Toward Personalized Hemoglobin A1c Estimation for Type 2 Diabetes

Abstract

A considerable gap is noted between a measured value of hemoglobin A1c (HbA1c) and its estimation from blood glucose (BG) levels; this gap needs to be resolved to improve glycemic control in people with diabetes. We aimed to develop clustering-based personalized models that estimate the HbA1c level from continuous glucose monitoring (CGM) value using a real-world clinical dataset and a novel machine learning (ML) approach. We used data from 101 insulin-treated Koreans with type 2 diabetes at Korea University Ansan Hospital (Ansan, South Korea). The dataset consisted of CGM values, activity level, nutritional data, antidiabetic medication for ten days, and HbA1c levels. Using the dataset, we optimized a novel ML approach, ${K}$ -means local nonlinear regressor (NLR), and developed two types of estimation models. The proposed models showed 0.877% and 0.857% mean absolute differences (MADs) in estimating HbA1c levels, whereas Nathan's and Bergenstal's general models showed 1.296% and 1.157%, respectively, and Grossman's personalized models showed either 0.952% or 0.923%. The proposed model outperformed commonly used models, and the performance improvement in our model came from appropriate personalization--estimating HbA1c higher for patients with poor glycemic control. We expect that the proposed CGM-based HbA1c estimator can serve as a more powerful motivator for glucose control. The proposed model, along with the CGM data, can provide real-time integrated information on glycemic control, which allows diabetic patients to adjust their lifestyles and physicians to modify management plans more promptly. Ultimately, this may result in better glycemic control than using conventional HbA1c measurement.