Studies On Non-destructive Quantitative Analysis Of Drugs Using ANN On NIR Spectroscopy

With widely application of the computer technology in analyticalchemistry, the modern analytical instruments and automatizationhave the fabulous development. Pharmaceutical analysis as theimportant section of the analytical chemistry has become the modernpharmaceutical analysis using the analytical instruments principally.Then the analysts have to face a problem: how to select the fittingmeasured method and the optimal measured process and how toextract more meaningful chemical information from the originalexperimental data while modern analytical instruments give a largeamount of measured data rapidly and precisely. This is the main taskfor the analysts.Near-infrared (NIR) spectroscopy and chemometrics were moreand more recognized and spread by the analysts and became asingle analytical technology owing to the broad and embeddedapplication in analytical chemistry with the combination ofmathematics, statistics and computers in 1980's, changing the ideathat theories trail practices step by step.NIR spectroscopy has proved to be a powerful analytical tool foranalyzing a wide variety of samples used in the agricultural, food,petrochemical, textile and pharmaceutical industries, especially theuse of NIR spectroscopy for the quantitative analysis ofpharmaceutical samples has been significantly increased during thelast decade. The NIR spectral region is generally defined as thewavelength range from 780nm～2500nm. It is customarily dividedinto two ranges, short-wavelength NIR spectroscopy (780nm ～1100nm)and NIR spectroscopy (1100nm～2500nm). In NIR region, itis possible to observe mainly overtone and multiple bands andcombinations of absorption bands normally occurring in the mid-IRregion and associated with the O-H,C-H and N-H bonds present inorganic molecules. Most compounds have absorption in this region.They are be combined with molecular inner structure,functionalgroup and molecular state, therefore the quantitative and qualitativeinformation may be obtained from NIR spectroscopy. Compared withthe traditional analytical methods, NIR spectroscopy technology hasa lot of advantages and is a simple rapid method. The mostadvantage of NIR is that the samples are not restricted inappearance, which can be determined no matter what they are gas,liquid and solid. We may determine the samples using diffusereflectance spectroscopy directly because of its strong penetrableability. The solid samples may be any shapes such as fruits, cornsand solid tablets. However, NIR spectroscopy has somedisadvantages resulting in weak, partly overlapped, non-specificbands that require use of a multivariate calibration technique in orderto correlate the spectral information they contain to the concentrationof the target analyte.Many multivariate calibration methods have been adopted suchas multiple linear regression (MLR), principal component regression(PCR) and partial least squares (PLS) in Chemometrics, and PLS isa usual tool for multivariate calibration because of the obtainedcalibration models, the ease of its implementation and the availabilityof a prior separation for analysis. However, in the presence ofsubstantial non-linear, PLS tends to give large prediction errors andcalls for more robust models. Non-linear calibration technique suchas artificial neural networks (ANN) has gained much focus in recentyears in the cases. The main advantage of ANN is its anti-jamming,anti-noise and robust nonlinear transfer ability. It has beendemonstrated that it is possible to obtain excellent results inmultivariate calibration problems using ANN. In the proper model,ANN results in lower calibration errors and prediction errors.The conventional determination of active compounds content bythe use of Chinese Pharmacopoeia and State Drug Standard inpharmaceutical analysis often calls for using various analyticaltechniques, including chromatographies (GC, HPLC) andspectroscopies (IR, UV). However, these methods usually requiredissolving the samples, separating them and determining theiringredients. In this paper the active components of six different drugswere determined using back-propagation artificial neural networks onNIR spectroscopy. It indicates that the ANN models have the bestresults.In this paper, five main parameters, namely, input nodes, hiddennodes, momentum, learning coefficient and number of iterations,were optimized and discussed;furthermore, the optimal ANN modelswere obtained. The present criterion of optimization is to make theerror of the training set the smallest;however, it is very easy tochoose an overfitting model, namely, the error of test set is notsmallest. In order to avoid establishing overfitting models, a newconcept, degree of approximation, was cited. The citation of degreeof approximation avoids bringing overfitting phenomenon. Someother samples monitor training set while the network is trained, andtraining and predicting all get optimization at last.Different pretreated techniques are usually applied to originalspectra. Original spectra contain not only the information from thecomponents of the samples, but many noises from any aspects.These noisy signals can interfere in spectral information, therefore,original spectra must be pretreated. The general pretreated methodsincluding first-derivation, second-derivation, standard normal variate(SNV) and multiplicative scatter correction (MSC) were studied anddiscussed detailedly in this paper. The results show that the optimumpretreated method of different drugs is different. Of all the pretreatedmethods, MSC results in the worst errors because of the uniformdimension of the sample granules.It is feasible to obtain quantitative information using ANN byprocessing NIR spectroscopy of both simple active component drugsand two active components drugs. Although ANN models establishedneed a little long time, unknown samples predicted need little time. Inthe two components drugs, the ANN models both two componentsdetermined simultaneously and two components determinedseparately were established. They have the familiar concentrationerrors and correlation coefficient;therefore, the two componentsdetermined simultaneously can be used.NIR spectra of the tables were measured directly and ANN tabletmodels were established. Then the tablet samples were ground upcarefully and turned into powder and ANN powder models wereestablished. Compared with tablet models and powder models, theirreliability was high. Therefore, satisfactory ANN model was likely tobe established by directly using tablet samples other than by usingground powder samples.Both ANN models and PLS models were established in compoundparacetamol and diphenhydramine hydrochloride powdered drug.Together with ANN and PLS multivariate calibrations, reasonablygood predicted indicators of paracetamol and caffeine were obtainedwith an improvement on the results when ANN was applied. Itilluminates that PLS calibration results in the poor predicted modelsdue to its overlapped bands, complicated absorptance, andnonlinearity. On the contrary, ANN method can result in the robustmodels due to its anti-jamming, anti-noise and nonlinear transferability. ANN method has the superiority in quantitative analysis ofcomplicated drugs.The network models of compound aminopyrine phenacetin tabletsand compound aspirin tablets were established. In these modelsprincipal component analysis (PCA) was added to the ANN algorithm,i.e., the scores of the principal component analysis of the responsedata of the calibration mixture were used as an input layer. PC-ANNarchitectures were constructed by using different number of principalcomponents. The use of scores reduced input nodes, so training timeof the network was shortened. In comparing PC-ANN with ANNmodels, both RSE and R2 of ANN models were worse than those ofPC-ANN models. The results indicate the PC-ANN models havemore advantages than ANN models.The methods of ANN on NIR spectroscopy are explored fornondestructive quantitative analysis of solid pharmaceutical samples.It is feasible to obtain quantitative information by processing NIRspectroscopy with ANN models. ANN must have been expansiveforeground and applied values.