing antibiotic selection.