(P-102) Determining the Accuracy of Developing New User Cohorts Using Electronic Health Record Data

Background: Electronic health record (EHR) data are ubiquitous and increasingly used in comparative effectiveness research. However, with few exceptions (i.e., vertically integrated health systems/insurers), EHR-based databases may often be missing important patient information due to discontinuity of care. Accordingly, developing new user cohorts from EHR data is challenging.

Objectives: We sought to determine the accuracy of developing new user cohorts from EHR data, using linked EHR-claims data as the gold standard. We further sought to assess whether misclassification of new users differed in certain contexts, e.g., across therapeutic areas or drug classes with varying characteristics.

Methods: We linked longitudinal Medicaid & Medicare claims data to the OneFlorida Clinical Research Network, a member of the Patient-Centered Outcomes Research Network. OneFlorida contains structured EHR data from major healthcare systems in Florida. We included patients having at least 365 days of overlapping EHR and Medicaid or Medicare (but not dual) enrollment. Using the EHR prescribing data, for each drug exposure, we identified the date of first prescription (i.e. index date). Patients were required to have ≥365 days of overlapping EHR and Medicaid or Medicare enrollment prior to the index date (i.e. 12-month baseline). Such patients were “presumed new users.” Among these, patients without any evidence of dispensing in the 12-month baseline period according to the claims data were identified as “true new users.”

Results: We included 138,340 patients with linked EHR-Medicaid claims data & 107,112 patients with linked EHR-Medicare claims data. In both populations, we observed high misclassification of new users, with substantial variation across drug therapy categories and classes. Among Medicaid patients, the proportion of presumed new users that were true new users was 39.2% for antihypertensives (range across antihypertensive classes,44.8% to 65.0%); 34.8% for antidiabetics (range, 40.3% to 85.2%); 74.6% for anticoagulants (range, 59.6% to 97.4%); 43.9% for antilipidemics (range, 43.0% to 89.9%); and 51.5% for antibiotics (range, 67.2% to 86.5%). Generally similar findings were observed in the Medicare population, in which EHR-based new-user cohorts were less often misclassified with short-term vs. long-term medications.

Conclusions: Substantial challenges exist in identifying true new users using EHR data, particularly for routine, chronic medications, e.g., antihypertensives and antidiabetics. Our results suggest that misclassification of new-user cohorts more generally, may be context-specific. Future research is needed to identify effective approaches to minimize such misclassification, particularly when linked claims data are unavailable.