(147) Development and Validation of a Novel Tool to Predict Hemoglobin A1c among Patients with Type 2 Diabetes, Using Administrative Claims Data

Background: Among patients with type 2 diabetes (T2DM), elevated hemoglobin A1c (HbA1c) levels are associated with increased risk of complications and may influence prescribing decisions for glucose-lowering medications. Although HbA1c is considered an important confounder and risk factor in many pharmacoepidemiologic studies, it is typically not available in claims data. To date, no prior study has developed an algorithm based on claims data to predict HbA1c level among patients with T2DM.

Objectives: To develop and validate a claims-based tool for ascertaining the A1c value in the population with T2DM.

Methods: Our study cohort included adults with established T2DM and at least 2 valid laboratory assessments of HbA1c within 180 days, in a national EHR network linked to the Optum Clinformatics® Data Mart Database (from 2004/1/1- 2022/3/1), randomly divided into a training dataset and validation set in a 70:30 ratio. Using the least absolute shrinkage and selection operator (LASSO), we selected among 124 investigator-specified variables to develop models for predicting continuous HbA1c. We assessed model performance by area-under-the-receiver operating characteristic curves (AUC) when predicting a binary outcome of HbA1 < 7 vs. >=7. We also measured positive predictive value (PPV) when predicting HbA1 >=7 in the general population and the high-risk subgroups with cirrhosis, cardiovascular disease (CVD), chronic kidney disease (CKD), and in those aged >=65 years.

Results: We identified 1,820,785 patients in the training dataset and 780,336 patients in the internal validation set. In both datasets, the median age was 67 years, 50.6% of the population were female and the median observed HbA1c was 6.6%. The final model included 117 selected variables, and the AUC for predicting HbA1c >= 7 was 0.79, with PPV=0.80. The model had robust performance when applied to subgroups of patients with T2DM and cirrhosis (AUC=0.76, PPV for predicting A1c >=7: 0.80), CVD (AUC=0.78, PPV for predicting A1c >= 7: 0.80), CKD (AUC=0.77, PPV for predicting A1c >=7: 0.80), or age >= 65 years (AUC=0.77, PPV for predicting A1c >=7: 0.81).

Conclusions: The HbA1c prediction algorithm could be used to improve phenotyping of T2DM in administrative claims datasets without laboratory data.