Beyond the Mean: A Machine Learning-Based Trend Analysis of CGM Metrics for Improved HbA1C Estimation in Type 2 Diabetes.

Abstract

Background: Hemoglobin A_1C (HbA_1C) is the gold standard for assessing long-term glycemic control in people with diabetes. Increasing use of continuous glucose monitoring (CGM) has led to adoption of the glucose management indicator (GMI) as a CGM‑based HbA_1C estimate, but GMI often differs from laboratory HbA_1C, especially in type 2 diabetes. This discordance may be associated with the fact that GMI, as a measure of central tendency, fails to capture temporal glycemic trends and variability that relate to HbA_1C formation.

Objective: To evaluate whether combining CGM-derived metrics capturing variability, excursions, and temporal trends improves estimation of laboratory-measured HbA_1C in type 2 diabetes.

Methods: A machine learning framework was applied to CGM data from a three-month randomized trial, including 159 participants with type 2 diabetes. Participants had ≥70% CGM data coverage and valid end-of-trial HbA_1C. From a standardized 90-day CGM window, 51 metrics were extracted. Benchmark models (mean glucose and GMI) were compared with models developed using forward and exhaustive feature selection with threefold cross-validated multiple linear regression.

Results: Benchmark models yielded R-squared = 0.53. A forward selection model including five metrics (GMI at night, night-to-overall mean glucose ratio, glycemic risk assessment diabetes equation, time in tight range [3.0-7.8 mmol/L], time above range [13.9 mmol/L] at night) improved R-squared to 0.60. The best-performing model (substituting GRADE at night for GMI at night) achieved a similar R-squared (0.61). Nighttime and hyperglycemia‑related metrics were consistently selected.

Conclusion: Continuous glucose monitoring‑based HbA_1C estimation improves when variability and temporal patterns are included. Nighttime hyperglycemia adds notable predictive value, though further validation is needed.