Classification of EEG Data Using Convolutional Neural Networks and the Scaled Reassigned Spectrogram

Sammanfattning: Electroencephalography (EEG) is a medical technique for measuring brain activity through several channels connected to the scalp. Interpreting EEG data is a difficult problem because of the large amount of noise contained in the data. Using spectral methods on EEG data can improve the ability to interpret the data, especially using a time-frequency method called the scaled reassigned spectrogram which has been shown to perform well on data with similar properties to a model containing Gaussian envelope transients. A technique for classification of data transformed by time-frequency methods is to use convolutional neural networks (CNN), which are known to be successful at image classification. In this thesis, spectral methods and CNNs are combined for use on EEG data in order to identify the location of a sound, whether a person hears the sound in the left or in the right ear. The best classification results obtained in this thesis were for a single channel near the right ear without transforming the data at 60.13%, and using singular value decomposition (SVD) on four channels near each ear and the scaled reassigned spectrogram with a result of 59.47%. These are both significant results.