When designing a self–tuning controller for multivariable systems a proper representation of the model structure is important, particularly if the interactions between loops are significant. A popular transfer function structure used to describe multivariable processes is the P–canonical form structure where loop interactions are treated as feedforward couplings. However, polynomial–based controllers can also be applied to multivariable systems by designing several single-input single–output controllers, and compensation for cross–coupling between the different loops can be achieved by treating these interactions as feedforward measurable disturbances. This is the theme of this paper which considers the extension of the Generalized Predictive Control algorithm (GPC) to this technique. Following a derivation of the control strategy, called Generalized Predictive Control with Feedforward (GPCF), it is applied to a realistic nonlinear model for anaesthesia in a series of simulations. These results are compared with those obtained using the multivariable GPC version with a P–canonical form representation for the discrete multivariable model. The GPCF scheme is shown, in this case, to offer advantages over the multivariable GPC in terms of transient responses, interaction reduction, control quality, and computational burden.