The Analysis of EEG Spectrogram Image for Brainwave Balancing Application Using ANN

Transcription

1 2011 UKSim 13th International Conference on Modelling and Simulation The Analysis of EEG Spectrogram Image for Brainwave Balancing Application Using ANN Mahfuzah Mustafa 1,2 1 Faculty of Electrical & Electronics Engineering University Malaysia Pahang Kuantan, Pahang, Malaysia mahfuzah@ump.edu.my Mohd Nasir Taib 2,3, Zunairah Hj Murat 2,3, Norizam Sulaiman 1,2, Siti Armiza Mohd Aris 2 2 Faculty of Electrical Engineering 3 Biomedical Research Laboratory for Human Potential University Teknologi MARA Malaysia Shah Alam, Selangor, Malaysia dr.nasir@ieee.org Abstract The purpose of this paper is to analysis EEG spectrogram image using Artificial Neural Network (ANN) for brainwave balancing application. Time-frequency approach or spectrogram image processing technique is used to analyze EEG signals. The Gray Level Co-occurrence Matrix (GLCM) texture feature was extracted from spectrogram image and passed through Principal components analysis (PCA) to reduce the feature dimension. The experimental result shows that ANN was able to analysis EEG spectrogram images with an optimized model in training by varying neurons in the hidden layer, learning rate and momentum. Keywords EEG, spectrogram image, GLCM, PCA, ANN I. INTRODUCTION The brain is the most important organ in the human body. The main function of the brain is to manage the entire process in the body for example to control heart rate, hearing and speech. It is estimated that brain consists of billion cells called neurons. Neurons produce electrical power in term of brainwaves to control the movement of the whole human body[1]. Brainwaves are normally categorized into four groups known as Delta, Theta, Alpha and Beta frequency bands [1]. Beta is the highest frequency band with the lowest amplitude while Delta is the lowest frequency band having the highest amplitude. Human brain is divided into two main regions which is right and left hemisphere. The left hemisphere is dominant in activities involving language, speech, arithmetic, and analysis whereas the right hemisphere is superior in perceiving, thinking, remembering, emoting and understanding [2, 3].The discovery of the left and right brain function was opened completely new fields of brain research. Using both the right and the left brain will yield optimum balanced lifestyles resulting in happiness and good health [4]. The aim of this paper is to analysis GLCM texture feature that was obtained from EEG spectrogram image using ANN for brainwave balancing application. The result was verified using established brain dominance questionaire. II. RELATED WORK Scientifically, the brainwave signal can be measured using Electroencephalography (EEG). The EEG signals are characterized by amplitude (voltage) and frequency. The frequency varies in each band, Delta range within 0.5 to 4 Hz, Theta range from 4 to 8 Hz, Alpha range from 8 to 13 Hz and Beta range from 13 to 32 Hz [5]. However, the raw EEG signal need to be analysed in order to extract useful information for specific research. Usually, EEG signal is extracted via three methods, namely time, frequency or time-frequency based. Generally, EEG raw signals are in time-based format. To analyse in frequency-based, typically the signals need to be transformed into Fourier Transform (FT). In this paper, EEG signals were analysed based on time-frequency image processing technique or called spectrogram. The most often technique used to analyse signal in time-frequency based is Short Time Fourier Transform (STFT). The STFT is to perform a FT on the signal, then mapping the signal into a two-dimensional function of frequency and time. Most of the EEG signal analysis have been carried out using time-frequency based, and mainly in signal processing area [6], thus, very few in image processing area. Nevertheless, a group of researchers used the spectrogram in time-frequency to analysis heart abnormalities from electrocardiogram (ECG) [7]. The spectrogram was produced using STFT technique from the ECG signal. They extracted the Euler number and height and width of pulses from the spectrogram image. Back-propagation ANN was used to analysis heart abnormalities. Once a spectrogram image is obtained, various image processing tools, such as texture analysis, could be used to further analysis. Most common technique for the textural classification is GLCM. This technique commonly used to /11 $ IEEE DOI /UKSIM

