Studies On Nondestructive Quantitative Pharmaceutical Analysis Via NIR Spectroscopy Combined With Chemometrics

The Near-infrared (NIR) spectral region is generally defined as the wavelength range from 780nm~2500nm. It is customarily divided into two ranges, short-wavelength NIR spectroscopy (780nm~1100nm) and long-wavelength NIR spectroscopy (1100nm~2500nm). The spectra in this region are dominated by the absorption for overtones and combinations of fundamental vibration modes correspond mainly to O-H,C-H and N-H groups in compound. NIR spectroscopy has been proved to be a powerful analytical tool for analyzing a wide variety of samples that are used in agricultural, food, chemical and pharmaceutical industries, mainly due to its advantages over other analytical techniques, such as being expeditious, without destruction, low cost, being adaptable for almost all kinds of samples in all states and with little or no sample preparation. Frequently, the objective with this characterization is to determine the concentrations of different components in the samples. However, NIR spectra often contain serious systematic variation that is unrelated to the response data set, and the analyte of interest absorbs only in small parts of the spectral region. For solid samples this systematic variation is mainly caused by light scattering and differences in spectroscopic path length. Furthermore, the baseline and slope variations may often constitute the major part of the variation of the NIR spectra, and these variations may disturb the multivariate modeling and cause imprecise predictions for new samples. So the first step of a multivariate calibration based on NIR spectra is often to preprocess the original data.For the preprocessing of NIR spectral data, conventional methods that are commonly used including smoothing, derivation, multiplicative scatter correction (MSC) and standard normal variate (SNV). These signal corrections are different cases of filtering, practical effect of the derivation is that it removes an additive or multiplicative baseline, MSC is used to remove the variation that caused by differences in spectroscopic path length, and SNV is used to remove the variation that caused by light scattering. But common for all these methods is that they do not require a response variable in the preprocessing step, which is a prerequisite when orthogonal projection to latent structures (OPLS) method is applied. Being a generally applicable preprocessing and filtering method, OPLS provides a way to remove systematic orthogonal variation from a given data set. Compared with the original data, because the orthogonal variation is removed by applying OPLS method, the filtered data which is used as input data for the calibration model is simplified, thus the complexity of the calibration model is reduced and the predictive ability is improved.NIR spectroscopy also has some disadvantages resulting in weak, partly overlapped, non-specific bands, so it is hard to interpret of the data only depend on the conventional analytical method, and chemometrics methods must be applied in the qualitative and quantitative analysis in order to obtained precise result. Therefore, building up a multivariate calibration model using NIR spectral data is a key step for such quantitative analytical applications. The conventional chemometrics methods for the model establishment including: partial least squares regression (PLS), artificial neural network (ANN), support vector regression (SVR) and so on. PLS is the most commonly used method to make regression models, however, PLS is, in general, a linear calibration technique and it possesses some deficiencies in modeling of data sets containing strong nonlinear relationships; The distinct characteristic of ANNs is their ability to learn from examples and to get adapted with changing situations accordingly. In quantitative and qualitative analysis, ANNs have been more and more widely applied during the past several years, mainly due to their anti-jamming, anti-noise and robust non-linear transfer ability; SVR also possesses a strong ability of non-linear transfer, so it as a powerful new tool for data classification and regression has rapidly gained widespread acceptance in many fields during recent years. Compared with ANN system, SVR usually can get better results when the amount of training samples is small, and as the number of parameters to be set is small, the SVR system is easy to control, making it very fast in training. In this paper, NIR technology has been applied for quantitative analysis of the active components in Ampicillin, phenoxymethylpenicillin potassium, diclofenac sodium and aluminium hydroxide powder samples, respectively, and very satisfying results have been obtained. The main content is described as following:1. OPLS method was applied to preprocess the original spectral data of Ampicillin powder, the filtered data was used as input data of ANN model, and with the aid of degree of approximation, the parameters that affected the network were studied and the optimal ANN model was established. The concentration of Ampicillin as the active component was determined with the established ANN model, and very satisfying result was obtained with this proposed method. In order to evaluate the OPLS-ANN model, the calibration models that use first-derivative and second-derivative preprocessed spectral data of the samples were also designed. Experimental results show that OPLS-ANN model is the best.2. On the basis of original spectral data, first-derivative, second-derivative, MSC, SNV and OPLS preprocessed spectral data of phenoxymethylpenicillin potassium samples, the ANN models were established. In these models, the concentration of phenoxymethylpenicillin potassium as the active component was determined respectively, and very satisfying results were obtained. Experimental results also show that OPLS-ANN model is the best.3. The ANN models and PLS models that use above preprocessed spectral data of diclofenac sodium samples have been established separately. The concentration of diclofenac sodium as the active component was predicted respectively with these ANN and PLS models. Research shows that both the ANN model and the PLS model that based on OPLS preprocessing have the smallest RSE and the best R, and compared with PLS models, ANN models can obtain much better results. In addition, PC-ANN model has also been established to predict the concentration of diclofenac sodium, and experimental results show that although the result of PC-ANN model is a little worse than that of OPLS-ANN model, in PC-ANN model the scores of the principal components are chosen as input nodes instead of the original spectral data, thus the complexity of the model is greatly reduced and the training time is shortened, so the PC-ANN model is also an effective analytical tool.4. The SVR model that use original NIR spectra of aluminium hydroxide samples has been established to predict the concentration of aluminium hydroxide as the active component. For the same purpose, the ANN and PLS models on the basis of original spectral data and different preprocessed (first-derivative, second-derivative, MSC, SNV and OPLS) spectral data of the samples have also been established. Experimental results demonstrate that SVR model possesses several advantages: few parameters to be adjusted in the SVR model, and the model-selecting process being easily controlled. Furthermore, the SVR model constructed with the original NIR spectra of the samples can provide a better result, which suggests that SVR model can handle higher dimensional data better even with a relatively low amount of training samples.Above research demonstrate that near-infrared spectroscopy combined with chemometrics is feasible in quantitative analysis of drugs, so the NIR technology will have broad development prospects and application value in the field of pharmaceutical analysis. Furthermore, through the comparison of different preprocessing methods, we can know that multivariate calibration model established with OPLS preprocessed spectral data of samples can obtain much better prediction results, and the experimental results also show that SVR model is more superiority in pharmaceutical quantitative analysis.