Abstract
Brain Computer Interface applications require single-trial decoding of brain activity. Electroencephalography (EEG) holds promise for such applications, since it is non-invasive and has high temporal resolution. However, due to low signal-to-noise ratio, classification of intentions on single- trial EEG with high accuracy has been a challenge. In this work we consider a task, where the subjects had to press a right (left) hand button rapidly and accurately in response to a right (left) arrow cue. Our goal was to examine if single- trial data from this experiment contained sufficient informa- tion to infer the type of intended movement. If so, could we design a classifier that would work well across different subjects? Does the EEG dynamics provide important infor- mation for this task? We have studied a classifier based on Fisher Linear Discriminant (FLD) applied to the pooled EEG data from all subjects. Our first finding is that the performance of FLD classifier that takes as input a single time-point EEG activity vector is relatively poor with classification error of 30 % at the optimal time. Next, we have constructed a classifier that uses the time course of the EEG traces. In addition, we performed dimensionality re- duction on the electrode array by using only the top 15-20 most informative electrodes. This method yielded a classi- fication success rate of 87 %, ranging from 80 % perfor- mance in 'poor' subjects to 95 % for the 'best' subject. For the task studied here, we show that single-trial EEG data contains sufficient information to perform an accurate bina- ry classification of intended movement. Interestingly, it was not necessary to build a separate classifier for each subject. On the other hand, information about the EEG dynamics was crucial for good performance. These results may bear important implications for Brain-Computer Interface appli- cations, as well as for understanding the nature of the EEG signaling of movement preparation.
