Improving the performance of online classifier by removing the error samples from offline training data

Brain-computer interface (BCI) plays an important role in helping the people with severe motor disability. In event-related potential (ERP) based BCIs, subjects were asked to count the target stimulus in the offline experiment, the recorded electroencephalogram (EEG) data was used to train the classification mode. However, subjects may make mistakes in counting the target stimulus or be affected by the non-target stimulus. The target trials may not contain expected ERPs and the non-target trials may contain unexpected ERPs, which was called error samples. This paper intends to survey whether the classification accuracy could be improved after removing these error samples from offline training data. The result showed that the online performance of BCI system could be improved after selecting the offline samples for training the classification mode.