The 2013 International Conference on Advanced Technologies for Communications (ATC'13) An Effective Procedure for Reducing EOG and EMG Artefacts from EEG Signals Nguyen Thi Anh-Dao‡ , Tran Duc-Nghia†† , Nguyen Thi-Hao† , Tran Duc-Tan† and Nguyen Linh-Trung† ‡ †† University of Technology and Logistics, Ho town, Thuan Thanh district, Bac Ninh, Vietnam Institute of Information Technology, Vietnamese Academy of Science and Technology, Ha noi, Vietnam † University of Engineering and Technology, Vietnam National University, Hanoi, Vietnam emails: linhtrung@vnu.edu.vn Abstract—Epilepsy is a neural disorder in which the electrical discharge in the brain is abnormal, synchronized and excessive Scalp Electroencephalogram (EEG) is often used in the diagnosis and treatment of epilepsy by examining the epileptic seizures and epileptic spikes By modeling the signal acquired at each electrode of the EEG measurement system as a linear combination of source signals generated in the brain, we can apply Blind Source Separation (BSS) techniques to separate the brain activity from other activities In this paper, we concentrate on applying SecondOrder Blind Identification (SOBI) algorithm to remove eye (EOG) and muscular (EMG) artifacts However, the disadvantage of SOBI is that it cannot provide the information about the order of sources, thus, an identification procedures of artifacts is further needed The effectiveness of this method has been examined and verified by simulated and experiment data Index Terms—epileptic seizures, EEG, EOG, EMG, timefrequency representations, under-determined blind separation Second-Order Blind Identification (SOBI) is one of the existing well-known and effective BSS algorithms and it has been applied widely in many applications [2] Some automatic detection methods based on SOBI and extended versions of SOBI have been proposed in previous studies [3]–[6] For examples, in [4], automatic removal of eye movement and blink artifacts from EEG data were considered In [5], spatially constrained ICA algorithms were proposed with applications in EEG processing These works are mostly complicated and time consuming In this paper, we proposed an effective scheme to remove electrooculographic (EOG) and electromyographic (EMG) artifacts from EEG This scheme works based on SOBI The disadvantage of SOBI is that it cannot provide the information about the order of sources Therefore, an identification procedures of artifacts is further needed I I NTRODUCTION II M ATERIALS AND M ETHODS Epilepsy is a neural disorder characterized by an enduring predisposition to generate epileptic seizures and its neurobiologic, cognitive, psychological, and social consequences An epileptic seizure is the abnormal, synchronized and excessive electrical discharge in the brain [1] Scalp electroencephalogram (EEG) is the recording of the temporal electrical brain activity through a set of scalp electrodes, thus it is useful to localize epileptogenic zones However, the EEG-based epileptic diagnosis faces difficulty because EEG signals are often disturbed by artifacts, such as: eye movements, muscle activity and heart activity Each recorded EEG signal can be modeled as a linear combination of source signals generated in the brain (seizures, background neural activity, artifacts, etc) Because it is not possible to turn off either the artifact sources or the cerebral sources it is not possible to record either or alone The artifact, thus, has to be removed from the combined recording by means of signal processing This has led to the development of several correction algorithms Hence, it is appropriate to we can apply blind source separation (BSS) techniques to separate the seizure from other signals in the EEG data BSS aims at recovering several source signals from several observed only linear mixtures of the source signals while the information about the mixing system is unavailable Assumptions for BSS to work depend on the mixing model and statistical properties of the source signals 978-1-4799-1089-2/13/$31.00 ©2013 IEEE A Noisy EEG data Within the scope of this research, we study two epileptic EEG dataset– simulated and real– disturbed by an eye-blink artifact, which is the most popular factor of distortion in the EEG recording process The first one contains four simulated EEG signals each of which is a mixture of three sources: EEG background, eye-blink artifact and seizure Ocular artifacts are the most relevant interference because they occur very frequently and their amplitude can be several times larger than brain scalp potentials As the eyeball moves, the electric field composed by cornea and retina changes and it produces the