2 process texture of image from various applications such as satellite, wood and ultrasound. The GLCM is a tabulation of grey levels frequency occurring in an image. One example of study using GLCM to analyse ECG signal is explained in[8]. The study is to detect sleep disorder breathing in human heart. Fuzzy was selected as classifier and the result shows 79.29% accuracy in training and 75.88% accuracy in testing. After GLCM process is done, texture features need to be extracted. The most popular texture feature extraction is proposed by Haralick[9] with 14 texture features for the photomicrographs of sandstones, photographs of land-use and satellite photographs of land-use application. Soh[10] also proposed texture feature extraction with 10 texture features for satellite photographs of sea ice. It has been shown elsewhere; the PCA was used for data reduction, regression and classification purposes. There is an example that PCA was used for data reduction[11]. The first three principal components which are contrast, dissimilarity and homogeneity were chosen from GLCM texture feature out of eight features. As a result, the first three components from PCA give better accuracy in classification than all eight GLCM texture features. ANN history begins in the year 1940 s and was initiated by McCulloch and Pitts, however it was popular in the 1980s[12]. The idea of ANN research was insipred from human brain or specifically human biological nervous system to process information. ANN is actually a mathematical model to solve a variety of problems in control, prediction, pattern recognition, and optimization. There are several issues in ANN design, including network model, network size, activation function, learning parameters, and number of training samples. Nevertheless, there is no general guideline to choose the best ANN architecture for a particular size of the training. Training unconstrained networks using standard performance measures such as the mean squared error might even give very unsatisfying result [13]. ANN is highly suited to process feature rich data[12-16]. There are studies using extracted EEG signal features to be fed into ANN, in various application. For example, ANN was employed to analysis epileptic seizure[14], Parkinson disease[15] and braincomputer interface[16]. III. METHODS A. Subjects The data collections were perform at Biomedical Research and Development Laboratory for Human Potential, Faculty of Electrical Engineering, Universiti Teknologi MARA Malaysia. The samples were collected from 51 volunteers, 28 males and 23 females with an average age of All volunteers were in healthy condition and not consuming medication prior to the test. This study was approved by ethics committee from Universiti Teknologi MARA. B. EEG Measurement The EEG data were collected with bipolar electrodes by using standard gold disc electrodes, with 2-channel Fp1 and Fp2 and reference to earlobes A1, A2 and Fz. The electrode connections are in accordance to International system with 256Hz sampling rate. The EEG signal was recorded for five minutes using g.mobilab, with wireless EEG equipment. The impedance was maintained below 5kΩ using Z-checker equipment. Prior to EEG recording, volunteers have to answer the eleven item Brain Dominance Questionaire[17]. Once the questionaire is completed, the score is calculated to determine the index of each sample. This index is produced from the previous experiment[4]. Table 1 shows the data per index. Index 3 is for moderately balanced brain, Index 4 is for balanced brain and Index 5 is for highly balanced brain. Data for Index 1 and Index 2, corresponding to unbalanced brain and less balanced brain, respectively are not available.the data was collected and processed using intelligent signal processing technique developed in SIMULINK and MATLAB. C. EEG Signal Pre-processing Figure 1 shows the flow diagram for EEG signal analysis. EEG signal pre-processing includes artifact removal and band pass filter. Artifacts occur when the volunteers blink his or her eyes. This artifacts were removed by means of a program designed using MATLAB tools by setting a threshold value. The threshold was set to eliminate data when the values are more than 100µV and less than -100µV. The band pass filter is designed using Hamming window with 50% overlapping for the frequency 0.5Hz to 30Hz. D. Short Time Fourier Transform (STFT) The spectrogram image is produced using STFT with image size 436x342 pixels for both Fp1 and Fp2 channel. In spectrogram image, each frequency band is set. The Delta band is set from 0.5Hz to 4Hz, Theta band is set from 4Hz to 8Hz, Alpha band is set from 8Hz to 13Hz and Beta band is set from 13Hz to 30Hz. The STFT is done by multiplying Fourier Transform (FT) of the EEG signal by window function. TABLE 1 DATA SAMPLE PER INDEX Index Samples STFT Images Index Index Index TOTAL

