Can Administrative Claims Data Be Used to Determine Pre-treatment HbA1c Among Adults with Type 2 Diabetes?

Background: When evaluating the effectiveness and safety of pharmacotherapies for type 2 diabetes, adjustment for pre-treatment hemoglobin A1c (HbA1c) is usually needed for proper confounding control. Traditional claims data, commonly used in pharmacoepidemiologic research, often lack clinical measurements on HbA1c, but it is unknown if other information in claims data can be used to determine pre-treatment HbA1c.

Objectives: 1) To measure the concordance of pre-treatment HbA1c categories based on electronic health record (EHR) data versus outpatient lab results linked to claims data, and 2) To use machine learning to predict EHR-based pre-treatment HbA1c from claims-based features.

Methods: We used data from the OptumLabs® Data Warehouse, containing de‐identified EHR and administrative claims data, including linked outpatient ("claims-affiliated") lab results. Emulating eligibility criteria from a prior published trial, we identified metformin-treated adults with type 2 diabetes who initiated combination therapy with liraglutide or glimepiride between Feb 2010 and Jun 2019. Eligible initiators also had a pre-treatment HbA1c result available in their linked EHR data. Among the subgroup of initiators with a claims-affiliated pre-treatment HbA1c result available, we measured the concordance of pre-treatment HbA1c categories ( <8%, 8%- <9%, ≥9%) from lab results in EHR versus claims data. Among all initiators, we derived and internally validated (using a 9:1 train-test split) an ensemble learner combining 6 machine learning algorithms to predict EHR-based pre-treatment HbA1c, as a continuous variable, from 46 claims-based features describing demographics, comorbidities, medications, illness severity, and claims-affiliated pre-treatment HbA1c results.

Results: Among 2,887 eligible initiators identified, 43% were female, the median (IQR) age was 62 (54-70) years, and the median (IQR) EHR-based pre-treatment HbA1c was 8.0% (7.4%-8.9%). Among 21% of initiators with a claims-affiliated pre-treatment HbA1c result, the concordance of EHR-based versus claims-affiliated pre-treatment HbA1c categories was excellent (weighted kappa 0.90, 95% CI 0.87-0.93). Among all initiators, when claims data were used to predict EHR-based pre-treatment HbA1c, the ensemble learner performed poorly (R2=18%), with notably better performance among adults with versus without a claims-affiliated pre-treatment HbA1c result (R2=59% versus 9%).

Conclusions: Pre-treatment HbA1c measured using claims-affiliated lab results had excellent concordance with the EHR, but when this information was unavailable, claims-based prediction models had poor ability to predict EHR-based pre-treatment HbA1c.