Reliability of Artificial Neural Network Predictions - A Case Study in Drug Release Profile Predictions

Artificial neural networks have been widely used in pharmaceutical research such as for estimation of process coefficients and pharmacokinetic parameters. In this paper, we present a study on the use of the Radial Basis Function-based Gaussian mixture model to predict dissolution profiles of a matrix controlled release theophylline pellet preparation. Performance of the network has been assessed using similarity factor—an index for profile comparison in pharmaceutical research. In addition, we also investigate the phenomena of interpolation and extrapolation of the test data sets that will affect the reliability of network predictions. The Parzen-window approach has been employed to determine, based on the calculated data densities, whether the trained network produces interpolated or extrapolated predictions. The experimental results are discussed and analyzed.