An Improved Unscented Kalman Filter Based Decoder for Cortical Brain-Machine Interfaces

Abstract

Brain-machine interfaces (BMIs) seek to connect brains with machines or computers directly, for application in areas such as prosthesis control. For this application, the accuracy of the decoding of movement intentions is crucial. We aim to improve accuracy by designing a better encoding model of primary motor cortical activity during hand movements and combining this with decoder engineering refinements, resulting in a new unscented Kalman filter based decoder, UKF2, which improves upon our previous unscented Kalman filter decoder, UKF1. The new encoding model includes novel acceleration magnitude, position-velocity interaction, and target-cursor-distance features (the decoder does not require target position as input, it is decoded). We add a novel probabilistic velocity threshold to better determine the user's intent to move. We combine these improvements with several other refinements suggested by others in the field. Data from two Rhesus monkeys indicate that the UKF2 generates offline reconstructions of hand movements (mean CC 0.851) significantly more accurately than the UKF1 (0.833) and the popular position-velocity Kalman filter (0.812). The encoding model of the UKF2 could predict the instantaneous firing rate of neurons (mean CC 0.210), given kinematic variables and past spiking, better than the encoding models of these two decoders (UKF1: 0.138, p-v Kalman: 0.098). In closed-loop experiments where each monkey controlled a computer cursor with each decoder in turn, the UKF2 facilitated faster task completion (mean 1.56 s vs. 2.05 s) and higher Fitts's Law bit rate (mean 0.738 bit/s vs. 0.584 bit/s) than the UKF1. These results suggest that the modeling and decoder engineering refinements of the UKF2 improve decoding performance. We believe they can be used to enhance other decoders as well.

Keywords: brain-computer interface; brain-machine interface; encoding model; neural decoding; neuroprosthetic; unscented Kalman filter.

Figure 1

Array implant locations and example recorded signals. (A) Implant location. (B,C) Photos taken during surgeries. (D,E) Sample waveforms. Each sub-panel shows waveforms from one channel, drawn as mean ± one standard deviation.

Figure 2

Experimental setup and behavioral tasks. (A) The monkey sat in a primate chair 55 cm before a computer screen and grasped a 6.5 cm tall joystick with 4 cm maximum deflection in its right hand. (B) Center-out task. The monkey alternatively moved the cursor to center targets and peripheral targets, located at random angles and fixed distance from center. (C) Pursuit task with Lissajous curve: the monkey kept the cursor within a target which moved continuously following a Lissajous curve. (D) Pursuit task with point-to-point trajectory: the monkey kept the cursor within a target which moved continuously between randomly selected points on the screen.

Figure 3

Offline reconstruction accuracy. (A) Mean ± SEM of signal-to-noise ratios. (B) Mean±SEM of correlation coefficients. (C) Reconstruction accuracy when pooling data from two monkeys. White bars show accuracy when using spike sorting. Gray bars show accuracy when using unsorted spikes derived by merging all sorted units on each channel. Right side bars show UKF1 augmented with each of: +A, acceleration; +PVI, position-velocity interaction; +T, target; +SH, spiking history of population. (D) Example reconstruction of x-axis velocity from one session of monkey M.

Figure 4

Encoding model prediction accuracy. (A) Histogram of spike count prediction accuracies measured by correlation coefficient. Dashed vertical lines indicate means for each encoding model. (B) Mean CC and SNR of each encoding model, including UKF1 augmented with each of: +A, acceleration; +PVI, position-velocity interaction; +T, target; +SH, spiking history of population. SEM was not calculated since data likely includes repeated observations of neurons. (C) Mean±SEM of CC from one session of each monkey when using all sorted units. (D) Mean ± SEM of CC from one session of each monkey when using units with mean spike height (peak to trough) in the top 10 percentile of all units in the session.

Figure 5

Position-velocity interaction in the encoding of a motor cortical single unit. (A) Illustration of position-velocity interaction tuning. Shading indicates firing rate. Each sub-panel depicts velocity tuning when the cursor was in a portion of the position work space, with the sub-panel's position corresponding to the cursor position. Location within each sub-panel corresponds to the 2D cursor velocity, with zero velocity in the center. See text for details. (B) Example spike waveforms from this single unit.

Figure 6

Comparison of decoders during closed-loop neural control of cursor. (A) Mean ± SEM of fraction of targets acquired. “Hand” indicates performance when monkey controlled the cursor using its hand via the joystick. (B) Mean ± SEM of movement time per peripheral target. (C) Mean ± SEM of Fitts's Law bit rate. (D) Example movement trajectories generated under UKF2 control. Peripheral targets and paths have been rotated so that all peripheral targets align. (E) Trajectories under Kalman control. (F) Trajectories under UKF1 control.