Please use this identifier to cite or link to this item:
Title: Automated identification systems based on advanced signal processing techniques applied on EEG signals
Authors: Sharma, Rajeev
Keywords: Electrical Engineering
Issue Date: 17-Feb-2017
Publisher: Discipline of Electrical Engineering, IIT Indore
Series/Report no.: TH62;
Abstract: The human brain is a highly complex system. The electroencephalogram (EEG) signals are commonly analyzed by experts to study the di erent states of the brain in di erent physiological and pathological conditions. The recorded EEG signal is a result of ring of neurons within the brain. In order to study these physiological and pathological conditions manual assessment is performed by the experts. The manual assessment of EEG signal is highly subjective due to the uneven perception of the experts in the interpretation of characteristic changes. Therefore, decision making which is independent of human intervention using automated identi cation system can yield accurate and repeatable results. Therefore, in this thesis, the automatic identi cation systems are developed for analysis of the di erent pathological and physiological conditions of the brain. These identi cation systems are developed using new methodologies which employ advanced signal analysis techniques. Epileptic seizure is the most common disorder of the human brain, which is generally detected using EEG signals. In order to perform epilepsy detection, epileptic seizure EEG signals are distinguished from the normal and seizure-free EEG signals. A methodology for classi cation normal and epileptic seizure EEG signals are is developed based on EMD. The empirical mode decomposition (EMD) is applied on EEG signals. The EMD generates the set of amplitude and frequency modulated components known as intrinsic mode functions (IMFs). In order to distinguish the normal and epileptic seizure EEG signals, two area measures are computed, one for the graph obtained as the analytic signal representation of IMFs in complex plane and another for second-order di erence plot (SODP) of IMFs of EEG signals. These computed area parameters show signi cant di erence for normal and epileptic seizure EEG signals. For the purpose of classi cation of epileptic seizure and seizure-free EEG signals, two-dimensional (2D) and three-dimensional (3D) phase space representations (PSRs) are used. Two measures are de ned namely, 95% con dence ellipse area for 2D PSR and interquartile range (IQR) of the Euclidian distances for 3D PSR of IMFs of EEG signals. These measured parameters show signi cant di erence between epileptic seizure and seizure-free EEG signals. In both classi cation problems least squares support vector machine (LS-SVM) is used as a classi er. The characteristics of the brain area a ected by partial epilepsy can be studied using focal and non-focal EEG signals. In this work, a method for the classi cation of focal and non-focal EEG signals is presented using entropy measures. The entropy features, namely, average Shannon entropy (ShEnAvg), average Renyi's entropy (RenEnAvg), average approximate entropy (ApEnAvg), average sample entropy (SpEnAvg) and average phase entropies (S1Avg and S2Avg), are computed from di erent IMFs of focal and non-focal EEG signals. The LS-SVM classi er is used for classi cation and method is applied for small subset of database. Considering the limitations of this approach like time complexity and smaller dataset, a new approach based on bivariate empirical mode decomposition (BEMD) is presented for computer aided diagnosis of focal EEG signals. This method is studied on entire dataset. The focal and non-focal EEG signals are decomposed using the BEMD, which results in IMFs corresponding to each signal. Then, bivariate bandwidths namely, amplitude bandwidth, precession bandwidth, and deformation bandwidth are computed for each obtained IMF. Interquartile range (IQR) values of bivariate bandwidths of IMFs are employed as the features for classi cation. In order to perform classi cation LS-SVM is used which resulted in signi cant classi cation accuracy. Also, a new method for detection of focal and non-focal EEG signals based on an integrated index, termed as focal and non-focal index (FNFI) is presented. This method involves application of discrete wavelet transform (DWT) on EEG signals. Then, di erent entropy measures, namely, average wavelet, permutation, fuzzy and phase entropies, are computed from approximate and detail coe cients of sub-band signals. We studied combinations of di erent feature ranking techniques and di erent classi ers for selection of most e ective features for classi cation of focal and non-focal EEG signals. Then, by employing top three most e ective features permutation, fuzzy and Shannon wavelet entropies FNFI is developed which discriminate focal and non-focal EEG signals using a single number. The proposed methodology can be helpful for identi cation of epileptogenic focus. Computer aided sleep monitoring device can e ectively reduce the burden of experts in analyzing the large volume of EEG recordings corresponding to sleep stages. The EEG signals are decomposed using iterative ltering method which result in modes. The discrete energy separation algorithm (DESA) is applied on the modes to separate amplitude envelope and instantaneous frequency functions. The Poincar e plot descriptors and statistical measures are computed from amplitude envelope and instantaneous frequency functions and used as input features for di erent classi ers in order to classify sleep stages. The two-class to six-class classi cation problems are formed by taking di erent combinations of EEG signals corresponding to various sleep stages. The methodology provided best results using random forest classi er. The EEG signals related to positive, neutral and negative emotions are studied and a methodology based on the tunable-Q wavelet transform (TQWT) for the recognition of human emotions is developed. The TQWT method is applied to emotion EEG signals and obtained sub-bands are used to extract the k-nearest neighbor (kNN) estimator based entropy. The kNN entropy features are used as input to the Na ve Bayes, J.48 decision tree, and random forest classi ers. The proposed emotion recognition methodology is found most e ective with random forest classi er. In this thesis, studying dierent advanced signal processing techniques, new methodologies are developed for the automated identication systems for identifying epileptic seizures, sleep stages and dierent emotions. The results of each methodologies are also compared with the other existing methodologies and the signicance of the results is discussed.
Appears in Collections:Discipline of Electrical Engineering

Files in This Item:
File Description SizeFormat 
TH_62_RajeevSharma.pdf7.22 MBAdobe PDFView/Open    Request a copy

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.