VIETNAM NATIONAL UNIVERSITY HANOI UNIVERSITY OF ENGINEERING AND TECHNOLOGY EEG DATA ANALYSIS USING EMPIRICAL MODAL DECOMPOSITION Phạm Văn Thành BACHELOR THESIS May, 2012 EEG DATA ANALYSIS USING EMPIRICAL MODAL DECOMPOSITION Phạm văn Thành Faculty of Electronics and Telecommunications University of Engineering and Technology Vietnam National University, Hanoi Supervised: Dr Nguyễn Linh Trung Document submitted for committee examination as a requirement for the Bachelor of Electronics and Communications Engineering degree at the University of Engineering and Technology May, 2012 AUTHORSHIP “I hereby declare that the work contained in this thesis is of my own and has not been previously submitted for a degree or diploma at this or any other higher education institution To the best of my knowledge and belief, the thesis contains no materials previously published or written by another person except where due reference or acknowledgement is made.” Signature:………………………………………… i SUPERVISOR’S APPROVAL “I hereby approve that the thesis in its current form is ready for committee examination as a requirement for the Bachelor of Electronics and Communications Engineering degree at the University of Engineering and Technology.” Signature:……………………………………………… ii TABLE OF CONTENTS INTRODUCTION BACKGROUND 2.1 Time-Frequency Presentation 2.1.1 Signals in classical representation 2.1.2 The need for Time-frequency representation 2.1.3 The techniques for Time-Frequency Presentation 2.1.3.1 The Wigner-Ville Distribution 2.1.3.2 The Spectrogram 2.1.3.3 The Reassigned spectrogram 2.1.3.4 Drawbacks of Wigner-Ville and Spectrogram methods .8 2.2 EEG Signal Characteristics 2.2.1 The Nervous System 2.2.1.1 Neurons 10 2.2.1.2 The Cerebral Cortex .11 2.2.2 Electrical Activity Measured on the Scalp .12 2.2.3 Normal EEG 13 2.2.4 Spikes in the EEG Signals .14 2.2.5 EOG in the EEG Signals .15 2.2.6 EEG Artifacts 15 2.3 Empirical Mode Decomposition 15 2.3.1 The sifting process 19 EEG DATA ANALYSIS USING EMPIRICAL MODE DECOMPOSITION 21 3.1 The method for measure real data 22 3.1.1 Devices connection .23 3.1.1.1 Connect the electrodes to an amplifier 23 3.1.2 Attach the electrodes to the patient 23 3.1.2.1 The electrodes on the cap .23 iii 3.1.2.2 The outside electrodes: 24 3.1.3 Choose the parameters on the machine 25 3.1.4 Surrounding environment 25 3.1.5 Measurement 26 3.1.5.1 Check the impedance of the electrodes 26 3.1.5.2 Standard (calibration) 26 3.1.5.3 Record 26 3.1.5.4 Observe the patient 27 3.1.5.5 Stimulation .27 3.1.6 Data assessment 28 3.2 Applying EMD into EEG signals processing .31 RESULTS AND DISCUSSIONS 35 4.1 Results 36 4.1.1 Model for Simulation 36 4.2 Discussion 46 CONCLUSIONS 50 References 53 Appendix A 55 iv Abstract Epilepsy is one of the most popular neurological disorders characterized and unexpected electrical disturbances of the brain The electroencephalogram (EEG) is an invaluable measurement for the purpose of assessing brain activities, containing information relating to the different physiological states of the brain It is a very effective tool for understanding the complex dynamical behavior of the brain This thesis presents the application of Empirical Mode Decomposition (EMD) for analysis EEG signals EMD is a technique, it can be compared with other analysis methods like Fourier Transforms and wavelet decomposition It is a method of breaking down a signal into IMFs without leaving the time domain It is useful for analyzing natural signals using the way to filters out functions which form a complete and nearly orthogonal basis for the original signal As we know, the EEG is a very complex signals, it contains many elements such as: background, EOG, EMG, spikes… So, the EMD decomposes the EEG signal into a finite set of bandlimited signals named Intrinsic Mode Functions After that, using Reassigned-Spectrogram to present the EEG signal together with IMFs on Time-Frequency domain and using Hillbert transformation for presents IMFs on the complex plane The area measured from the trace of the analytic IMFs, which have circular form in the complex plane, has been used as a feature in order to distinguish the background, EOG and spikes in epileptic EEG signals It has been shown that the area measure of the IMFs on the complex plane has given good discrimination performance between background, EOG and spikes v Acknowledgement I would like to express my