Research On Quantitative Model Of Hemoglobin Based On Near-Infrared Spectroscopy

Hemoglobin concentration is an important indicator for evaluating anemia in humans,which can be used as an important basis for clinical diagnosis of anemia.Clinical medicine uses biochemical analysis to detect the hemoglobin concentration of patients,but biochemical analysis not only has high cost,but also a wound which will increase the risk of infection,and can not be continuously monitored.NIR spectroscopy,as an emerging detection technology,has the advantages of non-invasive and continuous measurement,which made certain progress in the application of hemoglobin non-invasive detection.However,due to the large amount of spectral data,it directly affects the modeling speed and prediction accuracy.For this reason,in-depth study of spectral data processing methods and improving the speed and accuracy of spectral data modeling are the focus of current near-infrared spectroscopy quantitative analysis.In this paper,spectral analysis techniques,chemometrics and other methods are used to further study and discuss the data processing methods such as spectral preprocessing,modeling and characteristic wavelength selection in hemoglobin near-infrared spectroscopy quantitative analysis.The main research contents and conclusions as follows:The ability to predict the hemoglobin content of a single background by Partial Least Squares(PLS)model and BP neural network model was studied.The optimal PLS quantitative model is the SG_PLS model established by Savitzky-Golay smoothing preprocessing.The correction set and prediction set determination coefficients Rc and Rp of the model are 0.9999 and 0.9998,respectively.The RMSEC and RMSEP of the calibration set and prediction set are 0.4845 and 1.0727 respectively.The optimal BP quantitative model was obtained when the number of hidden layer neurons was 8 and the learning rate was 0.1.The Rc,Rp,RMSEC and RMSEP of the model were 0.9948,0.9843,1.086 and 7.9619.The results demonstrate that both models are able to accurately predict the hemoglobin content of a single background.The results of the PLS quantitative model and the BP quantitative model to predict the hemoglobin content of complex backgrounds was compared.The Rc,Rp,RMSEC,and RMSEP of the PLS model were 0.4884,0.4749,10.0128,and 9.3652.The Rc,Rp,RMSEC,and RMSEP of the BP model were 0.9999,0.9649,0.2213,and 3.0041.From the results,the prediction results of BP neural network are better for the prediction of complex hemoglobin.The PLS model and BP model in a single background were used to predict the hemoglobin of complex background and the model was modified.When the number of samples in the modified model reaches a certain proportion(30%),the Rp of the BP modified model reaches 0.9303,and the Rp of the PLS modified model reaches 0.6362.The results show that the updated model is more accurate for the prediction of complex background hemoglobin.The updated BP quantitative model has better prediction results and can meet the needs of complex background hemoglobin quantitative analysis.The characteristic wavelength selection method of hemoglobin quantitative analysis model was studied.The ability of equal interval partial least squares(iPLS)and principal component analysis(PCA)to select the characteristic wavelengths of hemoglobin spectra was compared.The results show that for the PLS model,the quantitative model established by the preferred wavelength of iPLS is determined to be optimal.The Rc,Rp,RMSEC and RMSEP of the model are 0.9994,0.9968,1.5223 and 3.4841.For the BP model,the iPLS_BP model established by the wavelengths selected by iPLS is also optimal.The Rc,Rp,RMSEC,and RMSEP of the model are 0.9996,0.9958,1.444,and 7.630,respectively.It is shown that the equidistant partial least squares method can effectively screen out the characteristic spectrum,simplify the model and improve the prediction accuracy.In summary,this paper studies the difference in predictive performance between PLS and BP neural network in hemoglobin quantitative analysis.For complex samples with individual differences,The prediction result of the BP model is more accurate.This study provides a solution to the problem of environmental interference and personal differences in hemoglobin measurement.