Comparing the predictability of different chemometric models over UV-spectral data of isoxsuprine and its toxic photothermal degradation products.

Isoxsuprine (ISX) is widely used for cerebral and peripheral vascular diseases. A comparative study was held among different multivariate calibration models for selective determination of a complex mixture of Isoxsuprine and four of its toxic photothermal degradation products that impair kidney and liver functions. The Partial Least Squares (PLS) and Artificial Neural Network (ANN) models were applied on the specific spectrum and on selected wavelengths using genetic algorithm (GA) technique as an efficient variable selection tool. The effect of GA on the model construction and performance was evaluated. The multilevel multifactor experimental design was adopted for the construction of the calibration set. Optimized parameters were used for the development of the different models. The performances of the developed models were assessed by predicting the concentration of eight different mixtures composing the validation set. Results were compared to one another and to the official method using one-way ANOVA statistical test to assure the validity of the constructed models. The lower chance of overfitting offered by ANN minimized the RMSEP relative to the PLS. On the other hand, the application of GA prior to model implementation affected the number of latent variables the prediction ability of both PLS and ANN models. The validated models were successfully applied as stability indicating assay methods for the selective determination of ISX and its photothermal degradation products in ISX raw material and market formulations.