K-medoid clustering of premotor firing patterns supports fine decoding of macaque reach-and-grasp

Cortical control of reach-and-grasp movements is one of the key challenges of brain machine interfaces (BMI). Recent advancements enabled tetraplegic subjects to control robotic arms through this approach. However, the fine control of such movements is usually automatic rather than controlled by the patient, and this might not be the optimal solution in real life environments. Here we show that fine properties of reach-and-grasp movement can be decoded thanks to K-medoid clustering. We recorded the activity of 71 neurons from the premotor cortex of a macaque performing different reach-and-grasp movements depending on the shape and the position of the object presented. The K-medoid algorithm selected in an unsupervised way four firing patterns associated with specific clusters of neurons' activity. Describing the cortical activity during each movement through the distances from these four medoids and feeding the resulting vector into a support vector machine (SVM) algorithm enabled the identification of the specific movement performed. This decoding approach could in future support a finer control of robotic arms - or other devices - through cortical BMI.