Movement direction decoding with spatial patterns of local field potentials


We show that movement direction can be decoded with high accuracy using the spatial patterns extracted from multichannel local field potentials (LFPs). Two monkeys were trained to execute center-out movement in 8 directions. During the task the LFP activity was recorded with two 64 channel grids from the pre- and primary motor areas. The LFP signals were decomposed into 4 sub-band components in the 0-4 Hz, 4-10 Hz, 14-30 Hz and 48-200 Hz frequency ranges. The sub-band activity was post processed with regularized common spatial patterns algorithm and fed to linear discriminant analysis for final classification. Directions of movement were estimated using a redundant hierarchical classification strategy that tested groups of directions against diametrically opposite groups. The grouping of directions was based on the spatial correlation that we observed between LFP signals corresponding to neighboring movement directions which is similar to the cosine tuning profile of single neurons. We found that the decoding power for 8 directions was 80% and 92% for the two subjects, respectively, in 0-4 Hz frequency band. Our best result of 92% nearly doubles the accuracy of the best results reported in the literature with similar set-ups. These results indicate that spatial patterns in LFP can be used to construct high accuracy brain computer interfaces.