Optimal feedback-controlled point process decoder for adaptation and assisted training in brain-machine interfaces

Abstract

Closed-loop decoder adaptation (CLDA) improves brain-machine interface performance by estimating the decoder parameters in closed-loop operation. By allowing the subject to use an initialized decoder, CLDA techniques infer the intended movement in closed loop, and refine the decoder parameters based on this inference and the recorded neural activity. In some cases, an initialized decoder may be far from optimal. This may cause the initial decoded trajectories to be biased towards a specific region of space, hence affecting the speed of parameter convergence by not allowing the decoder to explore the space. Moreover, this can lower the subject's motivation level due to low initial performance. Here we propose a new combined assisted training and CLDA algorithm based on an infinite-horizon optimal feedback control design for the BMI. We derive an adaptive optimal feedback-controlled point process decoder (FC-PPF) that estimates the subject's intended movement based on the visual feedback of the decoded kinematics and the target position. Using closed-loop BMI simulations that explicitly model the visual feedback delay, we show that decoder parameter values in the adaptive FC-PPF converge to the true values even when initialized arbitrarily. We also show that adaptive FC-PPF can be used as both a combined assisted training and CLDA technique or as a CLDA technique without assistance, hence providing a possible unified framework for closed-loop decoder adaptation.