Artificial neural networks applied to the in vitro in vivo correlation of an extended-release formulation: Initial trials and experience

Artificial neural networks applied to in vitro-in vivo correlations (ANN-IVIVC) have the potential to be a reliable predictive tool that overcomes some of the difficulties associated with classical regression methods, principally, that of providing an a priori specification of the regression equation structure. A number of unique ANN configurations are presented, that have been evaluated for their ability to determine an IVIVC from different formulations of the same product. Configuration variables included a combination of architectural structures, learning algorithms, and input-output association structures. The initial training set consisted of two formulations and included the dissolution from each of the six cells in the dissolution bath as inputs, with associated outputs consisting of 1512 pharmacokinetic time points from nine patients enrolled in a crossover study. A third formulation IVIVC data set was used for predictive validation. Using these data, a total of 29 ANN configurations were evaluated. The ANN structures included the traditional feed forward, recurrent, jump connections, and general regression neural networks, with input-output association types consisting of the direct mapping of the dissolution profiles to the pharmacokinetic observations, mapping the individual dissolution points to the individual observations, and using a "memorative" input-output association. The ANNs were evaluated on the basis of their predictive performance, which was excellent for some of these ANN models. This work provides a basic foundation for ANN-IVIVC modeling and is the basis for continued modeling with other desirable inputs, such as formulation variables and subject demographics.