Machine learning to predict urine culture antibiotic sensitivities in the emergency department.

Abstract

Background: Urinary tract infections (UTIs) are among the most common bacterial infections diagnosed in the emergency department. Treatment of UTIs is largely empiric because urine culture results are not rapidly available.

Objectives: We examined whether machine learning could predict antibiotic sensitivities of the urine cultures by using only data available during the clinical encounter.

Methods: We used extreme gradient boosting (XGBoost) to examine 62,963 emergency department patient encounters from January 1, 2017, through December 31, 2021. All encounters included a urinalysis and urine culture. We included 1303 variables in the full model and examined 21 antibiotics. An antibiotic was characterized as sensitive only if all bacteria in the culture were susceptible; if ≥ 1 bacteria was not susceptible, then it was characterized as intermediate or resistant.

Results: We predicted urine cultures to be sensitive vs intermediate or resistant with area under the receiver operating curve (AUROC) values ranging from 70 % (for amikacin) to 90 % (for linezolid) (median, 82 %) when negative urine cultures were characterized as antibiotic susceptible. AUROCs were as follows: nitrofurantoin (84 %); trimethoprim + sulfamethoxazole (80 %); ciprofloxacin (85 %); levofloxacin (85 %); first-generation cephalosporins (84 %); and third-generation cephalosporins (80 %). When models excluded urine cultures with no bacterial growth, AUROCs ranged from 66 % (for ampicillin) to 87 % (for amikacin) (median, 74 %). When models included only patients diagnosed with a UTI plus bacteriuria (≥10,000 colony-forming units per mL in urine culture), AUROCs ranged from 63 % (for ampicillin) to 85 % (for tetracycline) (median, 74 %).

Conclusion: XGBoost can predict bacteriuria antibiotic sensitivities.