Fourier-Bessel domain based new methods for automated classification of EEG signals

Abstract

A brain is a very complex organ in the human body. It consists of around 100 billion neurons. These neurons are responsible to generate electric potential inside and on the surface of the brain. The recording of electroencephalogram (EEG) signals is the most common and economical way for neurologists to observe the electrical activity from the brain. In particular, the EEG signals play a vital role in the diagnosis of various brain-related disorders and brain-computer interface (BCI) applications. Since these signals are also contaminated with noise and artifacts. Hence, adequate and accurate analysis of these signals is always a major challenge among researchers. The development of advanced signal processing based methods can be helpful for the analysis and classification of these signals. In this thesis, we have developed Fourier-Bessel series expansion (FBSE) based methods for analysis and classification of EEG signals. The classification of epileptic seizures EEG signals based on weighted multiscale Renyi permutation entropy (WMRPE) and FBSE based rhythms has been proposed first. In this method, the FBSE has been used to obtain the rhythms from the EEG signal and the WMRPE feature has been extracted from these rhythms. These computed feature values are then used in distinct classifiers such as random forest (RF), least squares support vector machine (LS-SVM), and regression for classifi cation using the 10-fold cross-validation technique. The classification performance has been optimized with feature ranking methods. The proposed method has been also tested for different signal to noise ratio (SNR) levels. In the second method, an automated focal epileptic EEG signals classification method has been developed with the help of FBSE based flexible time-frequency coverage wavelet transform. The features such as mixture correntropy and expo nential energy have been extracted with sub-band signals of each EEG signal. The classification task has been performed on these extracted features using LS-SVM classifier with a 10-fold cross-validation technique. The developed automated classi fication method has been optimized with probability (p)-value based feature ranking method and also tested for different SNR levels. The third method is based on the proposed Fourier-Bessel decomposition method (FBDM). The proposed FBDM decomposes the non-stationary signal into a finite number of Fourier-Bessel intrinsic band functions (FBIBFs). In addition to FBDM, we have also proposed a zero-phase filter-bank based FBDM to get fix number of FBIBFs in this work. The performance of the proposed FBDM has been evaluated with the help of Poverall quantitative measure and time-frequency representation (TFR) analysis of synthesized signals. The developed FBDM has been used for the classification of six different sleep stages using EEG signals. The convolutional neu ral network (CNN) classifier has been utilized for the classification of TFR images, which were obtained with the application of FBDM on a publicly available sleep EEG signals database. In our fourth method, we have proposed a Fourier-Bessel dictionary based spa tiotemporal sparse Bayesian learning (SSBL) algorithm with expectation maximiza tion (EM) method for efficient reconstruction of multichannel EEG signals. In this work, we have used the Fourier-Bessel dictionary with the SSBL-EM method as the Bessel functions have non-stationary characteristics and provide good compression for EEG signals. This proposed method has been tested for the steady-state visual evoked potential (SSVEP) based BCI classification using multichannel EEG signals and it provides 100% classification accuracy with 50% compression and 2 s time du ration. In this method, the classification has been performed with the L1-regularized multiway canonical correlation analysis (L1-RMCCA) method. The fifth method introduces a new method based on FBSE and CNN classifier for automated classification of schizophrenia EEG signals. The FBSE has been applied in a different way on EEG signals that provides a frequency matrix for each EEG signal. These frequency matrices have been utilized in CNN classifier for the classification of schizophrenia EEG signals. The proposed method achieves higher results in the classification of schizophrenia EEG signals and also exhibits the power to use it for the practical application of various diseases.