Unsupervised identification of stereotypical premotor firing patterns for the decoding of hand and mouth movements

Cortical decoding of motor intentions is a classic challenge in neural engineering. A candidate solution for tetraplegic patients is to decode motor intention from the activity of the frontal cortex. Here, we develop an algorithm to decode motor intentions from electrophysiological multi-single unit recordings. The algorithm identifies stereotypical spiking patterns for each of the recorded neurons and uses them as reference points to express and decode neuronal activity on different tasks. We tested the algorithm on recordings from the premotor cortex of two macaques performing different types of hand grasps and mouth movements associated with rewards. In the time window of highest performance, the decoder achieved average accuracies of 96%, 82%, and 67% in discriminating 2, 4, and 7 movement classes, respectively. Additionally, movement classes were decoded significantly above chance already in the window of 1000 to 500 ms before the grasp started, demonstrating decoding of motor intentions and the potential applicability of the approach for neuroprosthetic applications.