(035) Assessing the utility of claims for haemoglobin A1c performance measures in characterizing glycemic control in older adult Medicare beneficiaries with diabetes.

Background: Among patients with diabetes , hemoglobin A1c (A1c) provides information on time-averaged glycemic control over a 3-month period and is a strong predictor of both medication initiation and clinical outcomes. Administrative claims data often lack detailed lab results; to account for this limitation, investigators may use claims for physician performance measures (PM) comprising broad A1c categories, to proxy for glycemic control. Agreement between A1c lab results and these claims for A1c performance measures is unknown.

Objectives: We aimed to evaluate the agreement of A1c PM claims in characterizing glycemic control among Medicare fee-for-service (FFS) beneficiaries with diabetes.

Methods: We linked patients with at least one encounter in an electronic health record (EHR) from a large health system in North Carolina (USA) to Medicare FFS claims during 2015-2019. We identified individuals >65 years old with A1c lab results in the EHR who had claims within the prior 6 months indicating prevalent diabetes (2 outpatient, 1 inpatient, or 1 antidiabetic medication claim). PM claims (CPT code 3044F: A1c < 7%, 3045F: 7% ≤ A1c ≤ 9%, and 3046F: A1c > 9%) were paired with either (a) the most recent EHR-based A1c result (upto 90 days earlier); or (b) the closest EHR-based A1c result earlier that calendar year given no claims for A1c test (CPT code 83036, 83037) during that time interval. We calculated agreement between A1c lab results and PM claims across categories of A1c ( < 7%, ≥7% and ≤9%, or >9%), and estimated sensitivity and specificity of PM claims for classifying adequate glycemic control (A1c < 7% vs ≥7%).

Results: We identified 975 A1c eligible results with a corresponding PM claim. Overall, approximately 1% of PM claims could be paired with an EHR recorded A1c result that met inclusion criteria. 71% of A1c results were correctly classified and weighted kappa across the three-categories was 0.54 (95% confidence interval 0.50, 0.59). The PM code for A1c level < 7% had a sensitivity of 0.80 (0.77, 0.84) and a specificity of 0.83 (0.80, 0.86). Error probabilities reflecting misclassification between categories ranged from 7% to 31%. Sensitivity, and specificity were similar when restricting to PM claim/A1c pairs without another claim for an A1c test (n=680).

Conclusions: Although performance measure codes for A1c are primarily used in health maintenance or preferred provider organization plans, reporting in Medicare FFS claims may provide insight into glycemic control in the claim year and therefore, potential for adjustment for confounding by glycemic control when comparing the effects of alternative treatments for diabetes. Reported error probabilities, sensitivity and specificity may be useful for conducting additional quantitative bias analyses.