electrooculographic (EOG) signals Additionally, some neural activity is recorded by EOG electrodes because they are located near the head B Blind Source Separation and SOBI algorithm The assumption of independence among source signals can be relaxed to uncorrelation while using additional information about the source autocorrelation Thanks to using only second order statistic (SOS) information, the complexity of the SOBI algorithm and signal length can be reduced These algorithms have previously been applied to EEG seizure separation The classical linear mixing model can be written, at each instant k, as: 328 x = As, (1) The 2013 International Conference on Advanced Technologies for Communications (ATC'13) where x is a vector of M observed signals in EEG channels, A is the unknown full-column rank mixing matrix whose size is M ×N and s is the vector of N independent unknown sources In our study, we consider M > N because the number of EEG channels is larger than number of sources Thus, the separation problem here has a solution To estimate the original sources, it is need to calculate the following linear transformation: y = Wx = WAs, (2) where y is a vector of N estimated sources and W is a N ×M linear transformation matrix that allows the separation of the mixed signals in their independent components Thus, W should be the inverse of the matrix A; hence the sources can be prefectly recovered The most currently employed solution is to evaluate the number of linearly independent measures in the mixture by using some criterion based on the eigen-values of the covariance matrix of the measured signals However, it is difficult to obtain the exact inverse of the mixing matrix A Thus, source separation algorithms are focused on finding W such that G=WA be a permuted and scaled diagonal matrix It means that the sources can be recovered without information of their order and amplitudes In practical applications, we should not ignore the noise Therefore, Eq (1) should be rewritten as x = As + n, Expanding Eq (1) to x1 (k) = a11 s1 (k) + a12 s2 (k) + · · · + a1N sM (k) x2 (k) = a21 s1 (k) + a22 s2 (k) + · · · + a2N sM (k) xM (k) = aM s1 (k) + aM s2 (k) + · · · + aM N sM (k) Using SOBI we can estimate A Assume that we can identify two sources s1 and s2 that are EOG and EMG activities The EEG signals will be compensated as in the following: x1 (k) = x1 (k) − a11 s1 (k) − a12 s2 (k) x2 (k) = x2 (k) − a21 s1 (k) − a22 s2 (k) C Proposed scheme The disadvantage of SOBI is that we can not obtain the information of the order of sources and, thus, we can not compensate the noisy signal correctly In our study, we propose a method that integrates SOBI and source identification in order to remove EOG and EMG out of noisy EEG signals (5) xM (k) = xM (k) − aM s1 (k) − aM s2 (k) With a fixed-size window N , the energy of every output signals of the SOBI block is limited In the frequency domain, the energy of an EOG signal is maximum at low frequency while that of an EMG is maximum at high frequency Thus, we have exploited these properties to identify EOG and EMG Our scheme can be summarized in Algorithm Algorithm EOG and EMG identification using moving window Step 1: Initiate a moving window whose size is N =1000 ms Step 2: Calculate the weighting parameter for all channel w=E(f > 55Hz)/E(f < 20Hz) Step 3: EOG channel is determined by minimum of w and EMG channel is determined by maximum of w Step 4: Compensate the noisy EEG channel by using the Eq (5) Step 5: Continue with the next window (3) where n is a M ×1 vector of Gaussian noises Several BSS algorithms have been proposed and analyzed during the last decades Globally, source separation algorithms lay into two categories: those based on High Order Statistics (HOS) and those based on Second Order statistics (SOS) SOBI is one of the most representative algorithms of the SOS family The main advantages of these algorithms are their hypothesis are a priori verified for real EEG signals, which are band-limited and noisy These algorithms were already successfully applied for EEG separation, for example in [7], [8] Thus, we have included it into our analysis The first step of SOBI consists of whitening the signal part ˆ SOBI computes of the observation by a whitening matrix W ˆ of a set of covariance matrices Due to the joint diagonalizer U ˆ is unitary Then, the mixing matrix the whitening process, U can be calculated by the multiplication of pseudo-inverse of ˆ ˆ W ˆ # U the whitening matrix with the diagonalizing matrix A= # ˆ x ˆ (t) Finally, the source signals are estimated as ˆs(t)=A (4) The reason we use the ratio w=E(f > 55Hz)/E(f < 20Hz) instead of the energy is that the energy may be varied