Algorithm to Identify Type 2 Diabetes Using Electronic Health Record and Self-Reported Data.

Abstract

Aims: Identifying participants with type 2 diabetes (T2D) based only on electronic health record (EHR) or self-reported data has limited accuracy. Therefore, the objective of the study was to develop an algorithm using EHR and self-reported data to identify participants with and without T2D. Methods: We included participants enrolled in the Mayo Clinic Biobank. At enrollment, participants completed a baseline questionnaire on health conditions, including T2D, and provided access to their EHR data. T2D status was based on self-report and EHR data (International Classification of Diseases codes, hemoglobin A1c [HbA1c], plasma glucose, and glucose-regulating medications) within 5 years prior to and 2 months after enrollment. Participants who self-reported T2D but lacked corroborating EHR data were categorized separately ("only self-reported T2D"). After identifying participants with T2D, we identified participants without T2D based on normal HbA1c and plasma glucose. Participants who self-reported the absence of T2D but lacked corroborating EHR data were categorized separately ("only self-reported no T2D"). Using manual chart reviews (gold standard), we calculated the positive and negative predictive values (NPV) to identify T2D. Results: Of 57,000 participants, the algorithm classified participants as having T2D (n = 6,238), no T2D (n = 38,883), "only self-reported T2D" (n = 757), and "only self-reported no-T2D" (n = 9,759). The algorithm had a high positive predictive value (96.0% [91.5%-98.5%]), NPV (100% [98.0%-100%]), and accuracy (99.5% [98.3%-99.8%]). Participant age (median [range]) ranged from 52 (18-98) years (only self-reported T2D) to 67 (19-99) years (T2D) (P < 0.0001), and the proportion of women ranged from 45.3% (T2D) to 69.6% (only self-reported no T2D) (P < 0.0001). Most participants were of the White race (84.0%-92.7%) and non-Hispanic ethnicity (97.6%-98.6%). Conclusions: In this study, we developed an algorithm to accurately identify participants with and without T2D, which may be generalizable to cohorts with linked EHR data.