Phenotyping Diabetes Mellitus on Aggregated Electronic Health Records from Disparate Health Systems

Abstract

Background: Identifying patients with diabetes mellitus (DM) is often performed in epidemiological studies using electronic health records (EHR), but currently available algorithms have features that limit their generalizability. Methods: We developed a rule-based algorithm to determine DM status using the nationally aggregated EHR database. The algorithm was validated on two chart-reviewed samples (n = 2813) of (a) patients with atrial fibrillation (AF, n = 1194) and (b) randomly sampled hospitalized patients (n = 1619). Results: DM diagnosis codes alone resulted in a sensitivity of 77.0% and 83.4% in the AF and random hospitalized samples, respectively. The proposed algorithm combines blood glucose values and DM medication usage with diagnostic codes and exhibits sensitivities between 96.9% and 98.0%, while positive predictive values (PPV) ranged between 61.1% and 75.6%. Performances were comparable across sexes, but a lower specificity was observed in younger patients (below 65 versus 65 and above) in both validation samples (75.8% vs. 90.8% and 60.6% vs. 88.8%). The algorithm was robust for missing laboratory data but not for missing medication data. Conclusions: In this nationwide EHR database analysis, an algorithm for identifying patients with DM has been developed and validated. The algorithm supports quantitative bias analyses in future studies involving EHR-based DM studies.