A Pragmatic Machine Learning Model To Predict Carbapenem Resistance

• Published in: Antimicrobial agents and chemotherapy Vol. 65; no. 7; p. e0006321
• Main Authors: McGuire, Ryan J, Yu, Sean C, Payne, Philip R O, Lai, Albert M, Vazquez-Guillamet, M Cristina, Kollef, Marin H, Michelson, Andrew P
• Format: Journal Article
• Language: English
• Published: United States American Society for Microbiology 17.06.2021
• Subjects: Carbapenems - pharmacology; Epidemiology and Surveillance; Humans; Machine Learning; Predictive Value of Tests; Retrospective Studies; Risk Assessment; carbapenem; machine learning; antibiotic resistance; electronic health record; predictive modeling
• ISSN: 0066-4804; 1098-6596
• DOI: 10.1128/AAC.00063-21

Infection caused by carbapenem-resistant (CR) organisms is a rising problem in the United States. While the risk factors for antibiotic resistance are well known, there remains a large need for the early identification of antibiotic-resistant infections. Using machine learning (ML), we sought to develop a prediction model for carbapenem resistance. All patients >18 years of age admitted to a tertiary-care academic medical center between 1 January 2012 and 10 October 2017 with ≥1 bacterial culture were eligible for inclusion. All demographic, medication, vital sign, procedure, laboratory, and culture/sensitivity data were extracted from the electronic health record. Organisms were considered CR if a single isolate was reported as intermediate or resistant. Patients with CR and non-CR organisms were temporally matched to maintain the positive/negative case ratio. Extreme gradient boosting was used for model development. In total, 68,472 patients met inclusion criteria, with 1,088 patients identified as having CR organisms. Sixty-seven features were used for predictive modeling. The most important features were number of prior antibiotic days, recent central venous catheter placement, and inpatient surgery. After model training, the area under the receiver operating characteristic curve was 0.846. The sensitivity of the model was 30%, with a positive predictive value (PPV) of 30% and a negative predictive value of 99%. Using readily available clinical data, we were able to create a ML model capable of predicting CR infections at the time of culture collection with a high PPV.