Non-invasive blood glucose measurement based on visible light and embedded system

Blood glucose is an important physiological parameter. Regular evaluation of blood glucose level is of great significance for the monitoring and treatment of diabetes mellitus and its complications. Noninvasive detection is the ideal technology to achieve periodic blood glucose assessment, among which optical measurement method is the current research hotspot, but due to low SNR and low accuracy, optical noninvasive blood glucose detection method cannot be used in clinic. To solve the problems above, we designed a device for non-invasive blood glucose detection based on visible light and embedded system in this paper. We used the visible light source with a wavelength of 625nm and a high-quality camera to obtain the scattered image information of fingertips in a dark environment, and then feature vectors and dimensionality would be extracted from the original image by the method of Convolution Auto-Encode (CAE). Then, we used the theoryoriented method partial least squares regression (PLSR) and the data-oriented method gradient boosting regression (GBR) to establish the correlation model of the relationship between the scattering image feature vectors and blood glucose level, and the performance of the models were verified by the test set. We divided the prediction of blood glucose into three categories: less than 6.1, 6.1 to 7.8, and greater than 7.8, and the models would be trained to classify the blood glucose, and experiments show that the GBR performs better than PLSR, the accuracy of GBR in test set is up to 80 percent while the blood glucose greater than 7.8. Finally, the device is highly integrated centered on embedded system, and GBR has the advantages such as high precision, low cost, simple and convenient to use, which has great application value for the research of non-invasive blood glucose measurement.