FUZZY ARTMAP CLASSIFICATION FOR MOTOR IMAGINARY BASED BRAIN COMPUTER INTERFACE

Vickneswaran Jeyabalan*, Andrews Samraj**, Loo Chu Kiong**,

*

Faculty of Engineering and Technology, Multimedia University, Jalan Ayer Keroh Lama, Melaka, Malaysia vickneswaran.jeyab07@mmu.edu.my

**

Faculty of Information Science and Technology, Multimedia University, Jalan Ayer Keroh Lama, Melaka, Malaysia andrews.samraj@mmu.edu.my, ckloo@mmu.edu.my imaginary task produces extensive changes in the EEG and is easily to be detected [4]. These tasks are usually correlated to the mu-rhythm. The mu-rhythm can be observed in the Alpha frequency band of an EEG signal. The mu-rhythm is in the frequency range of 8Hz to 12Hz. Mu rhythm can be best measured from the EEGs channel C3, Cz and C4. Motor imagination signals are conventionally detected by observing the power attenuation of the EEG signal. The attenuation of the power is known as event related de-synchronisation (ERD) whereas the rebounding of the power is known as event related synchronisation (ERS). The ERD or ERS occurs contra laterally to the intended movements. In a normal ERD detection procedure, EEG signals are usually filtered in a narrow band, squared, low pass filtered and averaged over trial [8]. However, utilizing this method could cause useful information to be lost from the averaging over multitrials. In this paper we have implemented the adaptive recursive bandpass filter [13,14] to extract the motor imaginary features and to classify them according to left or right motor imagination movements using fuzzy ARTMAP (FA). Implementation of this method reduces the chances of information lost and increases the classification rate. The target signal, which contains mu rhythm, is first pre-processed by implementing a band pass filter. Then, the dominant signal which represents the motor movements is estimated by using the proposed adaptive recursive band pass filter. The filter trails the centre frequency of the dominant EEG signal and requires only one coefficient to be updated [13,14]. This is done in order to adjust the centre frequency of the filter band pass and to be approximated with the input signal [13,14]. The time function of the coefficient represents the feature for each signal. The feature vectors are then given as an input to the FA artificial neural network (ANN) for classification. FA is used because its a neural network (NN) architecture that consumes less training time and also creates a stable memory structure with fast online learning [7]. FA has also the ability for incremental learning, where data from new subjects are trained without needing to re-train data from previously trained subjects [12]. Whereas, other NN architectures have good generalization ability, but are difficult to train and consumes time [12].

ABSTRACT The rapid advancement of Brain Machine Interface (BMI) or Brain Computer Interface (BCI) research over the recent years is concentrated to the development of new technologies which adopts the easiest procedures since the expected beneficiaries are of disabled in nature. Most of the locked-in patients possess strong mental ability in terms of imagining and thinking but they are extremely unable to express their views physically. Hence a BMI which does not rely on physical movements or other muscle activity is definitely an added advantage in this arena. The objective of this paper is to identify and classify motor imaginary signals extracted from the left and right cortex of the human brain. The signals are captured using Electro-Encephalogram (EEG) from the C3, C4, and Cz channels of the scalp electrodes and is preprocessed to expose the motor imaginary signals. The features are extracted by implementing an adaptive band pass filter with the combination of frequency shifting and segmentation technique. The result of classification using FuzzyARTMAP articulates the effectiveness of our proposed technique. Positive results are obtained by using only features from one channel. The best results were found from the features of channel C4 and this proves the existence of neuro-science knowledge. Keywords: Brain Machine Interface,motor imaginary, adaptive filtering,Fuzzy ARTMAP

In recent years, the field of electroencephalogram (EEG) based Brain Machine Interface (BMI) or also know as Brain Computer Interface (BCI) has been evolving rapidly. The main objective of a BMI or BCI is to provide a new method of communication for people who have severe motor disabilities but being cognitively intact [1]. The EEG based BCI can be realised by using various types of input signals. For example, Visual Evoked Potential [18,19], motor imagination and P300 [3]. In this paper we concentrate in the development of a BCI based on the signals resulted from performing different mental tasks. The most commonly used task is motor imaginary, where the subject imagines a movement without actually performing it, and this has given encouraging results [5,6,11,17]. A motor

1. INTRODUCTION

In this paper we have also shown that good classification accuracy could also be obtained by just using signals from only one channel, which is either from C3 or C4. Figure 3 illustrates the signal analysis methodology of this paper. This method has been successfully implemented on the BCI Competition 2003 dataset IIIb, which consists of 140 labelled and unlabelled trials respectively, further information of the data set, is explained in the next section. The rest of the paper is organized into three sections. Section 2 discusses the methodology used to conduct the experiments. Whereas, section 3 and 4 comprises of the experimental results and conclusion respectively.

