Genetic algorithm optimized neural networks ensemble as calibration model for simultaneous spectrophotometric estimation of atenolol and losartan potassium in tablets

Improvements in neural network calibration models by a novel approach using neural network ensemble (NNE) for the simultaneous spectrophotometric multicomponent analysis are suggested, with a study on the estimation of the components of an antihypertensive combination, namely, atenolol and losartan potassium. Several principal component neural networks were trained with the Levenberg-Marquardt algorithm by varying conditions such as inputs, hidden neurons, initialization, training sets and random Gaussian noise injection to the inputs. Genetic algorithm (GA) has been used to develop the NNE from the trained pool of neural networks. Subsets of neural networks selected from the pool by decoding the chromosomes were combined to form an ensemble. Several such ensembles formed the population which was evolved to generate the fittest ensemble. Ensembling the networks was done with weighted average decided on the basis of the mean square error of the individual nets on the validation data while the ensemble fitness in the GA optimization was based on the relative prediction error on unseen data. The use of a computed calibration spectral data set derived from three spectra of each component has been described. The calibration models were thoroughly evaluated at several concentration levels using spectra obtained for 76 synthetic binary mixtures prepared using orthogonal designs. The ensemble models showed better generalization and performance compared with any of the individual neural networks trained. Although the components showed significant spectral overlap, the model could accurately estimate the drugs with satisfactory precision and accuracy, in tablet dosage with no interference from excipients as indicated by the recovery study results. The GA optimization guarantees the selection of the best combination of neural networks for NNE and eliminates the arbitrariness in the selection of any single neural network model, thus maximizing the knowledge utilization without the risk of memorization or over-fitting.