Development, Evaluation, and Multisite Deployment of a Machine Learning Decision Tree Algorithm To Optimize...

Area of Science:

• Clinical diagnostics
• Machine learning in healthcare
• Medical informatics

Background:

• Inappropriate urine culture (UC) testing poses challenges for diagnostic stewardship.
• Urinalysis (UA) parameters can potentially predict UC positivity.
• Developing data-driven tools is crucial for optimizing laboratory testing.

Purpose of the Study:

• To develop and validate the PittUDT algorithm for predicting UC positivity using UA data.
• To support a system-wide initiative to improve the appropriateness of UC testing.
• To identify low-risk urine specimens unlikely to yield pathogenic organisms.

Main Methods:

• A recursive partitioning decision tree algorithm (PittUDT) was trained on 19,511 paired UA and UC cases.
• Receiver operating characteristic (ROC) analysis identified key predictors: white blood cells, leukocyte esterase, and bacteria.
• The algorithm's performance was evaluated on a separate test dataset of 9,773 cases.

Main Results:

• The PittUDT algorithm achieved a negative predictive value above 90% on the test dataset.
• Key predictors (urine white blood cells, leukocyte esterase, bacteria) showed strong predictive ability (ROC AUCs 0.77-0.79).
• The algorithm resulted in a 30-60% negative proportion, with a false-negative proportion under 5%.

Conclusions:

• The PittUDT algorithm demonstrates adequate predictive ability for triaging urine specimens.
• This data-driven, rule-based approach can be implemented across various healthcare settings.
• Optimizing UA parameters for reflex UC protocols can enhance antimicrobial stewardship and reduce costs.