This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.
Epilepsy is one of the serious brain disorders, which affects many patients around the world. The implementation of Independent Component Analysis (ICA) based methods in obtaining the blind source signals are the accepted technique in the area of biosignal processing. The removal of artifacts and extraction of original brainwaves are considerable objectives for processing step of EEG signal analysis. The method of ICA based (JADE), SOBI, and Normalized Kurtosis based-blind signal separation (BSS) are applied here for original brainwaves extraction and seizure signal detection from the two epileptic EEG recordings. Of the three algorithms only SOBI was a little successful. It was found that the JADE and normalized Kurtosis based-BSS algorithms give better results.
Brain signals are originated from individual sources such as occipital area, auditory lobe, motor cortex and, etc so called Electroencephalogram (EEG). When epilepsy happens, the EEG is mixed with the epileptic seizures and contaminated by Electroocculogram (EOG) and Electromyogram (EMG) artifacts. Therefore, scalp epileptic EEG recordings contain normal EEG, epileptic seizure waves, and artifacts. Important processing steps for analysis of epileptic EEG recordings are to (1) remove artifacts and (2) detect the epileptic seizure signals. ICA is widely employed for these purposes [1-3]. An ICA based technique was proposed to remove ocular artifacts from high resolution EEG as simulation study . The non-Gaussianity and Gaussianity of contaminated EEG were measured by higher order cumulant and Gaussian distribution model in order to identify the independent signals [5, 6]. An ICA based system was proposed in  to remove a wide variety of artifacts from recorded EEG signals as off-line processing. A singular value decomposition based method was applied for removal of muscle and eye movement artifacts from EEG signals , and a total of eight (8) EEG fragments were estimated, identified, and filtered by the CCA and SVD algorithms . The validation of SOBI-recovered components was proposed in  throughout the two experiments, one took advantage of the fact that noise source associated with individual sensors could be validated independently by the SOBI process, and one utilized the fact that the time course and location of primary somatosensory cortex activation by median nerve stimulation characterized using converging imaging methods.
In the present study, the focus is on the two stages (1) decomposition of epileptic EEG recordings to achieve independent brainwaves and (2) epileptic seizure signal detection through visual investigation of the rhythmic features obtained by JADE, SOBI, and normalized Kurtosis based- BSS. It must be mentioned that the two epileptic EEG signals recorded from two individual patients are employed in this study.
Those algorithms that used in this study are derived from ICA concept, which assume that each observed signal of a multi-channel recording with channels can be expressed by a linear superposition:
of source signals , i.e. number of components equals number of sensors . In this study, the sources are assumed statistically independent. This means that the joint probability density function of the signals are factorized. Therefore, the sources can be separated theoretically by estimating a demixing matrix . Estimates of the original sources are identified by applying the demixing matrix to the measured variables as follows:. The observed signals are applied for whitening process before the three algorithms are employed .
JADE: Of ICA algorithms, Joint Approximate Diagonalization Eigenmatrices is applied for independent brainwaves extraction and epileptic seizure signal detection purposes. Fourth-order Cumulants is used as the cost function. This ICA algorithm reduces the mutual information involved the cumulant matrices by looking for a rotation matrix such that the cumulant matrices are as diagonal as possible .
SOBI: Second Order Blind Identification algorithm  takes advantage of the temporal structure in the observed data. The basis of the SOBI algorithm is a set of time-lagged covariance matrices.
Normalized Kurtosis based BSS: Normalized-Kurtosis is used as the cost function to estimate the coefficients of demixing matrix . The gradient descent optimization algorithm is applied for adjusting the tap-weights of demixing matrix as follows:
Where is the vector of the complex conjugate whitened scalp EEG signals, where the plus sign is for sub-Gaussian while the minus sign for super-Gaussian signals.
The epileptic EEG signals were recorded from two individual patients. Eighteen channels were used to record the EEG data from over the scalp. The frequency sampling rate (256Hz/per channel) was used to digitize the scalp EEG data. Software package ICALAB was used to provide us with the JADE and SOBI algorithms . Normalized Kurtosis based-BSS algorithm was made by writing the MATLAB code. Then, three algorithms were applied for (1) independent brain signal extraction and (2) epileptic seizure signal detection. From video observation of patient (1), figure 1 shows that the start time of epilepsy attack during the sleep started from 2sec. There is no abnormal activity in electrode points of FP2-F8, F8-T4,â€¦,Cz-Pz between 0 and 2sec. The rhythmic waves started from 2sec and remained continue until 12sec. The propagated version of positive narrow pulse shown between 12 and 12.5sec can be seen in the channels of F7-T3, T3-T5, F3-C3, and C3-P3. It is called spike signal. The negative peaked pulses represented between 16 and 17sec are blinking signal and video observation confirms this. Figure 2 shows that the start time of epilepsy attack for second patient started from 2sec and the high amplitude of rhythmic and fluctuating waves can be seen after 2sec. The right shoulder started to shake from 6sec and remained continue until 20sec. Therefore, the areas of interest are the signals located between 2 and 16sec for patient 1, and between 2 and 20sec for patient 2, in order to detect the seizure source signals. Red and blue signals illustrate activity of the right and left sides of the brain, respectively. Green signals show the activity in central part of the brain. It must be pointed out that the seizure attack happened at different time for each patient.
