Random Forests For Non-invasive Blood Glucose Sensing With NIR And The Design Of Experimental System

As the development of economy and the lifestyle changes, diabetes have beenbecome the third severe chronic just behind the tumor and the cardiovasculardisease with the aggravation of the aging society, changes in dietary patterns, theincrease in obesity, and so on. Patients have to use the invasive blood glucosesensing to monitor the blood glucose data every day, strictly, which not only bringpsychological and physical pain sustained, but also cause infection if not carefulenough. In recent years, non-invasive blood glucose monitoring technology isdeveloping rapidly to satisfy the demand for real-time monitoring of blood glucosedata painless, and the near infrared spectroscopy method is the most promising andpopular method for non-invasive blood glucose measurements. However, due to theimpact of human skin and blood of the complexity of individual-specific, theexternal environment, and the noise instrument interface, Lambert-Beer law of thelinear relationship was severely disrupted, which results the traditionalstoichiometric algorithm based on the principle guiding, such as multiple linearregression, principal component regression, partial least squares, etc. areunderperforming in the actual testing. In2001, Leo Breiman proposed a data-basedmethod called Random Forests, which particularly suitable for the unclear prioritheory and the complex experimental environment. It is a hot research field ofstatistics in recent years which is already widely used and achieved excellent resultsin many fields such as the selection of feature gene, sites of protein binding, facerecognition, chlorophyll content, etc. This thesis is trying to introduce the RandomForests into the areas of non-invasive blood glucose sensing. We use20%Intralipidliquid to build a Monte-Carlo model, which represents the light propagation withdifferent concentrations of glucose, to get different concentrations of glucosediffuse reflectance spectral data array. Then create a RF model by setting the appropriate number of random split tree and the number of attributes, and analysiswhich part of bands are more important for the model. We verify the feasibility ofRandom Forests applied into the NIR non-invasive blood glucose sensing. Wecombine the theoretical requirements with the actual conditions designe aNon-invasive Blood Glucose Sensing with NIR system in the end of the thesis.