**Research on Key Methods of Effective Spectra Signal Extraction in Near Infrared Non-Invasive Blood Glucose Measurement**

Author:Li Xiao Li

Supervisor:li cheng wei

Database:Doctor

Degree Year:2019

Download:46

Pages:116Size:2247K

Keyword:entropy，MIC-PCA，Mode decomposition，Near infrared spectra

Diabetes is a diseases caused by insulin secretion defect or its biological function disorder and characterized by hyperglycemia,associated with energy metabolism disorder.Diabetic complications involve each organs and tissue of whole body accompanied with pathological changes,which lead sick and death.The diabetes seriously affect the life and life quality of patients.Blood glucose concentration is an important index for the initial diagnosis of diabetes.Blood glucose concentration detection is an indispensable key step in the treatment of diabetes.At present,the blood glucose concentration is usually detected by venous blood or minimally invasive fingertips with enzyme catalysis,which can not realize continuous real-time blood glucose detection and increase the pain and economic burden for patients.Therefore,it is of great significance to study the related technologies of non-invasive blood glucose detection.The near infrared non-invasive blood glucose detection become one of the research focus in the field of non-invasive blood glucose detection.The near infrared spectra can penetrate the skin and body tissue,and the blood glucose concentration and its near infrared absorption has good linear correlation.But the weak signal of glucose concentration and the large dimension of near-infrared spectrum data increase the difficulty of near infrared non-invasive blood glucose detection.In this paper,the related technical difficulties of near infrared non-invasive blood glucose detection method were discussed.Furthermore,the aims of this paper are listed following,First,to explore a kind of adaptive signal denoising method that slove the problem of poor accuracy of glucose concentration prediction model caused by noise interference;Moreover,to research a kind of effective method of spectral variable selection that slove the problem that the detection accuracy of the prediction model becomes low due to the model building with near-infrared full spectrum data of glucose;Last but not least,to exlpore a kind of effective feature extraction method of spectral variables to solve the problem that the nonlinear correlation between spectral data makes the establishment of complex model.For these purposes,the main research contents of the paper are as follows:In order to solve the problem that noise exists in the near infrared spectra data in the process of near infrared non-invasive blood glucose detection,which affects the accuracy of the glucose concentration prediction model,this paper proposed a new near infrared signal denoising method based on the adaptive noise ensemble empirical mode decomposition(CEEMDAN)and discrete Frechet distance evaluation criteria.First,the paper outlines the related theory of near infrared spectroscopy detection.Based on CEEMDAN and discrete freich distance theory,the adaptive separation of noise and signal is realized according to the frequency difference of noise and signal in multi-intrinsic function components(IMF)obtained after decomposition of near-infrared spectral signals.Then,the IMFs are classified by using the discrete Frechet distance,to determine the IMF dominated by the interfering noise and the useful signal respectively.Finally,the useful signal dominated IMF was summed up and the near infrared spectral signal was reconstructed,which reduce the interfering noise unrelated to the glucose concentration information.The experimental results show that the denoised comprehensive evaluated index was increased by 0.1 at most when the method was applied to the near-infrared spectroscopy of solutions with different glucose concentrations.The denoised comprehensive evaluated index was increased by 0.09 at most when the method was applied to 2% Intralipid simulation of different concentration of glucose solution.The presented method is compared with traditional spectra denoising method,the result of the reconstructed signal have high precision and stable filtering performance which realized the separation of signal and noise,and had good adaptability.In order to solve the problem of lack of correlation between some spectral bands and glucose concentration and low accuracy of the whole spectral data modeling,a method of near infrared spectral variables selection based on permutation entropy was proposed.By analyzing the physical basis and necessity of spectral variable selection,the proposed method that splitting the denoised near-infrared full spectrum data by sliding window,the permutation entropy of each spectral wavelength range was calculated,and the permutation entropy was used as the characteristic parameter to select the spectral variable.The experimental results show that the new method is applied to different concentration of glucose in vitro solution and 2% Intralipid simulation of different concentration of glucose solution,compared with traditional spectral variable feature selection methods,the new method based on permutation entropy selected variables were reduced to 13.9% and 14.7% of the full spectrum variables,that far small all over 1867 wavelengths spectral data.When the method was applied to the near-infrared spectroscopy of solutions with different glucose concentrations,the correlation coefficient of SVR modeling was increased by 10.1% at most and the prediction root mean square error was reduced by 13.7% at most.The correlation coefficient of PLSR modeling was increased by 18.5% at most and the prediction root mean square error was reduced by 4.2% at most.When the method was applied to 2% Intralipid simulation of different concentration of glucose solution,the correlation coefficient of SVR modeling was increased by 13.4% at most and the prediction root mean square error was reduced by 12.4% at most.The correlation coefficient of PLSR modeling was increased by 14.3% at most and the prediction root mean square error was reduced by 10.8% at most.The proposed method can select spectral variables related to glucose concentration information,reduce the influence of unrelated spectral variables on model prediction ability,and effectively improve the problem of low prediction accuracy of full-spectrum data modeling.To solve the problem of nonlinear correlation between spectral data in near infrared non-invasive blood glucose concentration detection,a principal component analysis method based on maximum information coefficient(MIC)for feature extraction of near-infrared spectral variables was proposed.By analyzing the theoretical model of traditional principal component analysis method,the concept of maximum information coefficient is introduced according to the characteristics of near infrared spectral data.The new method first combines the traditional principal component analysis model and MIC related theory to conduct standardized processing on near-infrared spectral data and calculate the maximum information coefficient between features.Then,the MIC matrix was used to calculate the eigenvalues and eigenvectors,and the principal components were selected by threshold method to extract the features of near infrared spectral variables.The experimental results show that the principal component number of solutions with different glucose concentrations is 9.For the near-infrared spectral data of 2% Intralipid simulation of different concentration of glucose solution,the principal component number was 8.The experimental results show that the correlation coefficient of PSO-SVR modeling was increased by 10.8% at most and the prediction root mean square error was reduced by 15.1% at most,when the method was applied to the near-infrared spectroscopy of solutions with different glucose concentrations.The correlation coefficient of PSO-SVR modeling was increased by 9.4% at most and the prediction root mean square error was reduced by 17% at most,when the method was applied to 2% Intralipid simulation of different concentration of glucose solution.It is shown that the principal components extracted by the new method could effectively reflect the spectral information of glucose,reduce the complexity of the model and improve the prediction ability of the model.