sincere thanks to my advisor Dr Nguyen Linh Trung, the professional of Faculty of Electronic and Communication, University of Engineering and Technology – Vietnam National University, Hanoi for the guidance and support given to me throughout the thesis Special thanks to Dr Tran Duc Tan and Ms Nguyen Thi Thuy Duong the lecturers of Faculty of Electronic and Communication for their help and guidance me in all thesis process I would like to thank doctor Hoang Cam Tu who help me to use, measure the EEG machine system and the method to detect spikes and read noise Thanks for all of people in the Signal Processing Lab for their help and discussed conversations At the end, I would like to thank my parents and my brother because their comfort and supporting are the power for me going to success vi List of Figures Figure 2-1: Example with a 128 point 2-component signals different TFDs are display in the case of 128 point Figure 2-2: A signal propagating down an axon to the cell body and dendrites of the next cell [11] 10 Figure 2-3: Three types of neurons Motor, sensory, and interneurons [12] 11 Figure 2-4: Corebral cortex and four lobes [13] 12 Figure 2-5: EEG signal in the different status was excited, relaxed, drowsy, asleep, and deep sleep 14 Figure 2-7: Empirical Mode Decomposition (EMD) of a three component signal 17 Figure 2-8: Illustration of the process of EMD technique breakdown signal into IMFs components [15] 19 Figure 3-1: International standard 10/20 22 Figure 3-2: The connection between electrodes with an amplifier 23 Figure 3-3: EOG attached 24 Figure 3-4: EMG attached Figure 25 Figure 3-5: Unipolar 27 Figure 3-6: Bipolar 27 Figure 3-7: The raw EEG Signal 29 Figure 3-8: EEG Signals before use filters 30 Figure 3-9: EEG Signals after use filters 31 Figure 3-10: analytic signal representations in the complex plane (window size = 512 samples 33 Figure 4-1: The Empirical Mode Decomposition of (a) background, (b) spikes, (c) EOG 39 Figure 4-2: EEG Signals and IMFs ((a) background, (b) spikes, (c) EOG)on TimeFrequency domain of Reassigned-Spectrogram method 42 vii Figure 4-3: analytic signal representations in the complex plane (window size = 512 samples) 44 Figure 4-4: The difference between background, spikes and EOG base on energy level and frequency level 46 Figure 4-5: background signal and IMFs 48 viii (a) (b) (c) Figure 4-4: The difference between background, spikes and EOG base on energy level and frequency level 4.2 Discussion From the result above, I have some comments: The EMD technique beak down signal into IMFs, the first IMF contains higher frequency elements and the frequency level of IMFs decreased from IMF1 to the end and the last IMF is a residue 46 Signal IMF1 IMF2 IMF3 IMF4 IMF5 IMF6 (Residue) 47 Figure 4-5: background signal and IMFs Besides, we also observed that, the signal contains spikes having high frequencies Meanwhile, the signal is background or contains EOG has lowerfrequency The use of EMD before computing Hilbert transform has been used in this thesis more useful than applying the Hilbert transform to the whole signal because it provides greater insight, it also detected significant differences in surface area between background, EOG and spikes of epileptic EEG signals In addition, the analytic signal representation of epileptic EEG signals in the complex plane not have a specific geometry, makes it difficult to define the circle that encloses 95% of the total data points The Central Tendency Measure (CTM) was used to determine the number of data points lies inside the circle of specified radius The EMD based decomposition makes possible to cover more than 95% of the data points that lie within the circle due to having circular form of analytic IMFs in the complex plane From the figure 4-3, we can see that, the trace of the analytic signals of each IMF always have a circular form So, the surface area of them can be estimated in the complex plane The table 4.2 shows area parameters minimum (min), median (med), and maximum (max) for each of the classes for different Intrinsic Mode Functions The value of estimated area is small in background EEG signals, median with spikes and bigger with EOG when we compare together From figure 4.4, we can see that, the values of estimated area of all IMFs of background are smaller than the area of all IMFs of spikes when compared them together The cause for the difference