Dimensionality reduction for EMG prediction of upper-limb activity in freely-behaving primates

Neural prosthetic systems aim to assist patients suffering from sensory, motor and other disabilities by translating neural brain activity into control signals for assistive devices, such as computers and robotics prostheses, or by restoring muscle contraction through functional electrical stimulation (FES). In a neuro-motor prosthetic device, the prediction of intended muscle activity is required for effective FES. It has been already known that upper-limb electromyogram (EMG) signals in primates, can be accurately predicted during repetitive tasks, by decoding the spiking-activity (SA) of single cells and multi-unit activity (MUA) in the motor cortical areas. Recent work now suggests that EMG signals can also be decoded by local field potential (LFP) recordings from the same areas. In such decoding schemes, the number of input variables is usually very large and no systematic way of performing effective variable selection has yet been suggested. In this work, we demonstrated for the first time that muscle activity decoding from SA and LFP signals in the primary motor cortex (M1) and the ventral premotor cortex (PMv) areas is feasible during naturalistic free behaviour, and we compared the decoding performance of spike-, LFP- and hybrid decoders. We also tested the relative information in a number of LFP frequency bands, and found that mid-high and high-frequency bands (70 - 244 Hz) conveyed the most EMG-related information. Finally, we compared the decoding performance of a group of sparse regression algorithms, and we showed that a method based on the variational bayes (VB) outperformed the conventional Wiener cascade filter for LFP-decoders in the case of limited amount of training data. For longer training datasets, the results from all methods were comparable