Development and Evaluation of a Machine Learning Recursive Partitioning Decision Tree Algorithm to Optimize Urinalysis Parameters to Predict Urine Culture Positivity

PittUDT, a recursive partitioning decision tree algorithm for predicting urine culture (UC) positivity based on macroscopic and microscopic urinalysis (UA) parameters, may significantly decrease the number of unnecessary UC. The reflex algorithm derivation included 38,361 paired UA and UC cases; the average patient age was 57.4 years and 70% of samples were from female patients. Receiver operating characteristic (ROC) analysis identified urine WBCs, leukocyte esterase, and bacteria as the best predictors of UC positivity, with area under the ROC curve of 0.79, 0.78, and 0.77, respectively. The test data comprised 9,773 cases, of which 2,571 (26.3%) had a positive UC result. Using the held-out test data set, the PittUDT algorithm met the pre-specified targets of a 30-60% total negative proportion (true negative plus false negative predictions) with a negative predictive value above 90%. These data show that a rule-based algorithm trained on UA data has adequate predictive ability for triaging urine specimens by identifying low risk urine specimens, which are unlikely to grow pathogenic organisms, with a false-negative proportion under 5%. Our study demonstrates the feasibility and performance of using a supervised machine learning approach to develop a human readable machine learning model to optimize UA parameters in a reflex protocol that can be deployed across multiple hospital sites and settings. Our work supports a UA reflex to UC protocol that has the capacity to optimize UA parameters to predict UC positivity and improve antimicrobial stewardship and UC utilization, a potential avenue for cost savings.