Networking properties of primary motor cortical neurons can predict decoding performance of upper limb movements in advance

Abstract

Brain-machine interfaces (BMIs) for upper limb movement restoration rely on motor cortical circuits coordinating complex arm movements. As BMI decoding performance can vary across days or sessions due to the variation of the recordings of motor cortical neurons, it is important to understand how decoding performance changes with certain properties of the ensemble of neurons. Generally, it is plausible to predict decoding performance by assessing how well neuronal activities encode kinematic information, which requires both neuronal and kinematic data. Yet, it is desirable to estimate decoding performance even before a session starts with no kinematic information so that one can flexibly adjust decoder types or the amount of training in advance. To this end, we examined the networking properties of primary motor cortical (M1) neurons during a pre-session period, where the subject did not begin an arm movement task, to find if these could provide predictive information about decoding performance on the same day (or session). We analyzed networking patterns within M1 neurons of a primate (a Rhesus macaque) during the pre-session periods of seventeen different sessions. The linear Kalman filter was used to decode M1 ensemble firing rates into 3D velocity, and decoding accuracy was calculated as correlation coefficient between predicted and actual movements. We adopted the graph theory to quantitatively assess networking properties. As a result, we found that M1 networking properties could predict across-session performance of 3D velocity decoding. Specifically, an increase in the global clustering coefficient of M1 network led an increase in decoding performance. The result suggests that the M1 networking properties before the session starts can predict decoding performance in BMIs to restore upper limb control.