from channel to channel III S IMULATED AND E XPERIMENTED R ESULTS A Simulated Results Figure shows an EEG signal without any artifacts We can see that there are two spike existed in this segment of 5000 samples (at 510th and 3200th positions) Moreover, an amount of additive noise was added to this signal Figure shows a 5000-sample signal whose 2000 first samples are EOG After that, the EEG is mixed with the EOG artefact to obtain different mixed signals The first mixed signal is shown in Figure It is easy to see that the EEG signal is now dominated by EOG Consequently, we will analyze the performance of EOG removal using several methods: Least Mean Square (LMS), Zhou [9], Total variation (TV), and our method (e.g., combination of SOBI and identification of artefacts) Figures 4, 5, 6, and show the filtered signal applied to the first mixed one by LMS, Zhou, TV, and our proposed method, respectively Using LMS can amplify two spikes but it could not remove EOG artifacts Results obtained by Zhou or TV are even worse On the other hand, our method offers very good results wherein it can eliminate the EOG artifact totally 329 The 2013 International Conference on Advanced Technologies for Communications (ATC'13) Signal is filtered by LMS Expected signal 500 600 500 400 400 300 300 200 200 100 100 0 −100 −100 −200 −200 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 5000 5000 Fig The filtered signal using LMS algorithm Fig EEG signal without artifacts Signal is filtered by zhou 2500 Noise reference signal 2000 1500 1000 500 −500 −1 −1000 500 1000 1500 −2 −3 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 5000 2000 2500 3000 Samples 3500 4000 4500 5000 Fig The filtered signal using Zhou algorithm Fig EOG signal affected in 2000 first samples and ensure the existence of the two spikes at at 510th and 3200th positions Similar to EOG, the simulation scenario is changed to mixing of the EEG activity with an EMG artefact We can also obtain a good result using our method, in comparison with other methods Figure shows the signal at channel before and after filtering using our method Signal + Noise 2500 2000 1500 B Experiment Results 1000 500 −500 −1000 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 5000 Fig The mixed signal between EEG and EOG Firstly, data were acquired from 19 channels with the sampling frequency is 256 Hz Then, raw data were stored on the hard disk for further processing The initial processing process is to filter the raw data by feeding to a 50-Hz notch filter and a band-pass filter that altogether pass the signal between 0.5 Hz and 75 Hz We choose the first input channel is the channel that is most affected by ECG artefact (channel F p1), three other channels have less affect by EOG (channel C3, O2 and F z) Figure shows the signals in channels affected by EOG It is obviously that the EOG is dominated in all four channels from 330 The 2013 International Conference on Advanced Technologies for Communications (ATC'13) Signal is filtered by zhou Signal + Noise channel (Fp1) 0.2 2500 2000 −0.2 500 1500 1500 2000 2500 3000 2500 3000 2500 3000 2500 3000 Signal + Noise channel (C3) 0.05 1000 1000 −0.05 500 500 1500 2000 Signal + Noise channel 10 (O2) 0.05 1000 −500 −1000 −0.05 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 5000 500 1000 1500 2000 Signal + Noise channel 19 (Fz) 0.1 −0.1 Fig The filtered signal using TV algorithm 500 1000 1500 samples 2000 Fig The signal at channels affected by EOG Signal is filtered by SOBI 500 400 Signal is filtered by Zhou of channel 0.02 300 −0.02 200 500 1000 1500 2000 2500 3000 3500 3000 3500 3000 3500 3000 3500 Signal is filtered by Zhou of channel 0.05 100 −0.05 0 500 1000 1500 2000 2500 Signal is filtered by Zhou of channel 10 0.05 −100 500 1000 1500 2000 2500 Samples 3000 3500 4000 4500 5000 −0.05 Fig The filtered signal using our algorithm 500 1000 1500 2000 2500 Signal is filtered by Zhou of channel 19 0.05 Simulation signal with EMG is filtered by SOBI 250 −0.05 before filtering after filtering 200 150 500 1000 1500 2000 Samples 2500 Fig 10 The signal at channels filtered by Zhou 100 50 −50 −100 −150 −200 −250 500 1000 1500 2000 2500 3000 Samples 3500 4000 4500 5000 Fig The signal at channel before and after filtering using our method 1th to 450th samples Figures 10 and 11 show the signals in channels filtered by Zhou and our method, respectively We can see that the EOG artefacts are eliminated in all channels By using our method, amplitude of the spikes exists in channels are larger than ones using Zhou’s method For EMG artefact, we choose the first input channel is the channel in which a spike existed (channel F 4), the second input channel