Predictive modeling of antimicrobial resistance in zoonotic Salmonella enterica: a retrospective machine learning analysis of national surveillance data (2000–2020)

Zoonotic pathogens are an important One Health issue from the perspective of antimicrobial resistance. However, surveillance datasets remain underutilized in this area, which can lessen the risks linked to antimicrobial resistance. This research examined how well machine learning techniques could accurately predict tetracycline resistance in Salmonella enterica isolates obtained from veterinary sources. Data collected between 2000 and 2020 from the Food and Drug Administration national antimicrobial resistance monitoring system was studied. This article focused on developing a classification–based predictive model for the classification of serovars of salmonella pathogens using machine– learning tools. Through using 5–fold stratified cross–validation was conducted for model performance evaluation along with clinical utility via prevailing negative predictive value and 95% confidence intervals. Tetracycline resistance was found to be overall 12.7%. The predictive performance of XGBoost was the best at AUC = 0.817 (95% CI: 0.802–0.829) than logistic regression (0.790). The internal stability of the model was strong and the negative predictive value was 94.9%, a good identification of susceptible isolates. Analysis of feature importance of resistance indicated that serotype was the most important variable, followed by temporal and spatial variables. The results shed light on the potential of gradient–boosted models for antimicrobial resistance risk stratification for antimicrobial resistance issues as well as antimicrobial stewardship, especially in the setting of guiding empirical therapy in the absence of laboratory results.