Decoding Neuronal Activity Using a Deep Neural Network to Predict Knob Supination Success

Abstract

Intracortical microelectrode arrays enable the acquisition of signals generated from individual neurons, offering a level of specificity and granularity that is highly useful in many applications. For example, these signals can be used as input to decoding algorithms which aim to estimate intent during volitional motor tasks in order to provide control signals for neuroprostheses. Previous studies have used a variety of linear and nonlinear decoding algorithms to predict fine movement of the upper extremity using cortical activity recorded from nonhuman primates and humans. Other studies have utilized muscle activity to decode wrist flexion and extension in the same animal models, however, the use of cortical signals to decode wrist supination has not been well explored and may hold clinical significance. In this study, we demonstrate the feasibility of decoding neuronal activity recorded from the motor cortex of rats to predict success during a knob supination task, i.e. wrist rotation past a 60° threshold. A deep neural network was trained using the firing rate of individual neurons during forelimb supination and was able to predict the outcome of individual trials with high accuracy. This algorithm was specific to each rat, meaning its prediction capacity was not generalizable across all subjects. These results suggest that the rat model is a viable means of developing and testing brain-machine applications for skilled motor tasks.