Abstract
This report describes progress toward developing a brain-computer interface for decoding human navigational movements from scalp electrode recordings. In our paradigm, subjects (n=11) sat in a swivel chair, viewed a virtual environment displaying non-spatial information specifying a right or leftward navigational turn, and rotated their body by pushing against the floor with their feet. Signals monitoring chair rotation, electroencephalographic (EEG) activity, and instructed direction were recorded. Our first goal was to apply legacy methods previously used for synchronous decoding of the direction of planar arm movements to decoding the direction of body turning movements. The legacy method 1) extracted 34,020 features from data epochs of -.5 to +1.5 seconds relative to turn specification using combinations of EEG electrodes, band powers, and sliding windows of different sizes, 2) selected the best 100 features using linear discriminant analysis, 3) classified direction using a support vector machine, and 4) computed classification accuracy on movements not used for building the classifier. The average accuracy for decoding left-right movement direction was 82%. Our next goal was to apply a more theoretically motivated information gain criterion for selecting the best features from the same set of extracted features. We found that classification accuracy asymptoted at 86% with an average of 255 features per subject. We then used the information gain criterion to investigate how to reduce the size of the necessary feature set without sacrificing classification accuracy. We found that EEG mean amplitude, a .5-second window size, and the electrodes in the right frontal quadrant carried more information than any other band power, window size, or quadrant. We then extracted this reduced set of 13 features from just the .5-second epoch preceding movement onset and found a decoding accuracy of 78%. This is the first time navigational turning movements have been decoded from surface EEG. The introduction of an information gain criterion for feature selection has advanced our ability to speed data processing by limiting the number of features needed to achieve classification of movement intent with pre-movement data, both of which are essential steps in moving toward a real time BCI application.