Spatial proximity based subspace decomposition for movement direction decoding of Local Field Potentials

Local Field Potentials (LFP) provides higher spatial resolution and SNR than EEG data and can be used to construct a Brain Computer Interface. In, we have shown that movement direction decoding can be done with about 90 % classification accuracy using spatial patterns (CSP) and Error Correction Output Codes (ECOC). However, a major challenge in this study is to make this method more robust to inter-session variability of the LFP data, where state-of-the-art results are in the high 70 percent. In, we have demonstrated that LFP features that are recurrent across sessions can be extracted using a subspace learning method and used to improve the CSP +ECOC classifier. In this work, we propose an extension of the subspace learning method that exploits the spatial topology of the channels. This allows us to learn spatially diverse features, while previously the subspaces were being learned independently of the channel layout. We proposed a method where a block of samples from neighboring channels is used to find the subspaces and decode the directions. This approach is analogous to analyzing an 8×8 pixel map in image processing. Furthermore, this method allows a spatio-temporal classification, and it is indeed observed that different directions were providing higher accuracies at different time blocks. The proposed method can boosts the accuracy by at least 6% to bring classification to the mid 80 percent. Furthermore, we show early results where adding a pilot trial from the test session can be used as a calibration to further improve the spatio-temporal classification.