Figure 1) A segment of epileptic EEG signal recorded from patient 1
Figure 2) A segment of epileptic EEG signal recorded from patient 2
Results and Discussion
Figure 3 represents the independent signals obtained by JADE technique. Rhythmic waves with several fluctuating features related to the seizure wave are shown in the signals 8, 12, and 15. Propagation and influence of the rhythmic waves can be seen in the signals 2, 3, 4, 10, and 16, between 7000 and 10000msec. The largest amplitude of rhythmic wave has taken place in the signal 15. Therefore, the result of visual investigation showed that the signal 15 is the seizure wave. Figure 4 shows the independent EEG signals obtained by SOBI technique. Rhythmic waves are illustrated in the signals 7 and 12 between 3500 and 10000msec. Propagation of the rhythmic waves in the signals from 1 to 6 is less than the signals from 1 to 6 shown in figure 3. Signal 12 shown in figure 4 contains the largest amplitude in the rhythmic waves compared with the signals 1 to 18 and it is known as the seizure signal.
Figure 3) Independent signals extracted by JADE technique from scalp EEG; patient 1
Figure 4) Independent signals extracted by SOBI technique from scalp EEG; patient 1
Figure 5 represents the result of application of Normalized Kurtosis based-BSS method to patient 1 EEG. The main part of investigation to identify the seizure signal based on the rhythmic wave feature has been surrounded by a window. Independent signal (12) rose gradually to approximately and dropped suddenly to between 4000 and 4500msec and this confirms the expert opinion of the neurologist, and hence the signal (12) is "epileptic seizure signal". The propagation of rhythmic waves in other independent signals shown in figure 5 is less than the propagation of rhythmic waves in other independent signals represented in figures 3 and 4. This means that the normalized Kurtosis based BSS extracts independent signals better compared with the JADE and SOBI. Figure 6 represents the independent EEG signals extracted by JADE technique from patient 2 EEG. Signal 9 shows (i) the spike signal in between 2000 and 2500msec and (ii) the sequential rhythmic waves between 3500 and 6000msec, which is called seizure wave. Influence and propagation of the seizure signal 9 can be observed in the signals 11 between 3000 and 5600msec. Signal 3 shows several positive and negative fluctuating waves located between 3000 and 6500msec. They are fast blinking signals and video observation confirms this. Signals 8 and 13 are EMG waves and their propagations are obviously seen in the signals 14 and 15 between 3500 and 8500msec.
Figure 5) Independent signals extracted by Normalized Kurtosis based-BSS; patient 1
Figure 7 illustrates the independent EEG signals obtained by SOBI method from patient 2 EEG. Continuity of the rhythmic waves is observed in the signal 10 with the negative (impulse) spike signal located between 2500 and 5000msec. Signal 3 shows the fast blinking signal, which has been occurred during the epilepsy and video observation confirms this. As shown in figure 7, signal 13 represents the EMG signal and its propagation is represented in the signals from 7 to 16 except signal 10.
Figure 6) Independent signals extracted by JADE technique; patient 2
Figure 7) Independent signals extracted by SOBI technique; patient 2
The results of investigation of the second patient's EEG by normalized Kurtosis based-BSS are shown in figure 8. It indicates that the extracted independent signal (10) contains the spike, which takes place between 2 and 2.7sec. Signal (10) oscillates rhythmically from 3.2 to 6.5sec and its amplitude increased rapidly to and dipped quickly to around. This confirms the expert opinion of the neurologist and introduces signal (10) as the "epileptic seizure signal". In figure 8, the propagation of rhythmic waves in other independent signals is less than the influence of rhythmic features in other signals shown in figures 6 and 7. In the other hand, several congested and sharp pulses related to EMG signal (artifact) have affected other extracted independent signals illustrated in figures 6 and 7, and hence the rhythmic waves cannot be identified simply. But in figure 8, the rhythmic waves (seizure signal) can be seen obviously in the signal 10. Therefore, the independent signals extracted by normalized Kurtosis based BSS are more reliable compared with the obtained signals by JADE and SOBI.
