Classification of diabetes and measurement of blood glucose concentration noninvasively using near infrared spectroscopy

Developing noninvasive blood glucose monitoring method is an to immense need to alleviate the pain and suffering of diabetics associated with the frequent pricking of skin for taking blood sample. A hybrid algorithm for multivariate calibration is proposed to improve the prediction performance of classification of diabetes and measurement of blood glucose concentration by near infrared (NIR) spectroscopy noninvasively. The algorithm is based on wavelet prism modified uninformative variable elimination approach (WP-mUVE) combined with least squares support vector machine (LSSVM), named as WP-mUVE-LSSVM. The method is successfully applied to diabetic classification experiment (in vivo) and blood glucose concentration measurement experiment (in vivo) respectively. Human tongue is selected as the measuring site in this study. To evaluate effectiveness of pretreatment method and quality of calibration models, several usually used pretreatment methods and kernel functions of LSSVM are introduced comparing with our method. Higher quality data is obtained by our pretreatment method owing to the elimination of varying background and noise of spectra data simultaneously. Better prediction accuracy and adaptability are obtained by LSSVM model with radial basis kernel function. The results indicate that WP-mUVE-LSSVM holds promise for the classification of diabetes and measurement of blood glucose concentration noninvasively based on human tongue using NIR spectroscopy.