The Detection of Ischemic Stroke on the PSD Manifold of EEG Signals

Abstract

The study of ischemic brain stroke detection by Electroencephalography (EEG) signal is the area of binary signal classification. In general, this involves extracting features from EEG signal on which the classification is performed. In this thesis, we investi- gate the employment of Power Spectral Density (PSD) matrix, which contains not only power spectrum contents of each signal which complies with what clinical experts use in their visual judgement of EEG signals, but also cross-correlation between multi-channel (electrodes) signals to be studied, as a feature in signal classification. Since the PSD matrices are structurally constrained, they form a manifold in the signal space. Thus, the commonly used Euclidean distance to measure the similarity/dissimilarity between two PSD matrices are not informative or accurate. Riemannian Distance (RD), which measures distance along the surface of the manifold, should be employed to give more meaningful measurements. Furthermore, two classification methods, binary hypothesis testing and K-Nearest Neighbors (KNN), are applied. In order to enhance the detec- tion performance, algorithms to find optimum weighting matrix for each classifier are also applied. Experimental results show that the performance by the kNN method us- ing PSD matrix as features with RD as similarity/dissimilarity measurements are very encouraging.