is most affected by EMG (channel P 4), and two other channels are less affected by EMG (channels C4 and F p1) Figure 12 shows signals in channels affected 331 The 2013 International Conference on Advanced Technologies for Communications (ATC'13) Signal is filtered by SOBI of channel −4 0.02 0 −0.02 500 1000 1500 2000 2500 3000 −5 3500 Signal is filtered by SOBI of channel 0 −0.05 500 1000 1500 2000 2500 3000 −5 −0.05 500 1000 1500 2000 2500 3000 3500 450 500 50 100 150 200 250 300 350 400 Signal is filtered by SOBI of channel 450 500 450 500 450 500 x 10 50 50 500 1000 1500 2000 Samples 2500 3000 3500 −4 200 250 300 350 400 100 150 200 250 300 Samples 350 400 Fig 13 The signal at channels filtered by our method Fig 11 The signal at channels filtered by our method ones In the future works, we will integrate the EMD method with this algorithm to enhance the performance Signal + Noise channel (F4) x 10 150 Signal is filtered by SOBI of channel −1 0.05 −0.05 100 −4 x 10 Signal is filtered by SOBI of channel 19 ACKNOWLEDGMENT 50 100 −4 150 200 250 300 350 400 450 This work was supported by Project QG-10.40 granted by Vietnam National University Hanoi 500 Signal + Noise channel (C4) x 10 R EFERENCES −2 50 100 −4 150 200 250 300 350 400 450 500 400 450 500 400 450 500 Signal + Noise channel (Fp1) x 10 −5 150 200 250 300 350 400 Signal is filtered by SOBI of channel 0 100 −4 0.05 50 x 10 −2 3500 Signal is filtered by SOBI of channel 10 −5 −4 0.05 Signal is filtered by SOBI of channel x 10 50 100 −4 150 200 250 300 350 Signal + Noise channel (P4) x 10 −1 50 100 150 200 250 300 Samples 350 Fig 12 Signals in channels affected by EMG by an EMG artefact These signals are treated by using our method The recovered signals are shown in Figure 13 It is very interesting to see that in Figure 12 we can not realize two spikes at 210th and 260th By using our method, these spikes are very clear, specially at three first channels IV C ONCLUSIONS This paper presents a new approach for minimization of EOG and EMG artefacts from EEG signals The simulation and experiment results demonstrated that our method shows the better performance comparison with some conventional [1] N V Thakor and D L Sherman, “Eeg signal processing: Theory and applications,” in Neural Engineering Springer, 2013, pp 259–303 [2] A Belouchrani, K Abed-Meraim, J.-F Cardoso, and E Moulines, “A blind source separation technique using second-order statistics,” Signal Processing, IEEE Transactions on, vol 45, no 2, pp 434–444, 1997 [3] R Ferdousy, A I Choudhory, M S Islam, M A Rab, and M E H Chowdhory, “Electrooculographic and electromyographic artifacts removal from eeg,” in Chemical, Biological and Environmental Engineering (ICBEE), 2010 2nd International Conference on IEEE, 2010, pp 163– 167 [4] C A Joyce, I F Gorodnitsky, and M Kutas, “Automatic removal of eye movement and blink artifacts from eeg data using blind component separation,” Psychophysiology, vol 41, no 2, pp 313–325, 2004 [5] M De Vos, L De Lathauwer, and S Van Huffel, “Spatially constrained ica algorithm with an application in eeg processing,” Signal Processing, vol 91, no 8, pp 1963–1972, 2011 [6] N T Thuy-Duong, N Linh-Trung, T Tran-Duc, and B Boashash, “Separation of nonstationary eeg epileptic seizures using time-frequencybased blind signal processing techniques,” in 4th International Conference on Biomedical Engineering in Vietnam Springer, 2013, pp 317–323 [7] J J Kierkels, G J van Boxtel, and L L Vogten, “A model-based objective evaluation of eye movement correction in eeg recordings,” Biomedical Engineering, IEEE Transactions on, vol 53, no 2, pp 246– 253, 2006 [8] S Romero, M Ma˜nanas, and M J Barbanoj, “Ocular reduction in eeg signals based on adaptive filtering, regression and blind source separation,” Annals of biomedical engineering, vol 37, no 1, pp 176– 191, 2009 [9] W Zhou and J Gotman, “Removal of emg and ecg artifacts from eeg based on wavelet transform and ica,” in Engineering in Medicine and Biology Society, 2004 IEMBS’04 26th Annual International Conference of the IEEE, vol IEEE, 2004, pp 392–395 332 ... a new approach for minimization of EOG and EMG artefacts from EEG signals The simulation and experiment results demonstrated that our method shows the better performance comparison with some... [9] W Zhou and J Gotman, “Removal of emg and ecg artifacts from eeg based on wavelet transform and ica,” in Engineering in Medicine and Biology Society, 2004 IEMBS’04 26th Annual International Conference... · · + aM N sM (k) Using SOBI we can estimate A Assume that we can identify two sources s1 and s2 that are EOG and EMG activities The EEG signals will be compensated as in the following: x1 (k)