A Machine-Learning Approach For Predicting Antibiotic Resistance in Pseudomonas aeruginosa. (October 2020)

• Record Type: Journal Article
• Authors: Magleby, Reed; Simon, Matthew; Calfee, David; Zachariah, Philip; Cohen, Bevin; Cazer, Casey; Wang, Fei; Larson, Elaine
• Is Part Of: Infection control and hospital epidemiology. Volume 41(2020)Supplement 1
• Page Start: s96
• Page End: s97
• Publication Date: 2020-10
• Subjects: Nosocomial infections -- Epidemiology -- Periodicals; Health facilities -- Sanitation -- Periodicals; Hospital buildings -- Sanitation -- Periodicals; Cross Infection -- Periodicals; Epidemiology -- Periodicals; Hospitals -- Periodicals; Infection Control -- Periodicals; 614.44
• DOI: 10.1017/ice.2020.596
• Languages: English
• ISSNs: 0899-823X

Abstract

Abstract : Background: Pseudomonas aeruginosa is an important nosocomial pathogen associated with intrinsic and acquired resistance mechanisms to major classes of antibiotics. To better understand clinical risk factors for drug-resistant P. aeruginosa infection, decision-tree models for the prediction of fluoroquinolone and carbapenem-resistant P. aeruginosa were constructed and compared to multivariable logistic regression models using performance characteristics. Methods: In total, 5, 636 patients admitted to 4 hospitals within a New York City healthcare system from 2010 to 2016 with blood, respiratory, wound, or urine cultures growing PA were included in the analysis. Presence or absence of drug-resistance was defined using the first culture of any source positive for P. aeruginosa during each hospitalization. To train and validate the prediction models, cases were randomly split (60 of 40) into training and validation datasets. Clinical decision-tree models for both fluoroquinolone and carbapenem resistance were built from the training dataset using 21 clinical variables of interest, and multivariable logistic regression models were built using the 16 clinical variables associated with resistance in bivariate analyses. Decision-tree models were optimized using K-fold cross validation, and performance characteristics between the 4 models were compared. Results: From 2010 through 2016, prevalence of fluoroquinolone and carbapenem resistance was 32% and 18%, respectively. For fluoroquinolone resistance, the logistic regression algorithm attained a positive predictive value (PPV) of 0.57 and a negative predictive value (NPV) of 0.73 (sensitivity, 0.27; specificity, 0.90) and the decision-tree algorithm attained a PPV of 0.65 and an NPV of 0.72 (sensitivity 0.21, specificity 0.95). For carbapenem resistance, the logistic regression algorithm attained a PPV of 0.53 and a NPV of 0.85 (sensitivity 0.20, specificity 0.96) and the decision-tree algorithm attained a PPV of 0.59 and an NPV of 0.84 (sensitivity 0.22, specificity 0.96). The decision-tree partitioning algorithm identified prior fluoroquinolone resistance, SNF stay, sex, and length-of-stay as variables of greatest importance for fluoroquinolone resistance compared to prior carbapenem resistance, age, and length-of-stay for carbapenem resistance. The highest-performing decision tree for fluoroquinolone resistance is illustrated in Fig. 1. Conclusions: Supervised machine-learning techniques may facilitate prediction of P. aeruginosa resistance and risk factors driving resistance patterns in hospitalized patients. Such techniques may be applied to readily available clinical information from hospital electronic health records to aid with clinical decision making. Funding: None Disclosures: None