Predictive Modeling for Early Hyperglycemia Detection in Type 2 Diabetes

Abstract

Type 2 Diabetes (T2D) is characterized by a combination of defects in insulin action and impairment in insulin secretion. Deficient insulin action causes people with Type 2 Diabetes to have difficulty controlling their blood glucose concentration (BGC) and experience periods of high (hyperglycemia) and low (hypoglycemia) BGC. Continuous glocuse monitoring (CGM) sensors and machine learning algorithms can automate the process of meal size estimation, improve the accuracy of the carbohydrate estimations, and reduce the involvement of the subject. The aim of this project was to use dynamic Partial Least Squares (PLS) regression to model blood glucose data from CGM devices and optimize the model parameters for best accuracy of the predictive modelling. The parameters were optimized by explicit enumeration and grid search approach to minimize the mean square error (MSE), and the lowest MSE was obtained with the number of latent variables as 5 and past horizon as 5 (25 minutes). Future research will develop the logic inference using the first- and second-order derivatives of the predicition curve that will sound the alarms based on the predicitions made in the current work.