3 E. Gray Level Co-occurrence Matrix (GLCM) The GLCM for the four different orientations 0 0, 45 0, 90 0 and were computed. Subsequently, texture feature were extracted for each GLCM. In this research, texture feature is the combination of Haralick and Soh technique. The 20 texture features are Autocorrelation, Contrast, Correlation, Cluster prominence, Cluster shade, Dissimilarity, Energy, Entropy, Homogeneity, Maximum probability, Variance, Sum average, Sum variance, Sum entropy, Different variance, Different entropy, Information of correlation 1, Information of correlation 2, Inverse difference normalized, and Inverse difference moment normalized. G. Artificial Neural Network (ANN) A feed-forward ANN was used to analysis EEG spectrogram image. The system has an 8 inputs and 1 output. To achieve the best ANN model, there are three parameters to be optimized, namely number of neurons in the hidden layer, learning rate and momentum. The optimum parameters can be achieved by finding the lowest mean square error (MSE). Many studies refer to MSE as the error goal [18-21]. The sigmoid was used for the ANN activation function. In each experiment, the parameter to be optimized is varied while the two parameters were fixed and MSE were observed. Finally, the best model for the experiment was selected for the final application. EEG signal Artifact removed Band pass filter Spectrogram image for (Δ-band, θ-band, α-band and β-band) IV. RESULT AND DISCUSSION Spectrogram images produced using STFT are as shown in Figure 2 (a)-(h). These figures illustrate Delta-band, Thetaband, Alpha-band and Beta-band for both Fp1 and Fp2 channels. Based on these figures, the spectrogram is texture shaped and each frequency band produce a different texture. Each EEG sample will produce eight images for both channels Fp1 and Fp2. The number of spectrogram generated as shown in Table 1. The GLCM generated for 0 0, 45 0, 90 0 and matrix orientation, with distance=1 for each spectrogram image and then texture feature from combination of Haralick and Soh were extracted. GLCM Texture feature PCA ANN (a) (b) Figure 1. Flow diagram for EEG signal analysis F. Principal Component Analysis (PCA) Output from GLCM texture feature generates big matrices. PCA was employed to reduce data dimension and to find optimum features for classification purposes. The optimum features will reduce execution time for the classification process. The first principal components contain most of the useful information, and the last principal components contain mostly noise. Therefore, these last principal components can be removed without significantly affecting the information content of the GLCM texture feature. (c) (e) (d) (f) 66

4 98.3%. Finally, the best network defined by 6 hidden neurons, learning rate of 0.8 and momentum of 0.2. (g) (h) Figure 2. Spectrogram images for (a) Delta band from Fp1 channel (b) Delta band from Fp2 channel (c) Theta band from Fp1 channel (d) Theta band from Fp2 channel (e) Alpha band from Fp1 channel (f) Alpha band from Fp2 channel (g) Beta band from Fp1 channel (h) Beta band from Fp2 channel 80 GLCM texture features were extracted and PCA is used to reduce this data dimension. Figure 3 shows percentage of eigenvalue produce by 80 principal components. The graph shows the percentage gradually decrease until last components. The first components show the highest percent, with 70% eigenvalue of covariance from original data. Some components have been chosen for the purpose of classification based on the results of PCA, and 8 principal components were selected because it produces high percentage of eigenvalue. Figure 4. Training performance and prediction accuracy with varying hidden layer size Figure 3. Graph of eigenvalue in percent Figures 4 to 6 illustrate the result for the optimization of ANN. Figure 4 shows the result for optimizing number of neurons in the hidden layer size. From this figure, it can be seen that the ANN with hidden layer 16, 12, 11, 10, 9, 7, and 6 may produced good prediction result. In this experiment, the network with hidden layer size 6 with MSE with accuracy rate 79.6% was selected. Figure 5 illustrates the result for finding the optimum learning rate and it is shown that learning rate of 0.2 and 0.8 may produced good prediction. The learning rate of 0.8 was found to be optimum with MSE with accuracy 97.9%. Figure 6 shows the result for the finding optimum momentum. From this figure, it can be seen that momentum value of 0.2 and 0.5. The learning rate of 0.2 was found to be optimum with MSE with accuracy Figure 5. Training performance and prediction accuracy with varying learning rate 67

