Factor analyzed hidden Markov models for estimating prosthetic limb motions using premotor cortical ensembles

Abstract

Research is underway to develop neural control of prosthetic limbs. Here we propose a quantitative framework based on factor analyzed hidden Markov models (HMM) to estimate the limb motion states from cortical neuron ensembles. Limb motion states are the movement steps in the execution of a behavioral task including baseline, pre-movement planning, movement execution, and final fixation on the target peripheral object. In order to model complex motion states, we use neural recordings from ventral premotor (PMv) and dorsal premotor (PMd) neurons in a non-human primate executing instructed reach-to-grasp behavioral tasks following visual cues including pushing a button, pulling a mallet, grasping a sphere and pulling a cylinder. We estimate a factor analyzed HMM to represent the motion states, which are also called as epochs, between baseline, pre-movement planning, movement execution, and final fixation on the target peripheral object. As an extension of standard HMMs, a factor analyzed HMM has a continuous hidden layer besides the common discrete hidden layers as seen in HMMs. The continuous hidden layer is composed of a low-dimensional representation of observations obtained via the factor analysis. We find that not only our framework can achieve high decoding accuracies for different epochs of four different behavioral tasks, namely, 0.88 (±0.006) for the 1st epoch, 0.96 (±0.002) for the 2nd epoch, 0.79 (±0.015) for the 3rd epoch, and 0.89 (±0.005) for the 4th epoch, it can also estimate the latencies between epoch transitions (<150 ms). Our framework may be useful in neural decoding complex movements of prosthetic limbs.