Study Of Chance Correlation And Wavelength Selection In Non-invasive Blood Glucose Sensing By Near-infrared Spectroscopy

Diabetes and its complications are great menaces to human's health, and its incidence is growing fast. Noninvasive blood glucose sensing can reduce the inconvenience and pain in patients'daily monitoring. Near-infrared spectroscopy is one of the most challenging methods to measure the blood glucose non-invasively. However, the structure of near-infrared spectra is complicated and the spectral variation from glucose is quite weak. And there have multiple correlations among the wavelengths. All these issues restrict the measurement accuracy of glucose concentration. In this paper, the chance correlation and wavelength selection in the glucose sensing are investigated.In the noninvasive blood glucose sensing by the near-infrared spectroscopy, chemometrics is applied to achieve the quantitative analysis of unknown samples. However, a certain degree of chance correlation will be usually introduced in the modeling and validation process, which will affect the stability of model. In this paper, normally distributed random numbers are used to simulate spectral data and reference concentration to investigate the probability level of chance correlation from the number of selected wavelengths and different cross validation methods. Results show that, chance correlations exist indeed in the process of modeling, and the proper number of wavelengths and the optimal cross validation method are chosen to reduce the chance correlation.As the body temperature and other physiological factors'changes will also lead to changes in near-infrared spectra, in vitro experiments of glucose aqueous solutions in different temperatures are conducted. In the experiments, the relationship between temperature and glucose concentration is obtained, which can be used to reduce the temperature effect in practice.Furthermore, the spectra overlap in near-infrared region is serious, which results in poor sensitivity of some wavelengths. In order to improve the prediction accuracy of the model, genetic algorithm and its modified algorithm are used to optimize the body's blood glucose spectrum. Results show that, it's feasible to apply the genetic algorithm in improving the model and the wavelengths selected by genetic algorithm are concentrated in the vicinity of the absorption peaks of glucose, which is consistent with physical expectations. In PLS model, the root mean standard error of prediction can be reduced by 11% and 16% by the traditional genetic algorithm and dynamic genetic algorithm, respectively.In conclusion, the prediction accuracy can be improved by reducing chance correlation and selecting wavelengths, and it is critical for the success of non-invasive blood glucose measurement.