Time in Range Estimation in Patients with Type2 Diabetes is Improved by Incorporating Fasting and Postprandial Glucose Levels.

Abstract

Time in range (TIR) as assessed by continuous glucose monitoring (CGM) measures an individual's glucose fluctuations within set limits in a time period and is increasingly used together with HbA1c in patients with diabetes. HbA1c indicates the average glucose concentration but provides no information on glucose fluctuation. However, before CGM becomes available for patients with type2 diabetes (T2D) worldwide, especially in developing nations, fasting plasma glucose (FPG) and postprandial plasma glucose (PPG) are still the common biomarkers used for monitoring diabetes conditions. We investigated the importance of FPG and PPG to glucose fluctuation in patients with T2D. We used machine learning to provide a new estimate of TIR based on the HbA1c, together with FPG and PPG. This study included 399 patients with T2D. (1) Univariate and (2) multivariate linear regression models and (3) random forest regression models were developed to predict the TIR. Subgroup analysis was performed in the newly diagnosed T2D population to explore and optimize the prediction model for patients with different disease history. Regression analysis suggests that FPG was strongly linked to minimum glucose, while PPG was strongly correlated with maximum glucose. After FPG and PPG were incorporated into the multivariate linear regression model, the prediction performance of TIR was improved compared with the univariate correlation between HbA1c and TIR, and the correlation coefficient (95%CI) increased from 0.62 (0.59, 0.65) to 0.73 (0.72, 0.75) (p < 0.001). The random forest model significantly outperformed the linear model (p < 0.001) in predicting TIR through FPG, PPG and HbA1c, with a stronger correlation coefficient 0.79 (0.79, 0.80). The results offered a comprehensive understanding of glucose fluctuations through FPG and PPG compared to HbA1c alone. Our novel TIR prediction model based on random forest regression with FPG, PPG, and HbA1c provides a better prediction performance than the univariate model with solely HbA1c. The results indicate a nonlinear relationship between TIR and glycaemic parameters. Our results suggest that machine learning may have the potential to be used in developing better models for understanding patients' disease status and providing necessary interventions for glycaemic control.