is because spikes have high frequency while background has low frequency So, the energy and area of spikes are bigger than background possibly due to the greater amplitude of EEG signals for spikes for all IMFs Next to, we can be easily to distinguish the difference between background and EOG because EOG and background have low frequency level, but at the IMF4 of EOG, it has big one dimension elements So, the energy of EOG is bigger than IMF4 of background In order 48 to compare the difference between EOG and spikes, we had known that spikes have high frequency, but EOG has low frequency So, the first IMF of spikes has higher energy than EOG, but the energy levels of IMFs of spikes are unpredictable fluctuation while the energy levels of IMFs of EOG increase very quickly 49 Chapter CONCLUSIONS 50 In this thesis, I explored the ability of the analytic signal representation of Intrinsic Mode Functions (IFMs) to discriminate EEG signals which contains epileptic spikes, EOG and normal EEG signals (background), the applying of EMD into analysis EEG signal is a promising method because EMD decompose EEG signals into IMFs, it will provide a set of proper rotations which makes possible to accurately identify the center and estimating surface areas in the complex plane The area parameter of the IMFs has provided the difference between spikes, EOG and background Based on it, we can see that the EOG has greater surface area than spikes and background because it has big one dimension elements Next to, spikes have surface area at average level because it has many high frequency elements and the last is normal EEG (background) because it contains low frequency and a few of one dimension elements The increased surface area in the complex plane for IMFs of EOG might due to bigger amplitude of EEG signal for EOG, spikes and background subjects respectively Besides, the use of Empirical Mode Decomposition (EMD) enabled the extraction of individual centers of rotation for each IMF and establishes the relation between the rotations corresponding to IMFs and rhythms of EEG signal From these characteristic, we can see that, the combined of EMD, Reassigned-Spectrogram and Central Tendency Measurement techniques It is useful tool in medicine because it will help doctors and everybody can know more about epilepsy disease and treatment when it begins appearing Nevertheless, this method also exits the problem: Spikes and EOG exit on difference channels but EMD analysis on only one channel So, we need to repeat many times, it will make difficult to observe Reassigned-Spectrogram standardizes the highest energy level is and the other elements was sorted based on this standard Therefore, it is easy to misunderstand about the level of energy of signal and IMFs if we not understand about the mechanism of the Reassigned-Spectrogram code and frequency distribution For the future of research, I will go to measure and distinguish others artifacts in epileptic EEG signals such as: distinguish EOG and EMG, spikes and seizures…and from those difference, the doctor can diagnose exactly for epileptic patients Besides, I also proposed to present EOG, spikes and 51 background on the same index for people can observe and detect the difference between EOG, spikes and background when we present it on the TimeFrequency domain 52 References [1] Ram Bilas Pachori*, Varun Bajaj, “Analysis of normal and epileptic seizure EEG signals using empirical mode decomposition”, School of Engineering, Indian Institute of Technology Indore, Indore 45017, Indian, 2011 [2] A.Delorme, S.Makeig, T.Sejnowski, “Automatic artifact rejection for EEG data using high-order statistics and independent component analysis”, CNL, Salk Institute, 2001 [3] Tomasz M Rutkowski, Rafal Zdunek, Andrzej Cichocki, “Multichannel EEG Brain Activity Pattern Analysis in Time-Frequency domain with nonegative matrix factorization support”, International Congress Series 8611, 266-269 (2007) [4] Boualem Boashash “Time frequency signal analysis and processing a comprehensive reference” Elseviev, Oxford, 2003 [5] Leon Cohen, “Time-Frequency analysis” Prentice-Hall, Inc Upper Saddle River, Nj, USA, 1995 [6] Leif Sornmo, Pablo Laguna, “Bioelectrical Signal Processing in Cardac and neurological Applications”, 2005 [7] ITU, “Internet protocol data communication service – IP packet transfer and availability performance parameters,” ITU-T