Comparison of artificial neural networks (ANN) with classical modelling techniques using different experimental designs and data from a galenical study on a solid dosage form

Abstract

Artificial Neural Networks (ANN) methodology was used to analyse experimental data from a tabletting study and compared both graphically and numerically to classical modelling techniques (i.e. Response surface methodology, RSM). The aim of this investigation was to describe quantitatively the degree of data fitting achieved and the robustness of the developed models using two types of experimental design (i.e. a statistical, highly organised design and a randomised design). To compare goodness of fit, the R 2 coefficient was used, whereas for the robustness of the models the R 2 coefficient of an independent validation data set was computed. Comparable results were achieved for both ANN and RSM methodology when using the statistical plan. However, the robustness of the models when developed based on a randomised plan was clearly better for the ANN methodology. Based on the results of this study, it appears that the ANN methodology is much less sensitive to the organisational level of a trial design and is therefore better adapted to the data analysis of the results of historical or poorly organised trials. All tablet properties determined were largely influenced by the dwell time during compression as well as by concentration of silica aerogel and magnesium stearate, whereas the other factors showed very much weaker effects.