2 Nonparametric methods
All nonparametric methods are based on training an empirical function to
replicate the input-output behaviour of the QSP system. The training
involves generation of a large amount of pseudo-data from the full-order
QSP model and then either fitting a user-defined empirical model (termed
here Empirical Approximation) to the data or by a machine learning
technique such as an artificial neural network (ANN). None of these
methods retain mechanistic characteristics of the full-order model and
therefore cannot be used for extrapolation to data that might arise from
inputs beyond the range of the training inputs (e.g. to a dose that is
not within the range of doses used to create the pseudo-data). These
approaches are useful when parametric methods are not possible due to
high model complexity or even when the structure of the model is
unknown, e.g., models provided by third parties. The only requirement of
nonparametric approaches is the ability to evaluate the original
full-order model at a given set of input variables.