Performance evaluation of linear brain machine interface decoders in neural state space

Brain-Machine Interfaces (BMIs) have the potential to restore lost sensorimotor functions in people with severe motor disabilities. Several BMI decoding strategies have been suggested to translate activity of motor neurons into control signals that ac-tuate artificial devices. Among these, the class of linear decoders, particularly Wiener filters, is known to perform well for simple tasks, but degrades considerably as a function of increasing task complexity. In this work, we study the mathematical properties of the solution subspace of Wiener decoders in an effort to derive a desired neural state trajectory that is optimal for a given decoder and a desired biomimetic kinematic solution. We show that the error between the desired neural trajectory and the actual one measured during the performance of a 2D reach task provides reliable estimation and prediction of the performance in the task space. We demonstrate a significant correlation between the error measure in the neural state space and the error measure in the task space, which allows potential future use of this error measure as a way to estimate the true motor intent and the extent of learning the decoder by BMI subjects, and possibly as a feedback signal to improve their online decoding performance.