Purpose: Identification of hospitalizations for infection is important for post-marketing surveillance of drugs, but the validity of using diagnosis codes to identify these events is unknown. Differentiating between hospitalization for and with infection is important, as the latter is common and less likely to arise from pre-admission exposure to drugs. We determined positive predictive values (PPVs) of diagnostic coding-based algorithms to identify hospitalization for infection among patients prescribed oral anti-diabetic drugs (OADs).
Methods: We identified patients initiating OADs within 2 United States claims databases (Medicare, HealthCore Integrated Research DatabaseSM [HIRDSM ]) and 2 United Kingdom electronic medical record databases (Clinical Practice Research Datalink [CPRD], The Health Improvement Network [THIN]) from 2009 to 2014. To identify potential hospitalizations for infection, we selected patients with a hospital diagnosis of infection and, within 7 days prior to hospitalization, either an outpatient/emergency department visit with an infection diagnosis or outpatient antimicrobial treatment. Hospital records were reviewed by infectious disease specialists to adjudicate hospital admissions for infection. PPVs for confirmed outcomes were determined for each database.
Results: Code-based algorithms to identify hospitalization for infection had PPVs exceeding 80% within Medicare (PPV, 83% [90/109]; 95% CI, 74-89%), HIRDSM (PPV, 89% [73/82]; 95% CI, 80-95%), and THIN (PPV, 86% [12/14]; 95% CI, 57-98%) but not within CPRD (PPV, 67% [14/21]; 95% CI, 43-85%).
Conclusions: Algorithms identifying hospitalization for infection utilizing hospital diagnoses along with antecedent outpatient/emergency infection diagnoses or antimicrobial therapy had sufficiently high PPVs for confirmed events within Medicare, HIRDSM , and THIN to enable their use for pharmacoepidemiologic research.
Keywords: ICD-9; diabetes mellitus; infection; pharmacoepidemiology; validation.
Copyright © 2017 John Wiley & Sons, Ltd.