Novel Methods For Spectral Data Quality Assessment And Wavelength Selection In Dynamic Spectrum

Noninvasive measurement blood component has become a hotspot in recent years.Spectral analysis method is one of the most promising orientation.Dynamic spectrum(DS)based on photoplethysmography(PPG)have drawn more and more attentions because it can eliminate the influence of individual difference and background tissues.In DS method,quality of spectral data is crucial to the precision of model.Therefore,screen high quality data is a priority in DS.A good method for spectral data quality assessment can improve the performance of DS.Wavelength selection has important implications for NIR spectralanalysis.DS method has made considerable development,but there are little research about wavelength selection in DS.Therefore,this dissertation propose a novel method for wavelength selection in noninvasive hemoglobin measurement by DS,and simplify the model.Based on the elaboration of the principle of DS,this dissertation mainly include following three parts.Firstly,we designed and made a light source system with light-compensation LED.Then we conducted the clinical experiments and collected a large amount of data,based on which we made the subsequent study and data analysis.Secondly,we start from the stability of PPG signal,and proposed a new method for spectral data quality assessment.After simulation,we analysis the data set of 427 subjects.Then we seperated the samples into two groups and established the model by artificial neural network(ANN)between the DS and HB concerntration.The results show that the correlation coefficient of prediction is 0.875,while that of the unscreened data is only 0.715.The results demonstrate that it can really improve the performance of modle,by screening the data according to the stability coefficient.The third aspect is about an wavelength selection method based on variable importance in projection(VIP).We screened out the independent variables with higher importance by analysis the explanation power of each independent variable in the PLS model,and remove the wavelengths with lower explaination power,therefore,we can simplified the model and reduce the computational cost.We analysis the hemoglobin concentration value of 232 volunteers.We reduced the wavelenths number from 586 to 64.After screening the wavelength,the mean relative error(MRE)of the predction set is 1.82%,which was a satisfactory result obtained from a small number of independent vaviables.Then we demonstrated the screened wavelengths' explaination power by bootstrap method.Meanwhile,this study has pointed out the sensitive wavelengths in DS method for the first time.Wavelength selection based on the variable importance in projection,which makes an important step to the practical use of DS,offers valuable references and fresh ideas to the research of related fields.