Abstract
To correlate the subject's electrical brain activity and his overt behaviour has potential applications on human performance monitoring and brain computer interfacing. Classification of subject's signals produced during the execution of same task is a particular problem which presents high difficulty regarding signal discrimination. This problem has been studied only by few groups during the last decade [55], [56]. Most of the reported research works regarding classification of brain signals are aimed to differentiate signals produced during the execution of different kinds of tasks. The diversity of assumed statistical and physiological models for representation of EEC activity has produced a diversity of results, some of them with a questionable reliability [17], [29]. Therefore, the problem of correlating EEC signals with the subsequent responses was approached in this project by using just general assumptions about the data. I applied subspace projection methods of pattern recognition for predicting the reaction times of subjects' responses during the execution of a target detection task in a cognitive psychology experiment. In order to correlate the reaction time of the response associated with an unknown signal, a two-class classification problem was defined. Five distinct classifiers were designed and tested. The first three classifiers were principally based on the the eigenvectors of the class correlation matrices. The other two classifiers were based on orthogonal subspace modelling of the classes. I carried out two sets of experiments. In the first set, general classifiers were designed with the pattern vectors of all subjects simultaneously. However, the prediction rates were only slightly over the random allocation of the unknown patterns I wished to classify. These results showed that general purpose classifiers concerning all subjects treated together are not feasible. The second set of experiments concerned classifiers fitted for each subject separately Two different methods for training and testing the classifiers were applied. The first one was the so-called "matched halving" method. The second procedure applied was the leaving one-out method. Both methods produced similar classification rates between 60% and 70% of correct results. However the earlier method showed better performance than the later one regarding earlier prediction of the reaction time and lower dimensionality of the pattern vectors. In addition, a correlation was found between the spatial allocation of electrode and best results. By using registers from electrode position at the frontal section at midline of the head best results were produced for 7 out of the 11 subjects. Key words: Event-related potentials, single trial classification, subspace projection methods.