In the present study, a simple method, based on diffuse reflectance FTIR spectroscopy (DRIFTS) and artificial neural network (ANN) modeling is developed for the simultaneous quantitative analysis of mebendazole polymorphs A–C in powder mixtures. Spectral differences between the polymorphs are elucidated by computationally assisted band assignments on the basis of quantum chemical calculations, and subsequently, the spectra are preprocessed by calculation of 1st and 2nd derivatives. Then ANN models are fitted after PCA compression of the input space. Finally the predictive performance of the ANNs is compared with that of PLS regression. It was found that simultaneous quantitative analysis of forms A–C in powder mixtures is possible by fitting an ANN model to the 2nd derivative spectra even after PCA compression of the data (RMSEP of 1.75% for form A, 1.85% for B, and 1.65% for C), while PLS regression, applied for comparison purposes, results in acceptable predictions only within the 700–1750 cm −1 spectral range and after direct orthogonal signal correction (DOSC), with RMSEP values of 2.69%, 2.68%, and 3.40% for forms A, B, and C, respectively. Application of the ANN to commercial samples of raw material and formulation (suspension) proved its suitability for the prediction of polymorphic content.
Read full abstract