5 Figure 6. Training performance and prediction accuracy with varying momentum V. CONCLUSION It has been shown that, ANN was able to analysis EEG spectrogram images with an optimized model in training by varying neurons in the hidden layer, learning rate and momentum. The prediction accuracy is 79.6% to 98.3%. The experimental result also shows that, PCA has able to reduce the original GLCM texture feature data. In the future, this optimized model will be used to classify EEG spectrogram image for balance brain application. ACKNOWLEDGMENT The author would like to thank the members of Biomedical Research Laboratory for Human Potential, FKE, UiTM for their cooperation and kindness and UMP for studentship of Mahfuzah Mustafa. REFERENCES [1] M.Teplan, "Fundamental of EEG measurement," Measurement Science Review, vol. 2, pp. pp.1-11, [2] R. Sperry, "Left-brain, right-brain," Saturday Review, vol. 2, pp. 30-3, [3] R. W. Sperry, "Some Effects of Disconnecting The Cerebral Hemispheres," in Division of Biology, California Institute of Technology, Pasadena California, 1981, pp [4] Z. H. Murat, M. N. Taib, S. Lias, R. S. S. A. Kadir, N. Sulaiman, and M. Mustafa, "The conformity between brainwave balancing index (BBI) using EEG and psychoanalysis test," International Journal of Simulation System, Science & Technology, vol. 11, pp , [5] D. Cvetkovic, "Electromagnetic and audio-visual stimulation of the human brain at extremely low frequencies," RMIT University, [6] Y. Goren, L. Davrath, I. Pinhas, E. Toledo, and S. Akselrod, "Individual time-dependent spectral boundaries for improved accuracy in timefrequency analysis of heart rate variability," IEEE Transactions on Biomedical Engineering, vol. 53, pp , [7] M. Saad, M. Nor, F. Bustami, and R. Ngadiran, "Classification of Heart Abnormalities Using Artificial Neural Network," Journal of Applied Sciences, vol. 7, pp , [8] A.-A. Mohammad, B. Khosrow, J. R. Burk, E. A. Lucas, and M. Manry, "A New Method to Detect Obstructive Sleep Apnea Using Fuzzy Classification of Time-Frequency Plots of the Heart Rate Variability," in Engineering in Medicine and Biology Society, EMBS '06. 28th Annual International Conference of the IEEE, 2006, pp [9] R. M. Haralick, K. Shanmugam, and I. H. Dinstein, "Textural Features for Image Classification," Systems, Man and Cybernetics, IEEE Transactions on, vol. 3, pp , [10] L. K. Soh and C. Tsatsoulis, "Texture analysis of SAR sea ice imagery using gray level co-occurrence matrices," Geoscience and Remote Sensing, IEEE Transactions on, vol. 37, pp , [11] H. Murray, A. Lucieer, and R. Williams, "Texture-based classification of sub-antarctic vegetation communities on Heard Island," International Journal of Applied Earth Observation and Geoinformation, vol. 12, pp [12] A. K. Jain, M. Jianchang, and K. M. Mohiuddin, "Artificial neural networks: a tutorial," Computer, vol. 29, pp , [13] M. Egmont-Petersen, D. de Ridder, and H. Handels, "Image processing with neural networks--a review," Pattern Recognition, vol. 35, pp , [14] K. P. Nayak, T. K. Padmashree, S. N. Rao, and N. U. Cholayya, "Artificial Neural Network for the Analysis of Electroencephalogram," in Intelligent Sensing and Information Processing, ICISIP Fourth International Conference on, 2006, pp [15] R. Rodrigues, P. Miguel, T. Teixeira, and J. Paulo, "Classification of Electroencephalogram signals using Artificial Neural Networks," in Biomedical Engineering and Informatics (BMEI), rd International Conference on, pp [16] H. Dongmei, Z. Hongwei, and Y. Naigong, "High Resolution Time- Frequency Analysis for Event-Related Electroencephalogram," in Intelligent Control and Automation, WCICA The Sixth World Congress on, 2006, pp [17] L. Mariani, "Brain-dominance questionaire," in Internet: [18] N. F. Güler, E. D. Übeyli, and I. Güler, "Recurrent neural networks employing Lyapunov exponents for EEG signals classification," Expert Systems with Applications, vol. 29, pp , [19] T. Ah Chung and A. D. Back, "Locally recurrent globally feedforward networks: a critical review of architectures," Neural Networks, IEEE Transactions on, vol. 5, pp , [20] R. J. Kuligowski and A. P. Barros, "Localized Precipitation Forecasts from a Numerical Weather Prediction Model Using Artificial Neural Networks," Weather and Forecasting, vol. 13, pp , [21] S. S. Panda, D. Chakraborty, and S. K. Pal, "Flank wear prediction in drilling using back propagation neural network and radial basis function network," Applied Soft Computing, vol. 8, pp ,