Recommendation Y.1540, Feb 1999 [8] [online] http://www.clear.rice.edu/elec301/Projects02/empiricalMode/ [9] [online] http://pami.uwaterloo.ca/~gsdharwa/b_c_i/eeg_signal.htm [10] [online] http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3116477/ [11] [online] http://en.wikipedia.org/wiki/Neuron [12] [online]http://willistonberg.pbworks.com/w/page/21091185/Shaffer's%20Neur ology%20Page [13] [online]http://www.innovateus.net/innopedia/what-are-functions-brain-cortex 53 [14] [online]http://people.ece.cornell.edu/land/courses/ece4760/FinalProjects/s2012 /cwm55/cwm55_mj294/index.html [15] [online] http://perso.ens-lyon.fr/patrick.flandrin/emd.html [16] Nguyen Van Du, “Estimation of the number of sources in Blind Source Separation and Application in EEG data”, Bachelor thesis, University of Engineering and Technology, 2007 54 Appendix A Reading EEG data information from file.xlsx Provide some important MATLAB codes filename = 'D:\Pham Van Thanh\package_emd\examples\NSIP2003\Phamviethoang.xls'; start_record = xlsread(filename,1,'B2'); duration time: D3 cell, sheet sig_time = xlsread(filename,1,'C2:D9'); time of a signal channel_number = xlsread(filename,1,'E2:E9'); channel in file.mat %% Loading prefiltered EEG data in file.mat load 63s_phamviethoang_fil K = 339 Fs = 256; 10 sig_time(:,2) = sig_time(:,2) + 1; 11 index = (sig_time - start_record)*Fs + K; 12 index = round(index); 13 spike1 = test(channel_number(1),index(1,1):index(1,2)); 14 spike2 = test(channel_number(2),index(2,1):index(2,2)); 15 eye1 = test(channel_number(4),index(4,1):index(4,2)); 16 eye2 = test(channel_number(5),index(5,1):index(5,2)); 17 bgrd1 = test(channel_number(7),index(7,1):index(7,2)); 18 bgrd2 = test(channel_number(8),index(8,1):index(8,2)); 19 spike = [spike1(1:512); spike2(1:512)]; 20 eye = [eye1(1:512); eye2(1:512)]; 21 bgrd = [bgrd1(1:512); bgrd2(1:512)]; % reading the start % reading the start/stop % reading the index of % The sampling frequency 55 Run EMD for a background EEG signal [N_spike, M_spike] = size(spike); Nf = 512;% # of frequency bins Nh = 1023;% short-time window length w = tftb_window(Nh,'Kaiser'); T = 1:N_spike; imf_spike = zeros(4,512,M_spike); imf_eye = zeros(4,512,M_spike); imf_bgrd = zeros(4,512,M_spike); % imf_bgrd_2 = zeros(5,512,M_spike); 10 for i=1:M_spike 11 [imf,ort,nbits] = emd(spike(:,i)); 12 i; emd_visu(spike(:,i),T,imf,i);%show the IMFs in the figure number 13 imf_spike(:,:,i) = imf(1:4,:); 14 [s,rs] = tfrrsp(spike(:,i),T,Nf,w,1); 15 [s,rs1] = tfrrsp(imf(1,:)',T,Nf,w,1); 16 [s,rs2] = tfrrsp(imf(2,:)',T,Nf,w,1); 17 [s,rs3] = tfrrsp(imf(3,:)',T,Nf,w,1); 18 [s,rs4] = tfrrsp(imf(4,:)',T,Nf,w,1); Compute the Central Tendency Measurement spike = hilbert(spike); imf_spike_1 = imf_spike(:,:,1)'; imf_spike_2 = imf_spike(:,:,2)'; % convert to analytic signal imf_spike_1 = hilbert(imf_spike_1); % separate the 3-dim matrices into 2-dim matrices (need to be done manually) 56 imf_spike_2 = hilbert(imf_spike_2); %% Compute the radius of spike signals and eye signals [N,M] = size(spike); r_spike = zeros(N,M); r_eye = zeros(N,M); 10 r_bgrd = zeros(N,M); 11 for j=1:M 12 for i=1:N 13 r_spike(i,j)=sqrt(real(spike(i,j)^2+imag(spike(i,j))^2)); 14 r_eye(i,j)=sqrt(real(eye(i,j)^2+imag(eye(i,j))^2)); 15 r_bgrd(i,j)=sqrt(real(bgrd(i,j)^2+imag(bgrd(i,j))^2)) 16 end 17 end 18 %% Compute the radius of spike IMF signal and eye IMF signal 19 [N_imf, M_imf] = size(imf_spike_1); 20 for jj1=1:M_imf 21 for kk1=1:N_imf 22 r_imf_spike_1(kk1,jj1)=sqrt(real(imf_spike_1(kk1,jj1))^2+imag(imf_spike _1(kk1,jj1))^2); 23 r_imf_spike_2(kk1,jj1)=sqrt(real(imf_spike_2(kk1,jj1))^2+imag(imf_spike _2(kk1,jj1))^2); 24 r_imf_eye_1(kk1,jj1)=sqrt(real(imf_eye_1(kk1,jj1))^2+imag(imf_eye_1(kk1 ,jj1))^2); 25 r_imf_eye_2(kk1,jj1)=sqrt(real(imf_eye_2(kk1,jj1))^2+imag(imf_eye_2(kk1 ,jj1))^2); 26 r_imf_bgrd_1(kk1,jj1)=sqrt(real(imf_bgrd_1(kk1,jj1))^2+imag(imf_bgrd_1( kk1,jj1))^2); 57 27 r_imf_bgrd_2(kk1,jj1)=sqrt(real(imf_bgrd_2(kk1,jj1))^2+imag(imf_bgrd_2( kk1,jj1))^2); 28 end 29 end 30 %% sort the radius matrices in ascending 31 r_spike = sort(r_spike,1); 32 r_imf_spike_1 = sort(r_imf_spike_1,1); 33 r_imf_spike_2 = sort(r_imf_spike_2,1); 34 r_eye = sort(r_eye,1); 35 r_imf_eye_1 = sort(r_imf_eye_1,1); 36 r_imf_eye_2 = sort(r_imf_eye_2,1); 37 r_bgrd = sort(r_bgrd,1); 38 r_imf_bgrd_1 = sort(r_imf_bgrd_1,1); 39 r_imf_bgrd_2 = sort(r_imf_bgrd_2,1); 40 %% Compute CMT of spike signals 41 L = 100; % CMT is a continuous function in [0,1], then L is the number of CMT sample 42 k = zeros(L,M); 43 k(:,1) = linspace(r_spike(1,1),r_spike(N,1),L); 44 k(:,2) = linspace(r_spike(1,2),r_spike(N,2),L); 45 CMT_spike = zeros(L,M); 46 for i = 1:M 47 48 for n = 1:L for m = 1:N 49 if (r_spike(m,i) [...]