Automatic epileptic EEG detection using DT-CWT-based non-linear features. (April 2017)

• Record Type: Journal Article
• Authors: Li, Mingyang; Chen, Wanzhong; Zhang, Tao
• Is Part Of: Biomedical signal processing and control. Volume 34(2017)
• Page Start: 114
• Page End: 125
• Publication Date: 2017-04
• Subjects: Epilepsy -- EEG -- DT-CWT -- Non-linear features; 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.2017.01.010
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The DT-CWT was explored to obtain five components of EEG instead of DWT in this paper. Three non-liner features were quantified in the form of Hurst exponent (H), Fractal Dimension (FD) and Permutation Entropy (PE). The influences of different types of filters used in DT-CWT have been considered in this work. In order to make the conclusion more convincing, DWT configured with four types of wavelet functions was employed as comparison. Abstract: The epilepsy is a type of common neurological disorder plaguing many people around the world. A novel method based on the dual-tree complex wavelet transform (DT-CWT), in this study, is proposed to develop a reliable diagnosis method for the epileptic EEG detection. We explore the ability of DT-CWT to decompose the original EEG into five constituent sub-bands, which are associated with non-linear features such as the Hurst exponent (H), Fractal Dimension (FD) and Permutation Entropy (PE). Furthermore, influences of different filter types on the DT-CWT are considered in this study as well. With these features, the support vector machine (SVM) configured with filters of the near-symmetric 13/19 tap filters (NS 13/19) and Q-shift 14/14 tap filters (QS 14/14) is found to achieve the preferable classification accuracy of 98.87%, which is visibly higher than that with discrete wavelet transform (DWT)-based features. Results demonstrate that the technique proposed by us can not only provide significant performance with less computational cost but also can implement simply. It will be a potential method for practical applications extended to the development of a real-time brain monitoring system.