Characterization and validation of electronic medical record data for pharmacoepidemiologic research

Purpose: Electronic medical records (EMRs) are essential for pharmacoepidemiologic research; however, information on the quality of Canadian hospital EMR data is scarce. We aimed to test data validity in the Epic EMR for key pharmacoepidemiologic themes using QT-prolonging medications and major adverse cardiac events (MACEs) as the research question. Methods: An entity relationship diagram (ERD) was developed to navigate >20,000 tables. Computational data validation (comparison with Epic SlicerDicer and Canadian Institute for Health Information Discharge Abstract Database (CIHI-DAD) data) and manual data validation (chart review) were applied with iterative adjustments. Percent agreement with 95% confidence intervals was used to estimate data validity. Results: The ERD demonstrated that 43 tables are needed for key pharmacoepidemiologic research data. A cohort of 70,078 adult inpatients was used for computational validation and 2,281 patient charts for manual validation. Demographics (for example, age, sex, gender, admitting service, and critical care transfers); exposures (medication administrations); many potential confounders (lab results, interacting drugs, and electrocardiograms [EKGs]); and most timestamping (admission, medication administrations, death, lab results, and EKGs) were validated with >95% agreement. Inadequate agreement for MACE (other than death) brought the primary outcome MACE agreement to <90%, requiring patient-level linkage to CIHI-DAD to improve agreement. Comorbidities, including hypertension and heart failure, also required this linkage. Conclusion: A world-leading proprietary EMR after the painstaking development of an ERD and validation of key data fields showed that not all data are of sufficient quality for pharmacoepidemiologic research. Linked human-coded data are required for some diagnoses.