Optimization of mid-infrared noninvasive blood-glucose prediction model by support vector regression coupled with different spectral features.

Mid-infrared spectral analysis of glucose in subcutaneous interstitial fluid has been widely employed as a noninvasive alternative to the standard blood-glucose detection requiring blood-sampling via skin-puncturing, but improving the confidence level of such a replacement remains highly desirable. Here, we show that with an innovative metric of attributes in measurements and data-management, a high accuracy in correlating the test results of our improved spectral analysis to those of the standard detection is accomplished. First, our comparative laser speckle contrast imaging of subcutaneous interstitial fluid in fingertips, thenar and hypothenar reveal that spectral measurements from hypothenar, with an attenuated total reflection Fourier transform infrared spectrometer, give much stronger signals than the stereotype measurements from fingertips. Second, we demonstrate that discriminative selection of the spectral locations and ranges, to minimize spectral interference and maximize signal-to-noise, are critically important. The optimal band is pinned at that between 1000 ± 3 cm and1040 ± 3 cm. Third, we propose an individual exclusive prediction model by adopting the support vector regression analysis of the spectral data from four subjects. The average predicted coefficient of determination, root mean square error and mean absolute error of four subjects are 0.97, 0.21 mmol/L, 0.17 mmol/L, respectively, and the average probability of being in Zone A of the Clark error grid is 100.00 %. Additionally, we demonstrate with the Bland and Altman plot that our proposed model has the highest consistency with portable blood glucose meter detection method.