The use of continuous glucose monitors (CGMs) among people with type 2 diabetes continues to expand beyond those treated with insulin or at high-risk for hypoglycemia. Defining precision CGM utilization strategies and augmenting A1c monitoring approaches is increasingly important to inform the application of episodic or continuous wearable data in clinical, payer, and digital health settings. We offer a method to predict 90-day glucose from 30 days of CGM data, offering an accelerated intervention window for people with T2D at low risk of hypoglycemia and no insulin treatment.
We used linear regression to predict 90-day glucose from initial median glucose and glycemic variability (Fig1A - eq 2). Of 5516 people with T2D in a digital diabetes program with rtCGM and no insulin use (2019 - 2022), 497 had sufficient data for inclusion (age 56.5 +/- 8.9; 49.5% F, 75% with > 81 days CGM wear in the first 90 days). The data was 80/20 train/test split. The model predicted median glucose at 77 - 90 days with MAE +/- 14.7 mg/dL, RMSE 21.9 mg/dL and adj R2 =0.424 for training; and MAE +/- 16.7 mg/dL, RMSE +/- 26.0 mg/dL for testing. Fig 1B shows the distribution of the MAE of full dataset across different initial median glucose levels. Hence, based on the plot, error IQR and outlier boundary increase beyond 177.6 mg/dL, the model may be best used with initial glucose level < 177.6 mg/dL.
Z.Li: Employee; UnitedHealth Group, Stock/Shareholder; UnitedHealth Group. C.Clark: Employee; UnitedHealth Group, Stock/Shareholder; UnitedHealth Group. S.Saha: Employee; UnitedHealth Group, Stock/Shareholder; UnitedHealth Group.