1. College of Biomedical Engineering, South-Central Minzu University, Wuhan 430074, China
2. Key Laboratory of Cognitive Science, State Ethnic Affairs Commission, Wuhan 430074, China
3. Hubei Key Laboratory of Medical Information Analysis and Tumor Diagnosis ＆ Treatment, Wuhan 430074, China
Abstract：Hemoglobin is an important physiological index of the human body. Abnormal concentrations of Hemoglobin will lead to various diseases. Infrared spectroscopy has the advantages of simplicity, non-destructive and rapidity. It is very suitable for the quantitative analysis of physiological parameters. However, the spectral background is complex, and the effective information is weak. How extract effective feature variables and build an accurate quantitative model is a difficult problem. To solve this problem, study the spectral data of blood samples and hemoglobin imitation solution samples, and through modeling and comparison, the best data set division method is SPXY by using SPXY method, K_S method, duplex method and equal interval division method to divide the data. Three data pre-processing methods of Savitzky Golay first-order derivative filter (S_G1) + wavelet transform, wavelet transform+S _G1 and standard normal variable transform (SNV)+S_G1 are traversed, and the best pre-processing method is SNV+S_G1. Combined with the series idea, the characteristic wavelength optimization method of combining Synery interval Patial Least Squares (SiPLS) and Successive Projections Algorithm (SPA) in series is proposed, and so the SiPLS-SPA-PLS prediction model is constructed. The model is verified with two data sets, and the advantages and disadvantages are judged according to the evaluation indexes and compared to the three quantitative models of full spectrum PLS, SPA-PLS and SiPLS. The experimental results show that: (1) using SiPLS-SPA-PLS for quantitative analysis, the values of RC, RP, RMSEC and RMSEP of blood samples are 0.993 6, 0.990 6, 0.199 2 and 0.184 6 respectively, and the values of RC, RP, RMSEC and RMSEP of imitation solution samples are 0.998 9, 0.998 5, 1.848 9 and 2.007 4 respectively. Compared with the three quantitative models of full spectrum PLS, SPA-PLS and SiPLS, the SiPLS-SPA-PLS model is the best. Because the values of RC and RP are the largest and the values of RMSEC and RMSEP are the smallest. This model can realize the quantitative analysis of hemoglobin more accurately. (2) The SiPLS-SPA-PLS quantitative model can screen the optimal wave band more accurately. The effective wave bands screened by the two samples are blood (1 144~1 264, 1 606~1 798 nm) and imitation solution (1 018~1 390, 1 600~1 700 nm). The influencing factors of the instrument are roughly the same. This method can accurately optimize the characteristic wavelength. (3) The model can extract effective variables, remove the influence of useless noise, select 28 spectral variables from 700 blood samples and 41 spectral variables from 1 201 hemoglobin imitation solution samples to improve the detection speed and prediction efficiency. In short, this method provides an idea for rapid and accurate detection of hemoglobin.
