State-space decoding of goal-directed movements

Bayesian inference methods hold great promise for the prediction of hand-movement trajectories in neural prosthetic devices. The accuracy of such probabilistic methods can be improved by incorporating meaningful priors, thereby appropriately constraining the space of possible states that the system can attain. In this work we review and extend. methods for constructing reach trajectories that incorporate prior information of the intended movement target. For computational tractability, we model arm motion as a linear dynamical system driven by Gaussian noise, conditioned on this end-point information. These assumptions, while biomechanically unrealistic, give rise to a priori model arm-paths that share many of the characteristics of natural arm trajectories. Moreover, in this model formulation we may compute the predicted arm position, given simultaneously observed neural data, using standard forward-backward computations familiar from the theory of the Kalman filter. Here we review an earlier recursive approach for computing such reach trajectories and present a new nonrecursive approach, with computations that may be performed analytically for the most part, leading to a significant gain in the accuracy of the inferred trajectory while imposing a very small computational burden. Finally, we discuss extensions of our approach, including the incorporation of multiple target observations at different times, and multiple possible target locations.