Prediction of carbapenem-resistant Klebsiella-positive inpatient cultures using data from an electronic health record
Electronic algorithms to identify carbapenem-resistant Klebsiella (CRK) could improve infection control (IC) and guide empiric antibiotic therapy. We investigated the use of the large Veterans Affairs (VA) Electronic Health Record (EHR) database to predict CRK-positive (CRK+) clinical cultures.
In- and outpatient candidate predictor variables were extracted from VA EHR records between 2008 and 2012. Cases were defined as CRK+ admissions on acute care wards. Matched CRK-negative controls were randomly selected among admissions with clinical cultures on the hospital and calendar month of the case. Approximately 150,000 factors from the database (VA EHR) were screened for an association with CRK+ culture. 200 of the most highly associated variables were then used for logistic regression analysis. A conditional logistic regression was trained on half of the data set using stepwise selection (SAS Enterprise Miner 7.1). The model was validated on the remaining half.
In 667,637 admissions with cultures, the overall prevalence of CRK+ cultures was 0.39%. The area under the receiver operating characteristic curve was 0.85 on the training set and 0.83 on the validation set (Figure 1). A total of 19 factors were retained in the model, including prior healthcare and antibiotic exposures, diagnoses, clinical cultures, and physiologic parameters. Multi-collinearity precluded reporting individual predictors. Using a cutoff CRK+ culture risk of 0.63 yielded a sensitivity, specificity, and positive predictive value of 39.8%, 97.1%, and 5.0%, respectively.
A model with good predictive characteristics was derived from the large VA EHR database--a “first step” towards improving CRK infection control and empiric therapy. Because of the rarity of events, a substantial trade-off exists between false positives and sensitivity. Therefore, this model may be most useful in the context of higher-prevalence wards or hospitals. Our findings are limited to CRK+ clinical cultures and further work will be necessary to address CRK-carriage. To improve the accuracy of CRK prediction and control cost, a combination of initial evaluation with predictive analytics followed by targeted laboratory testing for carbapenem resistance may be of value.
K. Damal, None
K. Khader, None
R. Bonomo, AstraZeneca: Grant Investigator, Grant recipient
Merck: Grant Investigator, Grant recipient
Rib-X: Grant Investigator, Grant recipient
Steris: Grant Investigator, Grant recipient
TetraPhase: Scientific Advisor, nothing
NIH: Grant Investigator, Grant recipient
VA Merit Review: Grant Investigator, Grant recipient
M. Evans, None
C. Nielson, None