Continuous Glucose Data Construction and Risk Assessment Application of Diabetic Retinopathy Complications for Patients with Type 2 Diabetes Mellitus

Abstract

Managing diabetes mellitus (DM) includes achieving acceptable blood glucose levels and minimizing the risk of complications from DM. The appropriate glucose sensing method is continuous glucose monitoring (CGM). Effective evaluation metrics that reflect glucose fluctuations can be realized. However, compared with self-monitoring of blood glucose (SMBG), CGM data are not easy to obtain. Therefore, this article studies a fusion model to achieve this objective, including Gaussian process regression (GPR) and long short-term memory (LSTM). Compared with the three commonly used LSTM, GPR, and support vector machine, the proposed model can construct accurate results. By using the constructed CGM data, the conventional metrics, such as the mean amplitude of glycemic excursion (MAGE), mean blood glucose (MBG), standard deviation (SD), and time in range (TIR), are calculated. These metrics and other variables are input into statistical methods to realize diabetic retinopathy risk assessment. In this way, the relationship between the glycemic variability of the constructed CGM data by the mathematical model and DR could be achieved. The utilized statistical methods include single-factor analysis and binary multivariate logistic regression analysis. Results show that fasting blood glucose, disease course, history of hypertension, MAGE and TIR are independent risk factors for DR.