Stockwell transform for epileptic seizure detection from EEG signals. (September 2017)

• Record Type: Journal Article
• Authors: Kalbkhani, Hashem; Shayesteh, Mahrokh G.
• Is Part Of: Biomedical signal processing and control. Volume 38(2017)
• Page Start: 108
• Page End: 118
• Publication Date: 2017-09
• Subjects: EEG classification -- Epileptic seizure -- Stockwell transform (ST) -- Amplitude distribution -- Kernel principal component analysis (KPCA); Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2017.05.008
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Time-frequency Stockwell transform (ST) is used for epileptic seizure detection. Feature vector is constructed based on distribution of amplitudes of ST in different sub-bands. Different frequency sub-bands have different accuracies and their combination achieves the highest performance. Nearest neighbor classifier with Chebychev distance achieves highest accuracy among different distances. Proposed method outperforms other methods including discrete wavelet transform and discrete Fourier transform. Abstract: Epilepsy is the most common disorder of human brain. The goal of this paper is to present a new method for classification of epileptic phases based on the sub-bands of electroencephalogram (EEG) signals obtained from the Stockwell transform (ST). ST is a time-frequency analysis that not only covers the advantages of both short-time Fourier transform (FT) and wavelet transform (WT), but also overcomes their shortcomings. In the proposed method, at first, EEG signal is transformed into time-frequency domain using ST and all operations are performed in the new domain. Then, the amplitudes of ST in five sub-bands, namely delta (δ), theta (θ), alpha (α), beta (β), and gamma (γ), are computed. In order to classify EEG signal as healthy, interictal, and ictal, we obtain the distributions of amplitudes of ST in different sub-bands. In this way, for each EEG signal, five feature vectors, each for one sub-band are obtained. Next, kernel principal component analysis (KPCA) is used to extract the informative features from the feature vectors. Finally, the distances between the informative features of test and training samples in different sub-bands are calculated and the weighted linear combination of them is applied to the nearest neighbor classifier. We consider different distance measures. The test sample is assigned to the class of the training sample which has minimum distance from it. The results demonstrate that the proposed method achieves higher efficiency in comparison with the recently proposed algorithms.