Overcoming measurement time variability in brain machine interface


We introduce a subspace learning approach for multi-channel Local Field Potentials (LFP), and demonstrate its application in movement direction decoding for 8 directions movement. We show that the subspace learning method can effectively address the issue of signal instability across recording sessions by extracting recurrent features from the data. We present results for movement direction decoding, where we trained on two recording sessions, and evaluated decoding performance on a third session. We combine our method with a classifier based on Error-Correcting Output Codes (ECOC) and Common Spatial Patterns (CSP) and found improvement in Decoding Power (DP) from 76% to 88% for a subject known to have strong inter-session variability. Furthermore, we saw an increase from 86% to 90% DP with another subject which exhibited significantly less variability.