Then the adaptive recursive band pass filtering approach is applied to the signals of the pre-processed signal. The centre frequencies of the dominant signals from channel C3 and C4 are estimated and tracked by the employed filter. The main advantage of this approach is that the adaptive filter updates only one filter coefficient in order to adjust the centre frequency of the filter band pass to be matched with the input data [13,14]. The coefficients time function represents the features for each signal. A fourth order Butterworth band pass filter is employed as the adaptive filter where the filter function, T (z), could be expressed as [15,16]. T(z)=

2. METHODOLOGY
This method was experimented on the BCI Competition 2003, dataset IIIb [2]. The data set was provided by the Department of Medical Informatics, Institute for Biomedical Engineering, University of Graz. The signals was acquired from the EEG channels C3, Cz, and C4 (figure 2), which was band pass filtered for a frequency range of 0.5 to 30Hz and sampled at 128 Hz. The signals were acquired from a 25-year-old female relaxing on a chair with armrest. The subject was given a task to control a feedback bar by means of imagining left hand of right hand movements. The experiment consists of 7 runs with 40 trials each. All runs were conducted on the same day with several minutes break in between. The data set consists of 140 labelled and 140 unlabeled trials with an equal number of left hand and right hand movements. Each trial consists of duration of 9 seconds. At the 3rd second a visual cue, an arrow pointing left or right is presented to indicate left or right motor movements is to be imagined (figure 1).

2 4 D0 + D2Z + D4Z 2 1 2 3 4 1 + F C (n)Z + ( F2C (n) + F '2 )Z + F3C(n)Z + F4Z 1

(1)

2.1. DATA SET

where
D2 = 2 D0 , D1 = D3 = 4 D0 , F = 2l (2l + 2) D0 , 1 2 2 F2 = 4l D0 , F '2 = 2(l 1) D0 ,
F3 = 2l ( 2l + 2) D0 , 2 F4 = (l 2l + 1) D0 , l = cot an ( BP )

The coefficient C (n) could be expressed as

C (n) = cos( ( H 2 ( n ) + H1 ( n )) cos( BP )

(2)

where
H1 ( n ) = normalised low cut off frequency, H 2 ( n ) = normalised high cut off frequency,
BP

= normalised bandwidth of the filter, assumed constant.

sec

Feedback period with Cue Trigger Beep

Figure 1: The timing scheme.

Signals of channel C3 and C4 from the 140 labelled trials were pre-processed by means of band pass filtering respectively. A 7th order Butterworth band pass filter was employed with a pass band of 9Hz with less than 1 dB of ripple and a stop band of 11Hz with at least 6 dB of attenuation. Signals from channel Cz is ignored because it contains very little significant discriminative features [10].

2.2. FEATURE EXTRACTION USING ADAPTIVE BAND PASS

From equation 1 and 2, it can be seen that C ( n ) is the only coefficient that has to be updated by the adaptive filter since it is also the only coefficient which is dependant with the centre frequency, ( H1 ( n ) + H 2 ( n ) )/2. Figure 4 shows the realization of the features,. In order for the filter, T (z) to be self-adjusted to the centre frequency of the input signal, the output power of the filter should be maximised [15,16]. A standard gradient approach is implemented to update the adaptive coefficient so that the expected output power could be maximized [13, 14]. The filter coefficient is updated by using an algorithm called recursive maximum mean-squared [13, 14]. Instability of the adaptive filter could be observed for low frequency waves such as the mu-rhythm. The instability is because the adaptive filter is a band pass filter and the EEG signal is a low pass signal. This problem is solved by introducing a high frequency

shifting process to shift the EEG frequency to the highest ones before the adaptive filtering [13,14]. Finally, the updated co efficient is used as the feature that represents the signals resulted from right or left motor imaginary.

