Study Of Outliers Determination In Non-invasive Blood Glucose Sensing With Near-infrared Spectroscopy

Noninvasive blood glucose sensing based on near-infrared(NIR)spectroscopy is highly sensitive to numerous interference factors such as secondary blood components,physiological condition in human bodies,environment of measurement and measuring device.The considerable existence of interfering information eliminate the specificity of glucose in the spectra,making spectra outliers and resulting in undesirable performance of the accuracy and robustness of the model.Concerning the specificity expression of glucose variation,statistical patterns of spectra and the fact that concentration of glucose changes in a certain range,this paper proposes three outlier determination methods based on isosbestic wavelength and orthogonal distance with elaboration of the corresponding verification experiments.Statistical patterns of differential absorbance are analyzed under the situation of single wavelength and the whole wave band.An outliers determination method based on differential absorbance at isosbestic wavelength is proposed with isosbestic wavelength caused by displacement effect of glucose and water used as glucose specificity expression.Another outliers determination method based on orthogonal distance is proposed with the first principal component of glucose variation resulting from robust principal component analysis used as glucose specificity expression.Meanwhile,simplified orthogonal method of outliers’ determination with robust analysis replaced by direct measurement of high-concentration sample is also proposed in order to shorten the time span of measurement process and reduce the influence of experimental environment variation and the system error of measuring device.NIR transmission spectra are measured under normal,temperature-affected and shift-affected conditions with glucose aqueous solutions and 2% Intralipid phantom solutions used as measured objects.Two-dimensional correlation spectroscopy and the proposed three outliers’ determination methods are applied to analyze the character of outliers and determine outliers separately.Results show that strong and characterized autopeaks can be observed in the asynchronous two-dimensional correlation spectra of glucose aqueous samples,but character of outliers of phantom samples are no longer obvious in two-dimensional correlation spectra.The method based on isosbestic wavelength is able to determine outliers in aqueous solution samples.However,necessity of the existence of isosbestic wavelength in the adopted measuring wave band leads to the impossibility of the method applied in outliers determination of the phantom samples.Both orthogonal distance method and simplified orthogonal distance method are able to detect outliers correctly in aqueous and phantom solution samples.The simplified orthogonal distance method needs less time for measurement,hence reduce the potential influence of change of measurement condition upon the specificity expression of spectra.Therefore,simplified orthogonal distance method is more practical than the other two methods.