Artificial neural networks in the optimization of a nimodipine controlled release tablet formulation

Abstract

Artificial neural networks (ANNs) were employed in the optimization of a nimodipine zero-order release matrix tablet formulation, and their efficiency was compared to that of multiple linear regression (MLR) on an external validation set. The amounts of PEG-4000, PVP K30, HPMC K100 and HPMC E50LV were used as independent variables following a statistical experimental design, and three dissolution parameters (time at which the 90% of the drug was dissolved, t(90%), percentage of nimodipine released in 2 and 8h, Y(2h), and Y(8h), respectively) were chosen as response variables. It was found that a feed-forward back-propagation ANN with eight hidden units showed better fit for all responses (R(2) of 0.96, 0.90 and 0.98 for t(90%), Y(2h) and Y(8h), respectively) compared to the MLR models (0.92, 0.87 and 0.92 for t(90%), Y(2h) and Y(8h), respectively). The ANN was further simplified by pruning, which preserved only PEG-4000 and HPMC K100 as inputs. Optimal formulations based on ANN and MLR predictions were identified by minimizing the standardized Euclidian distance between measured and theoretical (zero order) release parameters. The estimation of the similarity factor, f(2), confirmed ANNs increased prediction efficiency (81.98 and 79.46 for the original and pruned ANN, respectively, and 76.25 for the MLR).