An Analysis of Machine Learning Methods for Decoding the Direction of Movement Using Neural Data from the Motor Cortex of a Non-Human Primate

Abstract

Parkinsons, arthritis, and tremors are motor complications many people face as they become older. While advanced prosthetics can substantially mitigate these issues, many require both complex surgeries and a lot of time to get used to. To facilitate the implementation of prosthetics, we present several computational methods for performing neural decoding from a population of neurons in the motor cortex of a non-human primate. We first analyzed the accuracy of traditional methods, specifically the population vector and maximum likelihood estimation (MLE); then we developed and compared different machine-learning models, such as feedforward neural networks, random forest classifiers, and support vector machines (SVMs), and compared their accuracies to those of the more traditional methods. These findings shall help accurately decode the direction of movement in a mammal, which could be used in developing brain-computer interfaces and prosthetics.