Determination of glucose concentration from near infrared spectra using least square support vector machine

One of the many challenges for translating noninvasive glucose measurement into clinical practice is the calibration of the measuring instrument. In this work, least squares support vector regression (LS-SVR) has been used to develop a multivariate calibration model for determination of glucose concentration from near infra-red (NIR) spectra. The behaviour of developed model is studied on NIR spectra of a mixture composed of glucose, urea, and triacetin which spans from 2100 nm to 2400 nm with a spectral resolution of 1nm. The proposed model improved the standard error of prediction (SEP) from 49.4 mg/dL in case of Principal Component Regression (PCR) and 27.5 mg/dL in case of Principal Least Squares Regression (PLSR) to 19.4mg/dL.