Using Linked Data: Agreement in Diagnosis for Patients with Asthma or Diabetes

OBJECTIVES: Data linkage is a valuable resource for observational studies. This study evaluates the agreement between identification of patients with asthma and diabetes using diagnosis codes from electronic health records (EHRs) and claims data.

METHODS: Data from the Oracle EHR Real-World Data (105.2M patients) and a national US claims data set (233.2M patients) were linked (35.4M). A linked study cohort included patients with ≥1 EHR encounter and ≥1 claim from 01/07/2021-30/06/2022 (15.2M). Two subgroups were identified using diagnosis codes: patients with ≥2 diagnoses with a minimum 30 days from first diagnosis of asthma or of type 2 diabetes. Patient age was restricted to 2-89 years (asthma) or 18-89 years (diabetes). Descriptive analyses were performed.

RESULTS: A total of 111,694 and 962,450 patients with asthma and 253,669 and 1,699,360 patients with diabetes were identified from EHR and claims, respectively. Diagnostic agreement between data sources was low (<3%) for patients with asthma or diabetes. Specifically, 84,407 patients met the diagnostic criteria for asthma in both data sources, representing 75.57% and 8.77% of patients with asthma identified from EHR and claims, respectively. For diabetes, 201,189 patients met the diagnostic criteria in both data sources, representing 79.31% and 11.90% of patients with diabetes identified from EHR and claims, respectively. Among patients identified in both data sources, agreement of ≥1 diagnosis date was 74.7% (asthma) and 51.1% (diabetes). Differences in demographic characteristics of patients identified from EHR and claims datasets were primarily for missing data. For example, race was missing in 3.6% (asthma) and 3.3% (diabetes) of patients identified from EHR compared to 15.1% (asthma) and 16.5% (diabetes) from claims.

CONCLUSIONS: Cohorts derived from linked data may be smaller than anticipated although represent more accurate patient crossover. Data linkage maximizes available information, however, requires careful examination of attrition and characteristics by data source independently and jointly.