Advanced Feature Extraction for Classification of Long-Term Epileptic Electroencephalography Records
|Tila||Julkaistu - 10 toukokuuta 2019|
|Nimi||Tampere University Dissertations|
First, a signal decomposition technique, applicable to physiological signals, is devised which can be used as a preliminary step for feature extraction. This is performed by proposing a novel time-frequency transform based on rational functions, namely, rational short time Fourier transform (RSTFT). A sparse decomposition method is then proposed by reconstructing the input signal into several components using non-overlapping sub-sets of the RSTFT coeﬃcients. Sparse representation of signal components is then obtained by inducing L1 regularization penalty on the RSTFT coeﬃcients during the reconstruction phase. The eﬀectiveness of the proposed sparse decomposition method is evaluated in the classiﬁcation of long-term EEG records for purposes of epileptic seizure detection and sleep stage scoring.
Another part of the thesis investigates the detection of seizures in textural images constructed using a proposed scheme for mapping of EEG signals into gray-scale image domain. The proposed mapping strategy makes it plausible to correlate textural changes of the obtained images with seizure activities in EEG records. The textural analysis is then carried out using the well-known gray-level cooccurrence matrix (GLCM) and Haralick’s feature extraction method resulting in a compact representation of EEG epochs. All the methods proposed in this thesis are evaluated using public EEG data-sets freely available on-line. The results obtained by the proposed methods are comprehensively compared with the other conventional dedicated feature extraction techniques using several classiﬁers.
The main contribution of the thesis is in the adaptation of conventional feature extraction techniques, commonly used in the textural analysis of images, to be applicable in EEG signal analysis. Additionally, the discriminatory power of feature descriptors is improved by representing EEG signals using their sparse rational components. The proposed rational local Gabor binary pattern (LGBP)-width feature outperforms competing methods in both seizure detection and classiﬁcation problems. Moreover, its perform consistency in patient/non-patient speciﬁc scenarios demonstrates its ability to tackle the curse of variability in seizures.