Advanced Signal Processing and Machine Learning Approaches for EEG Analysis

Abstract : Electroencephalography (EEG) offers a non-invasive brain-imaging technology with potential to extract user intent from brain signals. This can offer a potential method for dispersed soldiers to communicate silently with one another. The usual interface for acquiring EEG signals may house 128 or more electrodes. Each EEG signal may be sampled at KHz sampling rates and may last for a few seconds. Thus the number of samples used to represent each trial can be large. The goal of this short-term innovative research (STIR) project was to investigate innovative sample and channel (i.e., EEG electrode) selection methods to reduce the storage and computational complexity in analyzing EEG signals. In experiments aimed at determining the redundancy in imagined speech EEG signals, it was observed that EEG data has limited spatial redundancy, but large temporal redundancy. In another set of experiments, we investigated the classification of two imagined speech syllables (namely "Ba" and "Ku") from imagined speech EEG signals. Using all good channels, only one of the seven volunteer subjects produced "better than chance" classification accuracy of about 60%. By selecting specific electrodes, two subjects yielded better-than-chance results with recognition rates close to 60% for all trials. Overall classification rates appear to have improved slightly by the selection of electrodes, indicating that imagined speech classification performance can be improved by careful selection of EEG electrodes.