Figure 8) Independent signals extracted by Normalized Kurtosis based-BSS; patient 2
Epileptic seizure signals
In order to find the source of the seizure signals, extracted independent signals: (15) represented in figure 3, (12) represented in figure 4, (12) shown in figure 5, (9) illustrated in figure 6, (10) depicted in figure 7, and (10) represented in figure 8 are accepted as the seizure signals, and hence the rest of the extracted independent signals are removed from scalp EEGs data.
Figure 9 represents the result of extraction of epileptic seizure source signal by taking into account the signal 15 represented in figure 3, which was obtained by JADE method. Propagation of the rhythmic waves illustrated by blue signals is observed between 3000 and 10000msec in the red signals. For example, signals T4-T6 and C4-P4 are the propagated version of signals T3-T5 and C3-P3 with the same rhythmic waves represented between 8000 and 10000msec. Therefore, it is a bit difficult to realize the seizure source signal in figure 9, because both right and left (red and blue signals) sides of the brain contain the same components of the seizure signals.
Figure 10 shows the result of epileptic seizure signal separation by considering on the signal (12) shown in figure 4, which was obtained by SOBI. Red signals located between 3000 and 10000msec are known as the seizure signals. Propagation of the red signals can be obviously seen in the signals T3-T5, T5-O2, F3-P3, and C3-P3 shown between 2000 and 6000msec, and between 8000 and 10000msec.
Figure 9) The result of seizure source signal extraction via JADE; patient 1
Figure 10) The result of seizure source signal extraction via SOBI; patient 1
Figure 11 illustrates the result of seizure source signal detection by considering on the signal 12, represented in figure 5, which was obtained by normalized Kurtosis based-BSS. It can be observed that the main location of the epileptic seizure source signal is around (electrode) points T3-T5 and C3-P3 because of the maximum amplitude of the rhythmic signals. The spread of seizure signal from T3-T5 and C3-P3 could be observed in other channels of FP2-F4, C4-P4, and Cz-Pz. The diagnoses of the hospital using their sophisticated equipment confirm that the main location of the seizure source signal is T3-T5.
Figure 12 represents the result of seizure source signal detection by taking into account the signal 9 shown in figure 6, which was obtained by JADE. Several fluctuating waves are observed in both positive and negative peaked points of the signal Cz-Pz located from 3500 to 5500msec. The seizure source signal is identified by the electrode point Cz-Pz. Propagation of the seizure signal is shown in the signals T3-T5 and F4-C3 between 3500 and 5500msec.
Figure 11) Extracted seizure source signals via Normalized Kurtosis based-BSS; patient 1
Figure 12) The result of seizure source signal extraction via JADE; patient 2
Figure 13 shows the result of seizure source signal separation by taking into account the signal 10 represented in figure 7, which was extracted by SOBI. A few negative and positive peaked pulses related to the EMG signals have contaminated the seizure signal located in electrode point Cz-Pz. Several propagated version of rhythmic waves are obviously observed in the signals T6-O2, FP1-F7, F7-T3, T5-O1, FP1-F3, C3-P3, P3-O1, and Fz-Cz, between 3000 and 5000msec. Therefore, it is a bit difficult to realize the start location of the seizure signal. Figure 14 illustrates the result of epileptic seizure source detection by considering on the signal 10 shown in figure 8, which was obtained by normalized Kurtosis based-BSS. The main location of the epileptic seizure source signal is electrode point Cz-Pz. The result shows that the seizure signal recorded from point of Cz-Pz spread out into the other scalp EEG channels in less than 1 to 2 sec. Therefore, the propagation of the seizure signal (10) represented in figure 8 can be seen in other channels of Fz-Cz, P3-O1, T5-O1, and C3-P3, respectively.
Figure 13) The result of seizure source signal extraction via SOBI; patient 2
Figure 14) Extracted seizure source signal via normalized Kurtosis based-BSS; patient 2
This paper describes the comparison of application of ICA based method JADE, SOBI, and normalized Kurtosis based BSS in (1) extracting the original independent brainwaves and (2) seizure source detection. Two epileptic EEGs data recorded from the two individual patients are analyzed by three methods. Extraction of independent signals are performed better by Normalized Kurtosis based BSS, because of its resulting independent signals show that there is less propagation of rhythmic waves into other extracted independent signals compared with the resulting signals obtained by JADE and SOBI. In addition, extraction of the epileptic seizure signal performed by normalized Kurtosis based BSS is reliable than the epileptic seizure signals extracted by JADE and SOBI. However, the results obtained by JADE is more reliable that the results extracted by SOBI. It was found that the source of epileptic seizure signal is in the left side of the brain for patient 1, and electrode point Cz-Pz for patient 2.
The authors would like to acknowledge Consultant Neurologists from Hospital University Kebangsaan Malaysia for providing the epileptic EEG data and valuable clinical advice.