3.1. Results with MW-PLS
The NIR spectra of 230 human serum samples in the scanning area (780-2498 nm) were shown in Figure 1. As can be seen from the figure, absorption at about 2000 nm and 2400 nm has obviously strong noise. In order to obtain satisfactory results, wavelength selection must be carried out to overcome noise interference. For comparison, PLS model of the full spectrum region was first established. The corresponding SECV and RP,CV were 1.423 g L-1 and 0.935, respectively.
MW-PLS method was performed to optimize waveband and improve prediction accuracy. Depending on minimum SECV value, the optimal MW-PLS model was selected out. The corresponding waveband was 1504 to 1820 nm, located in the long-NIR region (1100 to 2498 nm). Prediction effects (SECV and RP,CV) and parameters of the above two methods were summarized in Table 2. The results showed that the predicted values were highly correlated with clinical measurements for the two methods, and comparing with optimal PLS model in full spectrum region, the optimal MW-PLS model achieved better prediction effect with fewer wavelengths.
3.2. Results withOPWC-PLS
The OPWC method was performed for screening information wavelength based on the steps mentioned in section 2.4. Firstly, 104 best partners for all 860 wavelengths were determined according to the results of LOOCV-BLR analysis, and PWS(1) with 104 wavelengths was obtained. Thus, the number of wavelengths was greatly reduced after the first projection. The correspondence between all 860 wavelengths and their best partners was shown in Figure 2. As shown in the figure, some wavelengths had the same best partner, such as the 2156 nm and 2190 nm as best partners of other wavelengths appeared 3 and 8 times, respectively, so projection was not a one-to-one mapping function in the whole spectral region . Obviously, was a subset of and the projection continues.
Based on the corresponding relationship determined above, the best partner of was easy to be selected, and the PWS(2) was obtained. Repeated the same process for PWS(2), and PWS(3) was obtained. As the projection progresses, the number of wavelengths decreased gradually until the number of wavelengths for PWS(6) no longer changed. The PWS(6) was the OPWC and it had only 28 wavelengths. Figure 3 showed the 28 wavelengths and their best partners. As the figure showed, the 28 wavelengths are divided into 14 groups, and the two wavelengths in each group are the best partners for each other.
Based on PLS, the LOOCVs were performed for every PWS, and the corresponding minimum SECV value and number of wavelengths (N(s)) used are shown in Figure 4. As shown in the figure, the N(s) and minimum SECV values have almost the same trend. After the first projection, both of them decrease rapidly, and the remaining wavelengths are more important, so as the number of projections increases, they slowly decrease. This may be due to the removal of a large amount of noise and background information from the original spectrum after the first projection, so both the N(s) and minimum SECV values decrease rapidly. The partner wavelength subset of the original spectrum contains less redundant information, so the N(s) and minimum SECV values decrease slowly in the later projection iteration.
3.3 Comparison of the two methods
Screening the information wavelengths of GLB in the human serum of a multi-component complex system is difficult and complicated. The wavelengths selected by the OPWC-PLS and MW-PLS methods, which correspond to the information of GLB, were shown in Figure 5. As indicated in Figure 5, the wavelengths selected by the OPWC method have a wider distribution range and partially coincides with the wavelengths selected by MW-PLS. This may be because the local characteristics of MW-PLS method make some wavelengths cannot be detected, which reflects the complexity of NIR model optimization and the commonness and difference of different methods.
Figure 6 showed the relationship between the predicted and measured GLB values based on the MW-PLS and OPWC-PLS methods, respectively. The prediction effect and corresponding parameters N and F were summarized in Table 2. The SECV and RP,CV were 0.813 g L-1 and 0.978 with OPWC-PLS, and 0.804 g L-1 and 0.979 with MW-PLS, respectively. The results show that, like MW-PLS, the prediction effect of OPWC-PLS was also obviously better than that of the whole spectrum PLS, and the OPWC is an effective method for screening wavelengths. The phenomenon conveys that better prediction results can be achieved with fewer wavelengths. Thus one can conclude that it is very necessary to first perform wavelength selection before building a calibration model. The two methods had achieved almost the same good prediction results (SECV and RP,CV). However, the optimal OPWC-PLS model adopted only 28 wavelengths, while the other adopted 159 wavelengths. Therefore, the OPWC method has great prediction performance for wavelength selection.
The differences in prediction of the OPWC-PLS and MW-PLS methods for GLB illustrate that MW-PLS can achieve higher prediction accuracy, but it is time-consuming and employs more wavelengths, while OPWC-PLS can achieve similar prediction results with MW-PLS in less time。In addtion, MW-PLS, as a continuous wavelength screening method, is more suitable for determining the object with relatively concentrated molecular absorption bands; while OPWC-PLS, as a discrete wavelength screening method, may be more suitable for determining the object with relatively fragmented molecular absorption bands.