169. Development of a Real Time Electronic Algorithm to Identify Hospitalized Patients with Community-Acquired Pneumonia

BACKGROUND: Syndrome-based antibiotic stewardship can be limited by difficulty in finding cases for evaluation. We developed an electronic extraction algorithm to prospectively identify CAP patients. METHODS: We included non-oncology patients ≥18 years old admitted to The Johns Hopkins Hospital from 12/2018 to 3/2019 who 1) received common CAP antibiotics for ≥48 hours after admission and 2) had a bacterial urinary antigen and chest imaging ordered within 48 hours of admission that was not for assessment of endotracheal tube or central line placement. Charts of patients meeting these criteria were reviewed by 2 authors to identify true cases of CAP based on IDSA guidelines. Cases identified in 12/2018 (n=111) were used to explore potential indicators of CAP, and cases identified 1–3/2019 (n=173) were used to evaluate combinations of indicators that could identify patients treated for CAP who did have CAP (true CAP) and did not have CAP (false CAP). This cohort was divided into a training and a validation set (2/3 and 1/3, respectively). Potential indicators included vitals signs, laboratory data and free text extracted via natural language processing (NLP). Predictive performance of composite indicators for true CAP were assessed using receiver-operating characteristics (ROC) curves. The Hosmer-Lemeshow goodness fit test was used to test model fit and the Akaike Information Criteria was used to determine model selection. RESULTS: True CAP was observed in 41% (71/173) of cases and 14 potential individual indicators were identified (Table). These were combined to make 45 potential composite indicators. ROC curves for selected composite indicators are shown in the Figure. Models without use of NLP-derived variables had poor discriminative ability. The best model included fever, hypoxemia, leukocytosis, and “consolidation” on imaging with a sensitivity and positive predictive value 78.7% and specificity and negative predictive value of 85.7%. Table. Indicators evaluated to identify patients with CAP [Image: see text] Figure. ROC curves for composite indicators [Image: see text] CONCLUSION: Patients with CAP can be identified using electronic data but use of NLP-derived radiographic criteria is required. These data can be linked with data on antibiotic use and duration to develop reports for clinicians regarding appropriate CAP diagnosis and treatment. DISCLOSURES: All Authors: No reported disclosures