A Machine Learning Approach to Integrated Multidrug Resistance Surveillance and Clinical Decision Support
Principal Investigator: Casey Cazer
DESCRIPTION (provided by applicant):
Infections due to multidrug resistance (MDR) gram-negative bacteria (GNB) are endemic in the United States and their prevalence is increasing. MDR surveillance at the population level and early knowledge of MDR at the patient level are critical to improve treatment outcomes and the judicious use of antimicrobials. In this context, we aim to develop an integrated system for MDR surveillance and prediction to support antimicrobial treatment decisions and stewardship. However, the exponentially large number of potential MDR phenotypes impedes MDR analyses. Machine learning algorithms are well-suited for tackling this problem. First, we will build an early-warning system for MDR phenotypes of clinical significance using an unsupervised machine learning technique, association mining, to identify and quantify patterns within phenotypic and genotypic resistance data. We will analyze 508 E. coli bloodstream infection isolates, examining both phenotypic and genotypic MDR. By sequencing 200 of these isolates, we will confirm the genetic mechanisms responsible for the observed phenotypic associations. Second, we will develop supervised machine learning models to predict specific MDR phenotypes from medical record data. This will improve clinicians’ empiric antimicrobial use choices by providing information on MDR prior to the completion of laboratory susceptibility testing. We will assess the clinical impact of our integrated system by comparing historical antimicrobial prescribing at a population and individual patient level to recommendations based on the MDR surveillance and MDR prediction models. The proposed work will form the foundation of an R01 proposal to create an expanded MDR GNB surveillance and antimicrobial stewardship system.