Orthogonal signal correction of near-infrared spectra for the prediction of glucose in human tissue

With the number of people being diagnosed with diabetes increasing globally, the development of a noninvasive method to monitor glucose levels has become an important field of research. One area that shows promise is the use of near-infrared (near-IR) spectroscopy in conjunction with mathematical and statistical methods for data analysis. An issue that arises with this methodology is the complexity of the skin, with many components that inhibit the ability to predict blood glucose levels. To address this issue, this work investigated the use of a data preprocessing method known as orthogonal signal correction (OSC) to remove information that is unrelated to that of glucose.

To begin, four OSC methods were tested and compared for use in predicting components in liquid samples that were measured in a longitudinal study over 617 days. From this study, the OSC method known as direct orthogonal signal correction (DOSC) was found to be most effective in improving predictions over time. Continuing with the liquid samples, a method to update the calibration model was explored through the use of additional known samples collected on the same day as the samples to be predicted. It was found that updating the calibration and using DOSC helped improve prediction performance significantly. Lastly, these developed methods were tested on noninvasive near-IR human tissue data to predict blood glucose concentrations across several days. As in the work with the liquid samples, the use of DOSC and model updating was found to be beneficial in achieving improved prediction performance. While the results obtained in this work are not yet sufficient for practical clinical implementation, the developed methodology shows promise to be used in future work as research toward a practical noninvasive glucose measurement continues.