Evaluation of non-stationary signal processing methods for binary EEG classification

Sammanfattning: Electroencephalogram (EEG) measurements are notoriously noisy and non-stationary and there are several specialized techniques for their analysis and interpretation. In this thesis, we implement a collection of stationary and non-stationary methods including coherence, Phase Locking Value (PLV), Phase Lag Index (PLI), and their imaginary counterparts. In particular, we use the Singular Spectrum Decomposition (SSD) algorithm to decompose each recording into interpretable components before computing a variation of the PLV. All methods are evaluated on simulated EEG data in relation to two research questions; one, how well do they manage to detect whether a subject is presented with a stimulus, and two, given that an auditory stimulus is present in one of the subject's two ears, how well can they determine the side. To measure performance, we train three classification algorithms on features extracted from the above-mentioned methods. We find that the imaginary coherence and imaginary PLV are the best predictors for answering research question two by estimating the sign of the phase difference, whereas the SSD algorithm yields the most important feature for stimulus detection. Lastly, we apply our methods to two sets of real EEG data where it is confirmed that imaginary coherence counteracts volume conduction. In addition, all classification algorithms perform more or less the same but the best one manages to predict the presence of an auditory stimulus with 68.7% accuracy, and the side that the stimulus originated in with 55.1% accuracy.