Abstract:

Since EEG signals have individual difference and are vulnerable to noise and artifacts, we propose an independent component analysis (ICA)-based method for the selection of optimal feature channels. This method applies the ICA to decompose channels’ data to N200, P300, ocular artifacts and other physiological signals. Whether a channel is suitable for feature extraction is decided by the influence of those signals that mentioned above to this channel. We apply our method and three other commonly used methods for feature channel selection to twelve subjects’ brain signals, and recognize N200 and P300 potentials. We find that our method achieves a 93.10% accuracy on average and it is 7.27%, 1.07% and 75.96% higher than the average accuracy of the other three methods respectively. We fit a relation curve between ICA weight and channel selection threshold based on the least square method, and obtain a high classification accuracy when predicting the optimal channels and recognizing the potentials from another three new subjects' data, which means that this prediction method has universality.

Key words: channel selection, independent component analysis(ICA), individual difference, artifact, classification accuracy