This study demonstrated particle size effect on the measurement of saikosaponin A in
Near infrared (NIR) reflectance spectroscopy is widely used for quality assessment of solid sample in areas of pharmaceuticals, agriculture, food, fruits, forage, and so on due to its rapid measuring speed, flexibility, and less or even no sample preparation [
However, for sample presentation of CHM, different particle sizes affected sample homogeneity, sample packing density, and sample surface, which all introduced uncontrolled variations that brought forth difference in optical path length and multiplicative light scattering effects [
In addition, the fact that sample presentation to the instrument (e.g., particle size) has been found to affect the characteristics of NIR spectra should be paid great attention, thus determining the robustness and accuracy of NIR as analytical technique. According to the effect of soil particle size (SPS) on the NIR measurement of exchangeable sodium (Na), NIR accuracy for soils with great particle sizes (SPS-0.212, 0.212 mm) was higher than soil with small particle sizes (SPS-0.053, 0.053 mm) [
Therefore, how to guarantee low noise and good NIR model performance with different granularity effect was worth clarification. Researches concerning this issue have done limited work to give conduction in CHM. David reported a method for quantifying the median particle size of a dry powder using preprocessing NIR spectra. A quadratic model was developed to explain these summations as a function of median particle size, since the effect of densification was minimal [
Nevertheless, this is only one paper on the particle size of CHM in NIR measurement, which illuminates the influence of granularity on NIR spectra characteristic of
All
A summary of tested samples.
Sample number | Origins | Growth pattern |
---|---|---|
1~5 | Shanxi | Unknown |
6~9 | Shanxi | Unknown |
10~14 | Shanxi | Cultivated |
15~19 | Shanxi | Wild |
20~25 | Shanxi | Wild |
25~30 | Hebei | Cultivated |
After being cleaned by brushing off soil dust from the surface,
About 1 g sample powder was packed into the sample cup. NIR spectra were acquired in reflectance mode with the Integrating-Sphere module of the Antaris I FT-NIR analyzer (Thermo Fisher, USA). Each spectrum was the average of 64 successive scans with air as the background. The spectral range was 10000–4000 cm−1 with 1.928 cm−1 data interval. To guarantee the analysis accuracy, each sample was analyzed in triplicate and the mean value of three spectra was used in the following analysis. To avoid the effects of environment condition in the laboratory, such as temperature and humidity, the room temperature was controlled at 25°C, and the humidity was kept at an ambient level.
The reference method used for SSA determination was the high performance liquid chromatography (HPLC) assay recommended by the Chinese Pharmacopoeia (ChP, 2010 Edition) for
Elution gradient used in the HPLC method.
Time/min | ACN (v/v) | Water (v/v) |
---|---|---|
0–50 | 25–90 | 75–10 |
50–55 | 90 | 10 |
55–60 | 25 | 75 |
60–67 | 25 | 75 |
All the computations were performed using TQ Analyst software package (version 8.0, Thermo Scientific, Madison, USA). Other data analyses were performed by Unscrambler 9.7 software package (Camo Software AS, Norway) and MATLAB version 7.0 (MathWorks Inc., USA). Some of the algorithms used in this paper were developed by us.
Figure
The chromatograms of
Figure
(a) Raw spectra of samples with different granularity. (b) Difference of NIR frequency range to the granularity.
Former research demonstrated that
It could be observed that
To avoid bias in sample selection, the Kennard-Stone (KS) algorithm was used to split the NIR data set into calibration and validation. Twenty concentration levels including 60 samples were used as the calibration set, and the remaining samples were the validation set, which was shown in Table
Concentration range of SSA in calibration and validation set (mg⋅g−1).
Sample set | Numbers | Concentration range | Average | Standard deviation |
---|---|---|---|---|
Calibration | 60 | 1.476–8.162 | 3.695 | 1.452 |
Validation | 30 | 1.601–5.807 | 3.727 | 1.269 |
Data-preprocessing techniques were investigated prior to calibration development. To optimize the spectra, the empirical multiplicative light scattering correction method, MSC, and SNV were applied. Then combination of derivative methods including first derivative (1D) and second derivative (2D) was used to reduce baseline variations observed in original diffuse reflectance spectra and to enhance spectral features. Meanwhile, smoothing methods including Savitzky-Golay smoothing filter (SG) and Norris derivative filter (ND) were employed to depress the background noise amplified by derivative. The optimum preprocessing method was determined by the lowest PRESS value (Figure
Plot of PRESS value against latent factors.
