Background: Escherichia coli is the most commonly isolated organism from dog urine cultures. Treatment guidelines advocate for amoxicillin as an empiric first-line treatment for sporadic cystitis. However, with rising antimicrobial resistance rates, it is critical to identify dogs at high risk for resistance to first-line antimicrobials so culture and susceptibility can be prioritized.
Objective: To determine if machine learning tools applied to clinicopathologic features can predict ampicillin-resistant (AmpR) Escherichia coli bacteriuria in dogs, a surrogate for amoxicillin resistance. Animals: 538 dogs.
Methods: Dogs with Escherichia coli bacteriuria and concomitant urinalysis from 2015-2021 were included. The Escherichia coli isolate was classified as resistant or susceptible to ampicillin based on VET01SCLSI breakpoints. Clinicopathologic features were collected (Table 1). These data were split into 80% training and 20% testing sets. Gradient-boosted tree machine learning models (MLM) were trained to predict AmpR, and performance on the test set was determined.
Results: Escherichia coli isolates were ampicillin susceptible in 398 and resistant in 120 dogs. The MLM accurately categorized 86.0% isolates in the test set, with a sensitivity of 93.9% (95% CI;88.2-99.7%) and specificity of 66.7% (95% CI;48.9-84.5%) for detecting AmpR. The area under the receiver operating characteristic curve was 0.956 (95% CI;0.917-0.994). The most contributive features included pyuria, antimicrobial administration in the preceding 30 days, and clarity of the urine. Conclusion and Clinical Importance: MLMs can predict the presence of AmpR Escherichia coli bacteriuria with high sensitivity using features available the day of clinical presentation. This may assist veterinarians in therapeutic and diagnostic decision-making.
