Kinematics-Dependent Neural Representations from Primary Motor Cortical Neural Population Activity for Motor Decoding

Abstract

While much of the recent study on intracortical brain-machine interfaces (BMIs) seeks to improve the quality of life of people with tetraplegia, it is still challenging to achieve high and stable performance for controlling external devices in clinical BMIs. Neural decoders used in BMIs for motor control infer kinematic information from measured firing activities, usually in the primary motor cortex (M1). One of the key approaches to enable motor decoding from M1 neurons is population coding. Although population coding provides fundamental insights into the instantaneous representation of kinematic variables in M1, clear kinematic information that allows the devices to keep performance is required to enable a decoder to produce more natural limb movements. To this end, I first need to understand the structure and kinematic tuning properties of the neuronal ensemble and find neural representations that more clearly represents motor information to improve decoding performance. As latent factors representing the neural states have recently emerged as one of the key interests in neural decoding, numerous studies have been conducted to discover neural representations that enable effective motor decoding. Although these neural representations could reflect neural states well for neural population activity, there is no guarantee that they clearly contain motor information. In this dissertation, therefore, I compare traditional decoding methods. Then, I suggest a new method as following order. First, I discuss the effects of neuronal ensemble properties on two types of motor decoders: a population vector algorithm (PVA) by linear basis function following the assumption of linearly independent in single neurons; an optimal linear estimator (OLE) and a linear Kalman filter (LKF) considering neural population activity jointly. This study emphasizes the importance of considering neural population activity jointly rather than a single neuron for motor decoding. Second, I introduce an approach to finding neural representations from neural population activity by jointly estimating multiple pairs of canonical variables with neural population activity and kinematic variables. This study only covers the effects of canonical variables on motor decoding. That is, I investigate whether our approach based on supervised learning is proper as a neural representation for effective motor decoding. This study highlights that supervised learning is more effective than unsupervised learning in terms of motor decoding. Third, based on the first and second studies, I propose a new approach to extracting kinematics-dependent neural representations from neural population activity. I first define kinematics-dependent latent components using a linear model and then estimate these components from neural population activity through a nonlinear mapping. This study shows that the proposed approach yields neural trajectories that clearly show changes in neural states before and after motion onset and yields higher decoding performance than counterparts. Our approach to improving neural state estimates by modeling kinematics-dependent latent components may enhance the stability of practical BMIs with noise. It may also be able to explicitly track neural states that describe various types of kinematic parameters. Taken together, kinematics-dependent neural representations have the potential to help improve decoding performance in clinical BMIs and to be employed towards diverse neurophysiological studies.