Regularised Model Identification Improves Accuracy of Multisensor Systems for Noninvasive Continuous Glucose Monitoring in Diabetes Management

Abstract

Continuous glucose monitoring (CGM) by suitable portable sensors plays a central role in the treatment of diabetes, a disease currently affecting more than 350 million people worldwide. Noninvasive CGM (NI-CGM), in particular, is appealing for reasons related to patient comfort (no needles are used) but challenging. NI-CGM prototypes exploiting multisensor approaches have been recently proposed to deal with physiological and environmental disturbances. In these prototypes, signals measured noninvasively (e.g., skin impedance, temperature, optical skin properties, etc.) are combined through a static multivariate linear model for estimating glucose levels. In this work, by exploiting a dataset of 45 experimental sessions acquired in diabetic subjects, we show that regularisation-based techniques for the identification of the model, such as the least absolute shrinkage and selection operator (better known as LASSO), Ridge regression, and Elastic-Net regression, improve the accuracy of glucose estimates with respect to techniques, such as partial least squares regression, previously used in the literature. More specifically, the Elastic-Net model (i.e., the model identified using a combination of and norms) has the best results, according to the metrics widely accepted in the diabetes community. This model represents an important incremental step toward the development of NI-CGM devices effectively usable by patients.