ICD
‐10‐
CM
‐Based Algorithms to Identify Bacteremia Incorporating Pathogen‐Specific Codes: A Validation Study
Madison G. Ponder, Alan C. Kinlaw, Elizabeth S. Dodds Ashley, Jennifer L. Lund, Emily J. Ciccone, Michele Jonsson Funk ABSTRACT
Purpose
Studying patients with bacteremia in real‐world data often requires using microbiologic data to confirm bacterial growth and species. When microbiologic data are unavailable, pathogen‐specific diagnostic codes may be used to classify the bacteremia type. We developed and validated ICD‐10‐CM‐based algorithms for identifying any bacteremia, Gram‐negative bacteremia, and Enterobacteriaceae bacteremia.
Methods
We identified patients ≥ 18 years with a hospitalization ≥ 2 days to any Duke Antimicrobial Stewardship Outreach Network hospital from January 1, 2021 to May 31, 2024. Blood culture results were the reference standard for classifying true cases of bacteremia. We evaluated 15 candidate algorithms developed based on previously published algorithms, in addition to the bacteremia ICD‐10‐CM code (R78.81) and pathogen‐specific ICD‐10‐CM codes. We estimated sensitivity, specificity, positive predictive value (PPV), and negative predictive value with 95% confidence intervals using generalized estimating equations to account for correlated hospitalizations.
Results
Among 907 958 hospital admissions (623 556 patients), algorithms with bacteremia and pathogen‐specific codes had high specificity (range: 97.3%–100.0%) and low sensitivity (range: 4.9%–58.1%). Algorithms with bacteremia codes and no pathogen codes were more sensitive (range: 67.4%–68.6%) and less specific (range: 91.1%–91.2%). In subgroup analyses, PPV for any bacteremia was highest among patients alive at discharge, age ≥ 65, and discharged on antibiotics with Gram‐negative coverage (68.9% [65.1%–72.7%]).
Conclusions
We developed and internally validated algorithms to identify bacteremia, Gram‐negative bacteremia, and Enterobacteriaceae bacteremia in adult patients. Studies seeking to maximize specificity could consider algorithms that include bacteremia and pathogen‐specific codes, whereas studies seeking to maximize sensitivity could consider algorithms only requiring a bacteremia code.