Designing a Multi Trial Classifier for EEG Signals: Classifying Rhythms Perceived

Classification of EEG-signals is error-prone, due to the the small differences in the measurements and the inherent presence of continuing brain dynamics. Often these dynamics are denoted as noise, and in this view, classification is difficult, due to the small signal to noise ratio. We investigate how multiple trials of our EEG data can be used to increase the classification rate. Two schemes are used to combine the measurements: i) combine probabilities of individual classifications; ii) average measurements before classification. The number of trials that was used was either fixed or flexible. Flexible means that as many trials as needed are used to get a certain reliability of the classification. It is found that combining probabilities works best for large variances and a flexible number of samples, and averaging a flexible number of measurements works best for small variance. The validation of the method is tested on an EEG data set in which a subject listened to two different rhythms. On a single trial, the classification rate was 80%, a classification rate of about 90% was achieved using the average of 2 trials, and a classification rate of approximately 95% was found for 3 trials. This coincided well with the predictions.