2218. Feasibility of Automated Prediction of Drug Resistance in Pneumonia in the Veteran Population.

Abstract

BackgroundTo enhance antimicrobial decision-making in community-onset pneumonia, several manually calculable clinical scores have been proposed to predict drug-resistant organisms (DRPs), most commonly methicillin-resistant Staphylococcus aureus or Pseudomonas aeruginosa. The Drug Resistance in Pneumonia (DRIP) score includes 10 features that can be calculated by providers at the bedside but is potentially automatically extractable from an electronic health record (EHR). We aimed to explore the feasibility of calculating an automated “eDRIP” score from the EHR in the Veteran’s Affairs (VA) population.MethodsWe extracted patient characteristics, features relevant to the DRIP score, and detection of DRPs among all inpatient admissions for pneumonia between August 29, 2010 and July 31, 2013 at VA hospitals nationwide using EHR data from the VA corporate data warehouse. We calculated an eDRIP score for each admission. We compared the prevalence of each electronically extracted feature to that reported in a separate study of manually extracted factors performed at the Salt Lake City VA by Babbel et al, and to the original DRIP score validation cohort from Webb et al.ResultsAmong 101,462 pneumonia admissions across 114 VA hospitals, 4% had a DRP detected on culture, 25% had an eDRIP ≥ 4, and 50% received broad-spectrum antibiotics. The Salt Lake City VA demonstrated slightly lower prevalence of eDRIP factors than the national population (table). Within the Salt Lake City VA, the EHR cohort and manually extracted Babbel cohort demonstrated similar prevalence of detected DRP’s, DRIP ≥ 4, and 8 of 10 features involved in the DRIP score (table). The eDRIP identified fewer hospitalizations with poor functional status and residence in long-term facilities.ConclusionIn a large population of veterans admitted for community-onset pneumonia, automated extraction of an eDRIP score from the EHR was promising, though in need of revision. While some extracted features had similar prevalence to manual review, others differed by a factor of 10 or more, which may reflect issues with data extraction. Further work is needed to optimize feature extraction and compare electronic to manual DRIP scores to determine its utility within the VA population. Disclosures All authors: No reported disclosures.