Magnetorheological elastomers (MREs), as a class of smart materials, have a property called Magnetostriction means mechanical properties, including deformation of MREs, could be changed in response to an external magnetic field. Because of the controllable deformation, MRE is a suitable candidate for rendering the loss of haptic feedback in Robot-Assisted Cardiovascular (RCI) applications. In the recently-designed such force feedback systems, i.e. TorMag, the effect of matrix shear modulus and filler volume percentage was not studied comprehensively. Tormag also exposed limitations in force range. In the current study, a previously proposed and validated constitutive model of MREs was adopted. Then, twelve MREs with three silicon rubber matrices and four filler volume fractions were fabricated and characterized to improve the limitations mentioned above in Tormag. The average relative error between analytical force range and experiment was 10.2\%, while the maximum force range was 5.29 N (stiffest matrix and 40\% filler), and the minimum range was 1.06 N (softest matrix and 10\% filler). Increasing filler percentage from 10\% to 40\% increased the force feedback range up to 288\%. The state-space analysis of Tormag revealed that this system did not fully cover the required force range and zero force rendering. As an approach, structural optimization of the system is performed using the local and global optimization process. Next, a neural network (NN)-based model as the control framework was proposed and validated to obtain the necessary force for the desired input data. Then, a nearest neighbour search (NNS) method was added to the NN model to find the required magnetic field for a force-displacement profile as input. The proposed neural network accurately predicted the force-displacement behaviour of three types of MREs ($R^2=0.97$, mean-absolute-error=1.26 N). Also, the NN+ NNS model successfully obtained the required magnetic field (mean-absolute error=3.64 mT).