Overcoming Individual Discrepancies, a Learning Model for Non-Invasive Blood Glucose Measurement

Abstract

Non-invasive Glucose Measurement (NGM) technology makes great sense for the blood glucose management of patients with hyperglycemia or hypoglycemia. Individual Discrepancies (IDs), e.g., skin thickness and color, not only block the development of NGM, but also become the reason why NGM cannot be widely used. To solve this problem, our solution is designing an individual customized NGM model that can measure these discrepancies through multi-wavelength and tune parameters for glucose estimating. In this paper, an NGM prototype is designed, and a learning model for glucose estimating with automatically parameters tuning based on Independent Component Analysis (ICA) and Random Forest (RF) is presented. The clinic trial proves that the correlation coefficient between estimation and reference Blood Glucose Concentration (BGC) can reach 0.5 after merely 10 times of learning, and rise to 0.8 after about 60 times of learning.