Real-Time Linear Prediction of Simultaneous and Independent Movements of Two Finger Groups Using an Intracortical Brain-Machine Interface

Abstract

Modern brain-machine interfaces can return function to people with paralysis, but current hand neural prostheses are unable to reproduce control of individuated finger movements. Here, for the first time, we present a real-time, high-speed, linear brain-machine interface in nonhuman primates that utilizes intracortical neural signals to bridge this gap. We created a novel task that systematically individuates two finger groups, the index finger and the middle-ring-small fingers combined, presenting separate targets for each group. During online brain control, the ReFIT Kalman filter demonstrated the capability of individuating movements of each finger group with high performance, enabling a nonhuman primate to acquire two targets simultaneously at 1.95 targets per second, resulting in an average information throughput of 2.1 bits per second. To understand this result, we performed single unit tuning analyses. Cortical neurons were active for movements of an individual finger group, combined movements of both finger groups, or both. Linear combinations of neural activity representing individual finger group movements predicted the neural activity during combined finger group movements with high accuracy, and vice versa. Hence, a linear model was able to explain how cortical neurons encode information about multiple dimensions of movement simultaneously. Additionally, training ridge regressing decoders with independent component movements was sufficient to predict untrained higher-complexity movements. Our results suggest that linear decoders for brain-machine interfaces may be sufficient to execute high-dimensional tasks with the performance levels required for naturalistic neural prostheses.

Real-Time Brain-Machine Interface Achieves High-Velocity Prosthetic Finger Movements using a Biologically-Inspired Neural Network Decoder

“…The ReFIT Kalman Filter, introduced by Gilja et al 6 , is a two-step training process that first computes the weights of a classic Kalman filter and then modifies the weights when the prosthesis direction is not toward the actual target. In this study, we find that the ReFIT neural network decoder substantially outperforms our previous implementation of the ReFIT Kalman filter [20][21][22] with >60% increase in throughput by utilizing high-velocity movements without compromising the ability to stop. This enables the use of shallow artificial networks, which may resemble biological motor pathways, for motor decoding applications and may be the bridge toward high-velocity, naturalistic robotic prostheses.…”

“…1b. We have recently demonstrated online real-time decoding of these 2 degrees of freedom using a ReFIT Kalman filter 21 , and primarily compare our novel algorithm to that approach.…”

“…Limitations. As opposed to using our typical center-out task finger task 21 , we increased the difficulty to better challenge the decoders and elucidate differences between two wellperforming decoders. Although these results could apply to a variety of real-word finger tasks, other tasks and prostheses (i.e., robotic arms) were not explicitly tested.…”

“…For Monkey W, only the first trial was discarded as there were fewer total trials since W was less motivated to complete trials. To visually illustrate the peak performance of the RN on our typical center-out task 21 , one additional day was included using the RN on this task.…”

“…ReFIT neural network decoder outperforms optimized RK decoder. Our implementation of the ReFIT Kalman filter for finger control utilizes a physiological lag and does not include hyper-parameter tuning (i.e., gain and smoothing parameters) [20][21][22] . However, other work suggests RK performance can be improved without lag (providing the RK updates to the virtual fingers without delay) 24 and by optimizing the online gain and smoothing parameters for RK 25 .…”