In the past few years, several research groups have worked on the design of efficient motion compensation systems for cardiac robotic-assisted Minimally Invasive Surgery (MIS). By providing surgeons with a stabilized work environment, significant improvements of the precision and repeatability of their gestures can be achieved. The design of a motion compensation system requires the accurate measurement of the heart motion, which can be achieved using computer vision techniques for tracking cardiac structures on the heart surface. However, most works in the literature focus on the representation and localization of cardiac structures while few explore their motion dynamics. In this paper we study and implement different adaptive methods for predicting the future heart motion using Kalman filtering. By exploiting the quasi-periodic nature of the heart motion, we are able to increase tracking robustness and computational efficiency. The experimental results indicate the significant increase in tracking performance when heart motion prediction is employed.