Classification of EEG data using machine learning techniques

Abstract: Automatic interpretation of reading from the brain could allow for many interesting applications including movement of prosthetic limbs and more seamless manmachine interaction. This work studied classification of EEG signals used in a study of memory. The goal was to evaluate the performance of the state of the art algorithms. A secondary goal was to try to improve upon the result of a method that was used in a study similar to the one used in this work. For the experiment, the signals were transformed into the frequency domain and their magnitudes were used as features. A subset of these features was then selected and fed into a support vector machine classifier. The first part of this work tried to improve the selection of features that was used to discriminate between different memory categories. The second part investigated the uses of time series as features instead of time points. Two feature selection methods, genetic algorithm and correlation-based, were implemented and tested. Both of them performed worse than the baseline ANOVA method. The time series classifier also performed worse than the standard classifier. However, experiments showed that there was information to gain by using the time series, motivating more advanced methods to be explored. Both the results achieved by this thesis and in other work are above chance. However, high accuracies can only be achieved at the cost of long delays and few output alternatives. This limits the information that can be extracted from the EEG sensor and its usability.