After application of the best data pretreatments, four local PLS models were constructed with powder samples, which were screened through 40-, 65-, 80-, and 100-mesh sieve separately. To compare the prediction performance of every local model, test-set validation was performed and the result was shown in Table
Local model performance of different granularity.
Model | LVs | Cross validation | Test-set validation | RPD | ||
---|---|---|---|---|---|---|
RMSECV |
|
RMSEP |
| |||
40 | 4 | 0.682 | 0.7671 | 0.650 | 0.8519 | 1.95 |
65 | 4 | 0.574 | 0.8347 | 0.492 | 0.9221 | 2.58 |
80 | 3 | 0.567 | 0.8408 | 0.534 | 0.9070 | 2.38 |
100 | 3 | 0.664 | 0.7484 | 0.522 | 0.9162 | 2.43 |
Correlation diagrams between the NIR predicted values and the reference values of SSA content.
It was significantly found that local model performance was not gradually increasing with decreased granularity. The result demonstrated that model performance went down at 65 mesh and tended to be steady from 80 mesh to 100 mesh. The result showed that granularity and sample heterogeneity were both critical for NIR analysis. When grinding the solid sample, sample granularity should be considered. Furthermore, the local model was not very perfect though its correlation coefficient was greater than 0.9. To further improve model performance, granularity-hybrid calibration model was tried in the next section.
To develop a robust calibration model and realize model’s successful application, another way to defend variations of particle sizes is to construct a granularity-hybrid calibration model (GH model), including calibration set of every granularity (240 samples, 40, 65, 80, and 100 mesh). Then validation sets of every particle size were predicted by the GH model, as shown in Figure
GH model performance constructed with different data-preprocessing methods was exhibited in Table
Prediction performance of GH model.
Pretreatment methods | LVs | Cross validation |
Test-set validation ( |
Test-set validation |
Test-set validation |
Test-set validation | |||||
---|---|---|---|---|---|---|---|---|---|---|---|
RMSECV |
|
RMSEP |
|
RMSEP |
|
RMSEP |
|
RMSEP |
| ||
RAW | 5 | 0.763 | 0.6903 | 0.814 | 0.7657 | 0.747 | 0.8031 | 0.707 | 0.8292 | 0.635 | 0.8862 |
MSC | 2 | 0.716 | 0.7278 | 0.716 | 0.8293 | 0.654 | 0.8551 | 0.621 | 0.8628 | 0.538 | 0.9123 |
1D + SG | 7 | 0.602 | 0.8229 | 0.687 | 0.8420 | 0.621 | 0.8473 | 0.566 | 0.8909 | 0.566 | 0.8929 |
2D + SG | 4 | 0.624 | 0.8025 | 0.671 | 0.8621 | 0.596 | 0.8815 | 0.612 | 0.8651 | 0.540 | 0.9065 |
MSC + 1D + SG | 6 | 0.606 | 0.8276 | 0.575 | 0.8971 | 0.481 | 0.9279 | 0.524 | 0.9137 | 0.545 | 0.9146 |
MSC + 2D + SG | 3 | 0.678 | 0.7621 | 0.690 | 0.8538 | 0.664 | 0.8552 | 0.618 | 0.8648 | 0.522 | 0.9126 |
MSC + 1D + ND | 6 | 0.609 | 0.8821 | 0.672 | 0.8508 | 0.580 | 0.8873 | 0.631 | 0.8655 | 0.765 | 0.8371 |
MSC + 2D + ND | 6 | 0.566 | 0.8450 | 0.621 | 0.8757 | 0.527 | 0.9099 | 0.583 | 0.8863 | 0.603 | 0.8879 |
Correlation diagrams of GH model.
Effects of granularity on NIR were investigated; the results concluded that influence on NIR spectra was wavelength dependent. NIR intensity was proportional to particle size in the FCOT and CR region. After appropriate data preprocessing, the local PLS model of 65-mesh samples showed the best prediction ability for
The authors declare that they have no competing interests.
Min Du, Zhisheng Wu, and Yanjiang Qiao designed the study. Min Du, Zhisheng Wu, and Xinyuan Shi performed the statistical analysis. Min Du, Bing Xu, and Zhisheng Wu wrote the paper. All authors read and approved the final paper. Zhisheng Wu and Min Du contributed equally to this work.
This work was supported by the National Natural Science Foundation of China (81303218) and Doctoral Fund of Ministry of Education of China (20130013120006). The authors thank Tong Ren Tang Technologies Co., Ltd., Beijing, China, for the assistance in instrument usage and CMM supply. They also thank the Key Laboratory of TCM-Information Engineering of State Administration of Traditional Chinese Medicine, Beijing, China, for the assistance in data processing.