Electronic operational classification of hospital-onset bacteremia and fungemia in Epic Cosmos: A retrospective observational study.

BACKGROUND

Hospital-onset bacteremia and fungemia (HOB) has been proposed as an electronic surveillance outcome, but deidentified electronic health record data require transparent organism handling when chart review is unavailable.

METHODS

We analyzed deidentified HOB records from Epic Cosmos and retained the first positive blood-culture event per admission on or after hospital day 4. Organism text was standardized and linked to a locally developed workbook based on common-commensal organism logic and review by infectious diseases physicians and laboratory directors. Events were assigned to 4 algorithm-defined operational classes: likely pathogen, possible pathogen, likely contaminant, or unclassified. We summarized class distribution and organism patterns. A logistic model comparing likely contaminant with likely pathogen events was exploratory.

RESULTS

Among 77,261 event-level records, 69,303 first events per admission comprised the analytic cohort. Events were likely pathogen (44,239; 63.8%), possible pathogen (6,093; 8.8%), likely contaminant (663; 1.0%), or unclassified (18,308; 26.4%). Unclassified events lacked usable organism tokens. Staphylococcus epidermidis accounted for 10.4% of classifiable organism tokens and 66.5% of likely contaminant tokens. The likely contaminant class was intentionally narrow.

CONCLUSIONS

A transparent electronic algorithm classified HOB events into operational surveillance classes at scale. Site-level validation is needed before benchmarking or reporting.