Background: Blood cultures are frequently ordered in hospitalized patients, often to work up undifferentiated fever or leukocytosis. This leads to low rates of positive cultures and high burden of false positive results resulting in avoidable interventions and waste. A recent study in hospital medicine patients identified several predictors of blood stream infection (BSI), but the sample size was small and generalizability was limited. In the era of the Electronic Health Record (EHR), we are collecting a multitude of clinical data that can be leveraged to address this question on a larger scale. We used EHR data to investigate predictors of BSI in patients hospitalized on the medicine service at a large academic medical center.
Methods: Using EHR derived data, we created a retrospective cohort of all patients with blood cultures ordered during 2016 on the hospital medicine service. We excluded all culture episodes ordered in the ED or ICU. Blood cultures were initially adjudicated based on organism. Possible non-pathogenic results then underwent expert review to further characterize true and false positive cultures. Predictors of interest included vital signs, white blood cell count, lactate, admission diagnosis, age, central line presence, immunosuppression, and recent antibiotics. We then calculated likelihood ratios of each predictor for BSI.
Results: 3,131 blood cultures were ordered on the hospital medicine service in 2016, of which 165 were true positives and 59 false positives (contaminated). Significant predictors of true BSI included transient hypotension (LR 2.10, 95% CI 1.25-3.5), lactate >3 (LR 1.98, 95% CI 1.06-3.71), central line access (LR 1.50, 95% CI 1.18-1.91) and admission diagnosis of bacteremia/intravascular infection (LR 2.68, 95% CI 1.06-6.79). We did not find a relationship between recent antibiotics, fever, or leukocytosis and BSI. When stratified by admission diagnosis, we found lower likelihood of BSI in patients with pneumonia, urinary tract infection, or skin/soft tissue infection (Table 1).
Conclusions: This is the first study to leverage clinical EHR data on a large scale to investigate predictors of BSI. We demonstrated that commonly referenced factors (fever, leukocytosis) are not predictive of blood stream infection, while other factors (admission diagnosis, hypotension, elevated lactate, age <65 years, and presence of a central line) are associated with increased likelihood. This suggests a change in blood culture ordering practices may be indicated. We conclude that these objective and readily available EHR data can be used as predictors to create clinical decision support to help clinicians identify patients at high or very low risk of BSI.
.png)