... 2.3 Empirical Mode Decomposition Empirical Mode Decomposition (EMD) is a method of breaking down signal into Intrinsic Mode Functions (IMFs) without leaving the time domain It can be compared with other analysis methods like Fourier Transforms; wavelet decomposition and it is a useful for analyzing natural signals, which are most often non-linear and non-stationary 15 16 Figure 2-6: Empirical Mode Decomposition. .. of the decomposition the signal ( ) will be represented as follow: ( )= ( )+ Where M is the number of IMFs and ( ) ( ) is the final residue Each IMF is assumed to yield a meaningful local frequency and the difference IMFs do not exhibit the same frequency at the time 20 Chapter 3 EEG DATA ANALYSIS USING EMPIRICAL MODE DECOMPOSITION 21 3.1 The method for measure real data In this thesis, I use EEG machine... signals As a result, they can make to difficult to interpret the EEG signals or lead to wrong diagnosis In order to reduce the wrong diagnosis, we need methods which can detect the artifact Based on that, doctors can find out epilepsy pulses easier So, this thesis, focuses on the Empirical Mode Decomposition (EMD) technique to decompose EEG signals into Intrinsic Mode Functions (IMFs) From these IMFs,... Intrinsic Mode Functions (IMFs) From these IMFs, Central Tendency Measure (CTM) to detect which artifacts are exits in the EEG signals The purpose of this thesis is to measure and investigate the characteristic of EEG signal and distinguishes EOG, spikes and background of epileptic EEG signals using Time-Frequency toolbox, EMD technique and CMT method The thesis is organized as follow Chapter 2 introduces... Distribution 2.2 EEG Signal Characteristics The human brain has complex structure, it includes millions of neurons and each neuron generates a small of electric voltage fields The aggregate of these electric voltage fields will make enough amount of electrical for electrodes on the scalp can access and record Therefore, the EEG signal is a complex of many simple signals The amplitude of EEG signal fluctuated... alertness or dream sleep, while low frequency/large amplitude rhythms are associated with drowsiness and non-dreaming sleep states 2.2.3 Normal EEG EEG is generally described in terms of its frequency band Delta, theta, alpha, beta and gamma are the names of the different EEG frequency bands which relate to various brain states Delta waves, 0.1 to 3 Hz The delta waves are typically encountered during the deep,... identification of signs or symptoms of clinical seizures An Electroencephalograph (EEG) records the brain's activity; EEG is the most complementary test for clinical diagnosis Especially, it is able to identify types of epilepsy and brain lesion areas which caused epilepsy impulses At the moment, in Vietnam as many other countries, the EEG recording system are in very common, use the signals of these machines... 37 Table 4-2: Area of analytic signal representation of Intrinsic mode functions 45 ix List of Abbreviations CNS Central Nervous System CTM Central Tendency Measurement EEG Electroencephalograph EMD Empirical Mode Decomposition EOG Electrooculogram FFT Fast Fourier Transform HHT Hilbert-Huang Transform IMFs Intrinsic Mode Functions PNS peripheral Nervous System RSP Reassigned-Spectrogram SP... frontal and central regions of scalp Gamma waves, 30 Hz to 100 Hz The gamma waves are related to state of Motor Functions, higher mental activity 2.2.4 Spikes in the EEG Signals Spikes are transient waveforms that stand out from the background EEG with an irregular It is unpredictable when time spikes will appear in epileptic patients, the appearing of spikes in a people is a result for improving the people... biphasic and triphasic waveforms and the 14 waveform morphology is dependent on where the electrode is located on the scalp 2.2.5 EOG in the EEG Signals Eye movement produces electrical activity-the electrooculogram (EOG) which is strong enough to be clearly visible in the EEG The EOG reflects the potential difference between the cornea and the retina which changes during eye movement The measured voltage