Application of diffuse reflectance infrared Fourier transform spectroscopy combined with artificial neural networks in analysing enantiomeric purity of terbutaline sulphate bulk drug

A new, simple and rapid analytical method was developed to analyse enantiomeric purity of terbutaline bulk drug. Diffuse reflectance infrared Fourier transform spectroscopy (DRIFTS) was combined with artificial neural networks (ANNs) for data modelling. Series of 10 binary mixtures were made from the two enantiomers with different mass ratios, dispersed as a 5% (w/w) mixture in KBr and spectra were sampled between 650.16 and 3999.40 wavenumbers (cm-1). Collected spectral intensities from 1720 wavenumbers were reduced into 172 averaged spectral values from 10 consecutive wavenumbers. These 172 values were used as inputs for the initial ANN model. Following a unit penalty the number of inputs was reduced to 23, 6 and 3 spectral inputs with unit penalty factors of 10^-5, 5×10^-4 and 10^-3, respectively. The ANN model was built by comparing the prediction obtained from several high scoring ANN models. The number of hidden layers and hidden neurons, activation function and postsynaptic potential (PSP) function were optimised. The best model was a radial basis function network (RBF) with 23 selected inputs, one hidden layer with 10 neurons and one output neuron for the percentage of S-terbutaline enantiomer in the mixture. Square root activation function was used to transform the squared distance activation in the second layer of radial units to the actual distance as an output. Sensitivity reports of selected inputs showed that the most important peaks were found to be peaks at 1227 due to C---N stretch of amines and at 2346cm^-1 due to N---H stretch of secondary amino group. The level of R-enantiomer in binary mixtures varied from 5 to 100%. The results were in close agreements with the true values calculated from the masses of the enantiomers in the mixtures. The mean±S.D. recovery values was 99.3±12.74. Linear regression analysis of theoretical composition against predicted values gave slopes of 0.973, intercept value of 0.80 and correlation coefficients of R^2=0.992. Intercept was not significantly different from zero (t= 0.72) and slope was not different from unity (t=-1.23) indicating no method bias and absence of proportional error. The minimum quantifiable level (MQL) of 17.45% and the limit of detection (LD) of 5.23% (w/w) for the R-stereoisomer was calculated. The relatively high MQL and LD offer a quantifiable range of 17.43-100% of R-stereoisomer dispersed as 5% mixture in KBr.