Diabetes
Validation of an Algorithm Based on Electronic Health Record Data to Identify Participants of All of Us With Type 1 Diabetes Mellitus Anna M Pederson* Anna M Pederson Scott C Zimmerman Peter Buto Jingxuan Wang Kendra D. Sims Audrey R. Murchland Mabeline Velez M. Maria Glymour Alana T Brennan Jennifer Weuve
Introduction: Diabetes mellitus (DM) is a leading risk factor for morbidity and mortality, but diagnostic algorithms for distinguishing DM subtypes in large observational, electronic health records (EHR) based cohorts are insufficiently validated. We aimed to develop an algorithm to distinguish individuals with type 1 DM (T1DM) from type 2 DM (T2DM) using EHR subcohorts with additional information on diabetes type.
Methods: We extracted data on participants in the All of Us Research Program (AoU) (N = 341,209) with an EHR record of diabetes who also: completed a self-reported diabetes questionnaire (n=12,464), had a C-peptide laboratory measurement (n=596), or had an islet-specific autoantibody (ISA) measurement (n=393). Three “gold standards” were used to identify T1DM: (1) self-reports via survey (T1DM vs T2DM); (2) C-peptide levels (T1DM: ≤ 0.20 nmol/L vs T2DM: ≥ 0.40 nmol/L); and (3) ISA positivity (T1DM: antibody > 0). We identified all EHR encounters for T1DM or T2DM and determined the optimal number of T1DM encounters to maximize sensitivity and specificity of classifying participants as T1DM versus T2DM.
Results: Among self-reported diabetes, C-peptide, and ISA laboratory measures, we identified, respectively: 1,525, 110, and 182 T1DM cases; and 10,939, 486, and 211 T2DM cases. Among A0U participants with diabetes, sensitivity and specificity were optimized at the following thresholds for EHR encounters: (1) ≥1 encounter for self-reported T1DM (sensitivity=0.63; specificity=0.89); (2) ≥4 encounters for C-peptide-defined T1DM (sensitivity=0.75; specificity=0.80); and (3) ≥5 encounters for ISA-defined T1DM (sensitivity=0.57; specificity=0.72).
Conclusion: We present a validated algorithm for identifying T1DM among diabetics in AoU. Our algorithm demonstrates good, but not perfect sensitivity and only modest specificity. Future work should quantify the direction and magnitude of bias to studies in which both DM subtypes are exposures or outcomes.
