Performance prediction using EEG and trial-invariant characteristic signals.

One of the most important parts of all applications trying to discriminate between a person's different mental tasks using their recorded EEG data is the process of feature construction. A common practice for this is to exploit an apriori knowledge about the nature of the mental processes of interest and their impact on the EEG signals. However, the use of features constructed in this way is restricted to applications concerning the corresponding mental processes. We present here a novel method for EEG data classification which is very general as it makes no assumptions about the nature of the EEG signals. It is based on the construction of a characteristic signal for each class which remains as invariant as possible over the trials belonging to that class. We use the proposed method in combination with a novel method for channel selection in an oddball experiment to predict a person's quick or late response.