Integrating neural spiking and LFP activity to decode kinematics of the arm and hand during unconstrained reach to grasp movements

Many current brain-machine interfaces do not consider the behavioral context of the subject, rather, they assume the subject is constantly engaged in a single task. We investigated how incorporating information about state can improve the performance of a decoder. Unit spiking activity and LFPs were recorded from chronically implanted electrode arrays in primary motor and premotor corticies while a monkey performed a reach to grasp task. We applied an unsupervised clustering technique to LFP data to identify different neural states. Then, for each state, we fit a sparse Bayesian linear model with causal interaction terms to decode the joint kinematics of many degrees of freedom in the arm and hand. We used automatic relevance determination for variable selection and to avoid overfitting. We show that the state-based decoding model improves decoding performance over a model without state information. We further show that topology of interaction terms is different across different states.