Pitfalls of artificial neural networks (ANN) modelling technique for data sets containing outlier measurements using a study on mixture properties of a direct compressed dosage form.

An application of the Artificial Neural Networks (ANN) methodology was investigated using experimental data from a mixture properties study and compared to classical modelling technique (i.e. Response Surface Methodology, RSM) both graphically and numerically. The aim of this investigation was to quantitatively describe the achieved degree of data fitting and robustness of the developed models. For comparing the goodness of fit, the R2 coefficient was used, whereas for the robustness of the models an outlier measurement was integrated in the data set. Comparable results were achieved for both ANN- and RSM methodologies for data fitting. The robustness of the models towards outliers was clearly better for the RSM methodology. All determined mixture properties were mainly influenced by the concentration of silica aerogel, whereas the other factors showed very much lower effects. For that reason the physical properties of this excipient (e.g. its specific surface area) are of importance for the behaviour of the mixtures.