Partial Least Squares Regression On Light Transmittance Spectra For Purpose Of Predicting Glucose Concentrations Of Unknown Spectra

Abstract

This project is motivated by non-invasive glucose sensing, with direct applications to Diabetes. Glucose spectra in the mid-infrared region of light contain unique spectral features. The idea is to use shine Quantum Cascade Lasers (QCLs) into the skin, and the light that is reflected back should mimic the spectra of blood glucose. QCLs are chosen for their power, which is good for deeper penetration into skin, and for their tenability, which can take advantage of the unique spectral features mentioned above. As a preliminary study, spectra are produced from either QCLs or an FTIR Spectrometer transmitting through glucose solution of varying concentrations. The ultimate goal is to predict concentrations from spectra, and to this end, partial least squares regression (PLSR) is used. This regression forms PLS components mapping spectra to their known concentrations, and using these components to then map spectra of unknown concentration to their concentrations (though in this early stage, the concentration are known and used for comparison purposes to measure accuracy of the regression). This project considered pre-processing the spectra, in order to obtain more accurate predictions from PLSR.