Table 1: Classification of motor imaginary respective to the 3 fold cross validation. The shuffle seed is set to 1.
Training fold 1, 2 2, 3 1, 3 Test fold 3 1 2 Average: Classification (%) C3 89.13 95.74 91.48 92.12 Classification (%) C4 95.65 97.87 97.87 97.13

The FA is a supervised NN, which has very fast convergence and the capacity for incremental learning, compared to other NN such as the Perceptron. The FA network comprises of two Fuzzy ART modules, ARTa and ARTb [7,12]. Each of the modules contains two neuron layers, F1 and F2. ARTa and ARTb are connected by a series of connection between the F2 layers. These connections are weighted, W and forms a module called as map field, Fab. Figure 5 shows the graphically illustrated FA network. In order to train the FA, ARTa receives a feature vector that represents a pattern and ARTb receives a feature vector that represents the class of the pattern. When ARTa receives the pattern, a category will be chosen through the ART competition, where the relevant ARTb pattern class will be also applied. A comparison of the categories of ARTa will be made to the ARTb prediction through the map field, Fab. The behaviour of ARTa is controlled by the vigilance parameter, a . If ARTb is not able to predict the correct output for ARTa then a is increased to force the creation of a new category. This process is called as match tracking. Further literature could be found in [7,12]. In this paper, we trained and tested the FA using features from channel C3 and C4 separately. This is done in order to find the classification accuracy by using features from only one channel. A 3 fold cross validation was also employed to validate this approach. The input features were shuffled according to seed values ranging from 1 to 10, where the seed value determines the order of the shuffle. Then the feature vectors are complemented in order to prevent category explosion and gives the networks its fuzzy nature. After this process, the features together with its labels are introduced to the FA for learning. The learning rate for FA is set to fast, a =1 and with a vigilance value of 0.75. Once the network is trained, the test data are introduced to the FA. The results are presented in table 1 and table 2.

2.3. CLASSIFICATION USING FUZZY ARTMAP

Table 2: Average classification results of motor imaginary respective to the shuffle seed values.
Shuffle seed Average classification (%) for 3 fold cross validation C3 C4

1 2 3 4 5 6 7 8 9 10

92.12 91.44 86.46 87.17 90.71 93.58 86.43 94.31 90.67 92.17

95.65 99.29 89.99 96.38 91.43 92.14 95.04 99.28 97.87 93.62

Input EEG signals from channel. C3 and C4

Band pass 7th order Butterworth filter

Pre Processed signal

Adaptive Band pass filter. (4th order Butterworth filter)

Filter co-efficient, C(n)

Fuzzy ARTMAP classification

Figure 3: Signal analysis methodology.

3 5 cm

RESULTS AND DISCUSSION

C3

Cz

C4

Figure 2: Electrode positions on the scalp

Separate experiments were conducted with bio signal features from C3 and also C4. Since the 3 fold cross validation approach was employed, the data set was divided into 3 segments with equal number of trials. Two of the segments were trained and one of it is tested. This leave one out (group) method is done so that all the 3 folds can obtain the classification results. Table 1 shows the classification results for shuffle seed=1. This experiment is repeated for the seed ranging from 1 to 10. Table 2 shows the average

classification results respective to the seed values. From the results, we could see that the classification percentage varies according to different seed values. We could also observe that the results for features from channel C4 gives the highest percentage of 99.29% while features from channel C3 gives the highest percentage of 94.31%. Experiments using the same dataset were also trialled in [9], where signals from C3 and C4 were used together from specific time period to provide the best classification accuracies. Besides that, the training dataset was also used in [10] to obtain an error rate of 7.9% by using the leave one out cross-validation method. Both of the method in [9,10] utilizes features from channel C3 and C4 together in order to provide better classification results. In contrast, the results obtained from our proposed method proves to provide better classification accuracy by utilizing features from only one channel, which is either from C3 or C4.

Map field

Wab

Fab

a F2

F2

Wa
Fa 1

Wb
F 1
b

ARTa

ARTb

Figure 5: Fuzzy ARTMAP architecture

REFERENCES
0.4 0.2
Amplitude

0 -0.2 -0.4 0

4
Time(s)

0.1
Feature C(n)

0 -0.1 -0.2 -0.3 0 1 2 3 4

Time(s)

Figure 4: 10 realisation of raw motor imaginary signals (Top). Mean values of the extracted features after performing adaptive recursive band pass filtering (bottom).

