Advanced wavelet transforms based EEG signal processing methods for epilepsy diagnosis

Abstract

The human brain is highly complex in nature. The electroencephalogram (EEG) signals are generated due to the continuous ring of neurons inside the brain. Generally, neurologists perform manual monitoring of EEG signals in order to analyze the pathological and physiological states of the brain. However, the manual monitoring of long duration EEG signals is tedious and time consuming task and may be prone to human errors. In addition, the recorded EEG signals may be contaminated with background noise, and muscle artifacts. Besides monitoring, the accurate localization of epileptiform discharges is also considered to be very crucial in the diagnosis of patients undergoing resective epileptic surgery. For this purpose, stereo EEG (SEEG) is considered as the gold standard consisting of electrode shafts implanted within the brain volume. Apart from this purely clinical purpose, the SEEG signals can also be used to explore the complex brain responses in the presence of provoked electrical stimulations. In particular, cortical stimulation (CS) o ers a complementary tool to investigate the lesioned brain networks and to identify the epileptic foci with precision. In this perspective, advanced signal processing based methods may be useful for real time monitoring and analysis of EEG signals. A new method for the classi cation of epileptic seizure EEG signals is proposed. The Hilbert marginal spectrum (HMS) has been derived in the Empirical wavelet transform (EWT) domain. The computed HMS is compared with conventional Fourier spectrum for a synthetically generated multi-component frequency modulated signal. The energy and entropy features are extracted from these HMSs lying in separate oscillatory levels and ranked based on probability (p) values. Finally, the seizure and seizure-free EEG signals are classi ed using the random forest (RF) classi er. For the classi cation of seizure, seizure-free and normal EEG signals, the quality factor (Q) based multi-scale entropy measure is proposed. The Q -based entropy (QEn) is found by computing K-nearest neighbor (K-NN) entropies of cumulative subband signals obtained with the tunable-Q wavelet transform (TQWT). Most signi cant features are selected using wrapper-based feature selection method and the support vector machine (SVM) classi er is employed for the classi cation purpose. We have developed a method for epileptic seizure detection in long duration EEG signals. The joint instantaneous amplitudes (JIAs) and joint instantaneous frequencies (JIFs) of the multivariate signals is computed using EWT. The multivariate extension of EWT (MEEWT) has been proposed and studied on both synthetic signal, and real EEG signals. In a moving-window-based analysis, multivariate EEG signal epochs ( ve automatically selected channels) are decomposed and three features are extracted from each JIAs of multivariate EEG signals. A new feature processing step is proposed and joint features are computed in order to enhance the discrimination between seizure and seizure-free epochs. The proposed detection method has been evaluated over 177 hour of EEG records using six classi ers with good performance. An automatic approach has been proposed to detect EEG signals of non-focal and focal groups. The proposed approach can be useful in determining the area linked to the focal epilepsy. The two di erent subsets of publicly available bivariate Bern Barcelona focal and non-focal EEG signals database are studied in this work. In this method, at rst, the rhythms of EEG signals are extracted using EWT. Then, two-dimensional (2D) projections of the reconstructed phase space (RPS) are found for the rhythms. The area of 2D RPS plots are estimated as featuresusing central tendency measure (CTM) parameter. Finally, the least-squares SVM (LS-SVM) classi er is used to classify the focal and non-focal EEG signals. In another work, we have proposed a classi cation method of focal and non-focal EEG signals using TQWT based multivariate sub-band entropy. We have employed TQWT for analysing the subband signals of the analysed multivariate EEG signals and computed multivariate fuzzy entropy (mvFE) of the obtained multivariate sub-bands. The proposed method has been studied on the entire focal and non-focal EEG signals database in order to investigate the statistical signi cance of the proposed features in di erent time segmented signals. Finally, the RF and LS-SVM classi ers are used for classi cation of focal and non-focal EEG signals. We have also distinguished focal and non-focal EEG signals using novel multivariate multiscale spectral entropy (MSSE) features. The proposed multivariate MSSE is based on the application of MEEWT method. The MEEWT based HMS is derived and spectral Shannon entropy is computed in di erent frequency scales based on the selection of successive JIA and JIF functions. The variation of the MSSE values along the time axis is also presented with three dimensional (3D) time-scale entropy plots. The proposed method has successfully discriminated di erent multivariate synthetic noise signals and real focal and non-focal EEG signals. A novel ltering approach is proposed for suppressing the CS artifacts in SEEG signals using time, frequency as well as spatial information. The method is based on the novel combination of spatial ltering and source separation methods with TQWT. The CS artifacts are isolated within very few components/sources after applying the spatial ltering methods on multivariate SEEG signals. The artifacted components are then decomposed into oscillatory background and sharp varying transient signals using TQWT. The CS artifact is assumed to lie in the transient part of the signal. Using prior known time-frequency information of the CS artifacts, we selectively mask the wavelet coe cients of the transient signal and extract out any remaining signi cant electro-physiological activity. We have applied our proposed method of CS artifact suppression on simulated and real SEEG signals with convincing performance. The experimental results indicate the e ectiveness of the proposed approach.