Rieselbach Distinguished Paper Session I

Background/Significance:

Unnecessary antibiotic prescribing for suspected urinary tract infections (UTIs) remains common in emergency departments (EDs). Machine learning (ML)–based clinical decision support tools have been proposed as a strategy to improve antibiotic stewardship, but their influence on prescribing behavior and clinician confidence remains unclear.

Purpose:

To determine the impact of an ML–based clinical decision support tool on antibiotic prescribing and clinician confidence for suspected UTI among ED providers.

Methods:

We conducted a randomized controlled, scenario-based survey with a crossover design to evaluate the impact of an ML-driven UTI decision support tool on antibiotic prescribing decisions among providers practicing at Advocate Health Midwest EDs. Participants completed eight standardized clinical scenarios representing three cases: (1) a patient with uncomplicated dysuria (4 urinalysis variations), (2) an elderly patient with altered mental status (2 variations), and (3) a patient with diabetic ketoacidosis (DKA) (2 variations). For each scenario, participants were randomized to receive or not receive an ML-generated probability of positive urine culture (range: 10.24% to 82.7%). Prescribing decisions and self-reported confidence (5-point Likert scale) were compared between ML-exposed and non-exposed groups using Fisher's exact tests and Mann–Whitney U tests.

Results:

Of 474 surveyed providers, 224 (47.3%) completed the survey. Four scenarios demonstrated significant differences in prescribing rates. The largest effect occurred in the DKA case with moderately abnormal urinalysis (ML probability 82.7%): prescribing increased from 21.4% without ML to 68.5% with ML (+47.1%, p<0.001). Similarly, ML exposure increased prescribing in the dementia case (ML probability 71.8%): 41.8% vs. 82.9% (+41.1%, p<0.001), and in the dysuria case (ML probability 71.8%): 47.7% vs. 72.1% (+24.3%, p<0.001). One scenario showed decreased prescribing: the DKA case with less abnormal urinalysis (ML probability 10.2%) had prescribing rates of 29.7% vs. 16.2% (-13.5%, p=0.025). In every scenario, prescribing decisions shifted in alignment with ML probabilities. ML exposure increased clinician confidence (mean 3.21 vs. 3.29, p=0.040).

Conclusion:

ML-generated probabilities influenced prescribing decisions, with clinician behavior shifting toward displayed probabilities. Further research is needed to assess real-world effects on patient outcomes and antibiotic stewardship.