Non-invasive blood glucose detection by using multi-wavelength optical imaging

The improvement of human living standards has led to the increasing number of patients with diabetes, which makes it necessary to use inexpensive and convenient continuous non-invasive blood glucose detection technology. In this paper, we developed a multi parameter fusion optical imaging non-invasive blood glucose monitoring system based on multi wavelength photoplethysmography and denoising autoencoder neural network. This paper selected four detection bands, 625nm, 740nm, 850nm, and 940nm, to characterize the concentration information of glucose, melanin, fat, and other substances in the human body at different wavelengths. This paper used a microcontroller to synchronize the imaging CCD device with the light source, sequentially capture images under four wavelength light sources, and obtain IPPG signals of multiple wavelengths through image processing techniques, eliminating signal artifacts and baseline drift of IPPG signals. After processing the IPPG signal, 48 time-domain features including KTE and logE characteristics were obtained. This paper used calculated signal features as temporal features, images as spatial features, and a combination of IPPG signals generated at four wavelengths as spectral domain features. A robust blood glucose prediction model was established using denoising autoencoder neural networks and Late Fusion multimodal fusion techniques. Clarke Error Grid Analysis (EGA) showed that 91.43% of predicted blood glucose values were within the range of A, indicating that the predicted blood glucose values obtained through our method are acceptable in clinical practice, providing technical support for the development of video based optical imaging non-invasive blood glucose detection technology.