A two-level partial least squares system for non-invasive blood glucose concentration prediction

In this study, we propose and demonstrate a novel two-Level Partial Least Squares (2L-PLS) architecture for non-invasive blood glucose concentration measurement. A total of 290 Near-Infrared (NIR) spectroscopy readings from six laser diodes with discrete wavelengths of between 1500 nm and 1800 nm are obtained together with blood glucose concentration readings collected via Oral Glucose Tolerance Test (OGTT) experiments from a healthy volunteer over 4 days. While the conventional approach to predicting the blood glucose concentrations is to use a single Partial Least Squares (PLS) or non-linear PLS model, these systems do not achieve a high level of accuracy. As such, a 2L-PLS system consisting of one PLS model at the first level and three at the second level is proposed to enhance the prediction accuracy. A non-linear 2L-PLS system based on the same structure is also investigated in this study. The proposed 2L-PLS systems show improvements of 10 to 12% in the number of predictions that fall below a 5% error margin as compared to single-level PLS systems.