1. School of Mathematics and Statistics, Changchun University of Science and Technology, Changchun 130022, China
2. School of Physics, Changchun University of Science and Technology, Changchun 130022, China
3. Mathematical Experiment Demonstration Center of Changchun University of Science and Technology, Changchun 130022,China
Abstract:Near-infrared spectroscopy analysis technology has broad application prospects in biomedical engineering. Non-invasive and continuous measurement can monitor the human blood glucose level in real-time, which brings great convenience to diabetes patients, improves the quality of life of patients, and reduces the incidence of complications of diabetes. The idea of non-invasive blood glucose monitoring was put forward earlier, but there are still difficulties, such as low prediction accuracy low correlation between prediction value and label value: up to now, it has not met the clinical requirements. In recent years, spectral detection technology has developed rapidly, and machine learning technology has obvious advantages in intelligent information processing. Combining the two can effectively improve the accuracy and universality of non-invasive blood glucose medical monitoring models. This paper proposes a label sensitivity algorithm (LS), and a prediction model of human blood glucose content is established by combining the support vector machine method. We used a near-infrared spectrometer to collect dynamic blood spectral data at the index finger of four volunteers (28 groups of data for each volunteer) and used the multivariate scattering correction (MSC) method to eliminate the influence of partial light scattering. Considering the difference in the absorption of blood glucose to light of different wavelengths, In this paper, a feature wavelength selection method based on blood glucose concentration label difference is proposed, and a label sensitivity support vector machine (LSSVR) prediction model is constructed Experiments were designed to compare the model with partial least squares regression (PLSR) and discriminant support vector machine (FSSVR, The predicted values are all in the A-region of Clark grid with allowable error. The excellent performance of the LSSVR model provides a new idea for the early realization of non-invasive blood glucose monitoring.