Studies On Pharmaceutical Quantitative Analysis Based On SVM Using Wavelet Denoised Short-wave NIR Spectra

The application of Near-infrared spectroscopy (NIR) in rapid and non-destructive analysis of intact tablet medicament in the spectral region from 1100 to 2500 nm (long-wave NIRS) has become more and more popular and shown relatively dependable results. However, short-wave NIRS in a region from 780 to 1100 nm has several advantages over long-wave NIRS. The scanning speed of narrow range could be about four times faster than that of wide range (1100-2500nm). The transmission depth penetrating through solid samples can be up to several centimeters. In addition, a silicon detector that can be used below 1100 nm is relatively inexpensive and the wavelength range of most UV-vis spectrometers is extended to this region (780-1100nm). Long range analysis with an optical fiber chemical sensor can detect poisonous samples, or samples in containers. Because of the weak intensity of diffuse reflectance spectra of solid samples, an efficient modeling system is needed. Support vector machine was introduced and applied to modeling multivariate non-linear systems of calibration samples.Support vector machine (SVM) introduced by Vapnik is an elegant tool for solving pattern recognition and regression problems. Over the past few years, a lot of researches of SVM systems have demonstrated the dependable capability of this method in classification. Only recently has SVM been applied to data regression, giving dependable results. The support vector machine for estimating real-valued function was described by V. Vapnik. It is based on the solution of a quadratic optimization problem which represents the tradeoff between the minimization of the empirical error and the maximization of the smoothness of the regression function. Compared to neural networks such as ANN system, SVM has a solid theoretical foundation with fewer parameters to be set, the SVM system is easy to control, but over fitting phenomenon appears if the parameters have been set improperly.Furthermore, wavelet method has been employed to minimize the influence of noise. Wavelets are a relatively new technique in the field of chemometrics that is to use mathematical functions to divide the data into different frequency components, each component is studied with a resolution matched to its scale. Wavelets are analogous to the Fourier transform, but preserve the time information upon transformation and are more adapted to dealing with non-stationary data.1. In this study, short-wave NIR spectra of metronidazole powder in different calibration sample size (70 and 26) were investigated to evaluate the applicability of SVM method as a powerful tool for quantitative prediction with only small sample size. During the method development diverse spectra pretreatments were compared to calculate the best calibration models. Data treatment was performed using the raw data of the standard normal variate (SNV), the multiplicative scatter correction (MSC), the orthogonal signal correction (OSC) and 1st derivation. The SVM models with SNV pretreatments showed equal predict ability when measuring the large sample set as small one, and the models using MSC, OSC and 1st derivation as data pretreatments showed higher RSE in the small sample test than in the large one. The models using SNV pretreatment calculated lower RSE than using the others. The result showed that SVM method on SNV data could be useful for pharmaceutical quantitative prediction required less samples.2. Calibration models for the determination of cimetidine tablets have been developed using short-wave NIR in diffuse reflectance mode based on SVM using kernel functions as sigmoid, polynomial and RBF for model building and validation. SVM models using RBF kernel showed more reasonable result than the two others for nonlinear regression, and produced more accurate prediction. Wavelet denoising method was introduced to comparer with the SNV pretreatment. Both the RSE and the correlation coefficients with wavelet denoised pretreatment are better than with SNV of SVM models using three kinds of kernel function. The performance and robustness of SVM regression are compared to partial least square (PLS) regression. SVM regression produced more accurate prediction. These data thus suggest that NIRS is a tool which can be developed for the rapid prediction of the nutritional value of feedstuffs with a precision which makes it attractive for use as a routine quality control tool in feed mills. SVM using RBF kernel function could build accurate and robust calibration models for the nondestructive analysis in prediction of intact troche from wavelet denoised short-wave NIR data.3. SVM calibration models for the quantitative analysis of paracetamol tablets have been built using NIR in defused reflectance mode. Diverse pretreatment with different wavelets and decomposition levels were compared to calculate the best calibration models. It is shown that unobvious effect can be generated by diverse wavelets with same decomposition levels. The result showed that most SVM models gave relative reasonable results with decomposition level of 3, and the prediction of SVM models from lower decomposition level wavelet denoised data was better than that of upper levels. It may be concluded that in the wavelet denoising process at large number of decomposition levels for short-wave near infrared diffused reflectance spectra the noise have been removed with the removal of some physical and chemical information simultaneously.4. The wavelet mother db8 at level 3 was used to denoise the short-wave NIR spectra of Ciprofloxacin Hydrochloride powder with the two kinds of white noise removal(unscaled and scaled) , two threshold selection rules(minimax and fixed), and two threshold functions (soft and hard). During method development, raw data and diverse spectra pretreatments were used to calculate the best SVM calibration models. Among all the models built with the data that had been removed unscaled white noise, only the model from the combination of fixed threshold and hard threshold function denoised data pretreatment gave better result than SVM model from original spectra. All models by the removal of scaled white noise pretreatment gave better result than SVM model from original spectra, the predictions of models calculated by minimax threshold were better than by fixed threshold, and it showed unobvious difference with the predictions of models by soft and hard threshold functions. The best model with unsealed white noise pretreatment gave a slightly better prediction than the best model with scaled white noise. Result showed that, the removal of different kinds of noise can affect the model obviously, and the different ways in combination of threshold and threshold functions can also make some affection on the predictive ability of the model.