Neuro-Prosthetic Control

Project Overview

Neuroprosthetic is an example of neuroengineering resulted from the application of the knowledge in decoding the brain signals to assist paraplegic patients to move by volitional control signals recorded from the brain.

The neural code used in voluntary movement execution is known in neuroscience as "population vector code."
Arm movement trajectory can be predicted based on the the population vector code of a large ensemble of neurons recorded in the motor cortex.
With the ability to decode the motor-command signal in the motor cortex, arm movement direction and velocity can be specified based on the neural code.
Paralyzed patients will be able to move a robot arm based on the neural code extracted from the neurons in the motor cortex.

Primates are used as the experimental model to test these neuroengineering techniques to make voluntary arm movements by "thinking about it" without actual execution of arm movements by making vitrual 3-D arm movements via virtual-reality display feedback fitted in front of the animal. [Primate research is done by Dr. Andrew Schwartz at the University of Pittsburgh. Neural spike train data are provided by Dr. Schwartz.]
The goal is to develop real-time decoding algorithms to extract information from the implanted multi-electrode array in the motor cortex to drive the robotic arm.

The neural code used in voluntary movement execution is known in neuroscience as "population vector code."
Arm movement trajectory can be predicted based on the the population vector code of a large ensemble of neurons recorded in the motor cortex.

Real-time implementation of the algorithm is essential. It is computationally expensive to analyze a hundred channels of neural data simultaneously. Efficient computational algorithms are needed. Off-line analysis can be performed without any constraints on the real-time performance. But if the neuroprosthesis were to be practical, real-time responses are essential.
Incomplete data set. Ideally, there should be enough data points in the neural recordings to compute the predicted arm trajectory. In reality, the firing rates of the neurons are very low in the motor cortex. This puts constriants on the statistical sampling size problem to extract enough information about the intended trajectory to drive the robot arm. Ingenious methods using adaptive predictive algorithms can overcome these limitations.

Adaptive algorithm is used to "learn" from prior movement trials to adjust for the unknown coefficients of the parameters needed to generate the predictive movement trajectory.
Predictive algorithm is used to "project" what the future neural code would be before the actual code are filled in from the neural recordings. This allows us to overcome the incomplete data sample problem.
Adaptive-corrective algorithm allows us to correct the on-going errors from the estimations, and adjust the predictive algorithm accordingly.
See publication: Tam, D. C. (2003) Real-Time Estimation of Predictive Firing Rate. Neurocomputing, 52-54: 637-641. [Reprint.pdf]