CONCLUSION

As for the conclusion, we could observe that the features from the motor imaginary signals extracted using band pass filtering and adaptive filtering could be classified easily using FA. This method also enables us to obtain classification results by only using features from one channel at a time. For example, we could observe better classification accuracies from C4 rather than C3. Furthermore, this method is found suitable for single trials. This technique can be also extended in identifying other useful sources and components buried in the recorded EEG signal.

We would like to thank and acknowledge the Department of Medical Informatics, Institute for Biomedical Engineering, University of Graz for making the dataset available.

ACKNOWLEDGEMENT

[1] Baharan Kamousi, Zhongming Liu, and Bin He, 2005, Classification of Motor Imaginary Tasks for Brain-Computer Interface Applications by Means of Two Equivalent Dipoles Analysis, IEEE Trans.Neural Sys.and Rehab. Engineering,Vol 13. [2] BCICompetitionII,http://ida.first.fraunhofer.de/pro jects/bci/competition_ii, 2003. [3] E.Donchin,K.M.Spencer, and R.Wijesinghe., The mental prosthesis:Assesing the speed of a P300based brain computer interface,IEEE Trans.Rehab.Eng.,Vol 8,pp.174-179. [4] E.Curran, P. Sykacek, M. Stokes, and S. J. Roberts, Cognitive tasks for driving a braincomputer interfacing system: A pilot study, IEEE Transactions on neural systems and rehabilitation engineering, vol. 12,no. 1, mar 2003. [5] F. Babiloni, F. Cincotti, L. Lazzarini, J. Millan, J. Mourino, M. Varsta,J. Heikkonen, L. Bianchi, and M. G. Marciani, Linear classification of lowresolution eeg patterns produced by imagined hand movements,IEEE Trans. Rehab. Eng., vol. 8, June 2000. [6] G. P. D. Flotzinger and J. Kalcher, Braincomputer interface - a new communication device for handicapped people, Electroencephalography and clinical neurophysiology. Supplement, vol. 16, 1993. [7] G. A. Carpenter, S. Grossberg, and J. H. Reynolds, A Fuzzy ARTMAP nonparametric probability estimator for nonstationary pattern recognition problems, IEEE Trans. Neural Networks, vol. 6, pp. 3301336, 1995. [8] J. Kalcher and G. Pfurtscheller., Discrimination between phase-locked and nonphase-locked eventrelated EEG activity, Electroenceph. Clin. Neurophysiol., vol. 94, pp. 3814, 1995. [9] J. Vickneswaran, S. Andrew, CK. Loo, Classification of Motor Imaginary Tasks Using Adaptive Recursive Bandpass Filter, Proc. International Conference on Signal Processing and

Multimedia Applications(SIGMAP2008),pp. 113118, 2008. [10] Lemm, S.; Schafer, C. and Curio, G., BCI competition 2003-data set III: probabilistic modeling of sensorimotor mu- rhythms for classification of imaginary hand movements, Biomedical Engineering, IEEE Transactions Vol 51, pp.1077 1080, 2004. [11] N. Birbaumer, A. Kuebler, N. Ghanayim, J. Hinterberger, J. Perelmouter, J. Kaiser, I. Iversen, B. Kotchoubey, N. Neumann, and H. Flor, The thought translation device (ttd) for completely paralyzed patients, IEEE Trans. Rehab. Eng., vol. 8, 2000. [12] R.Palaniappan, R.Paramesran, S.Nishida, N. Saiwaki, A new brain-computer interface design using fuzzy ARTMAP. IEEE Trans. Neural Systems Rehabil. Eng. v10 ,No3. 140-148, 2002. [13] R. R. Gharieb and A. Cichocki., Segmentation and tracking of the electro-encephalogram signal using an adaptive recursive bandpass filter, Medical & Biological Engineering & Computing 2001, Vol. 39, pp. 237-248, 2001. [14] R. R. Gharieb and A. Cichocki, On-line EEG classification and sleep spindles detection using an adaptive recursive bandpass filter, In proceedings Acoustics, Speech, and Signal Processing (ICASSP01) IEEE International Conference, Vol 2, 2001. [15] R.V. Raja Kumar and R.N. Pal, A gradient algorithm for center-frequency adaptive recursive band pass filters, Proc. IEEE,Vol 73,pp.371372,1985. [16] R.V. Raja Kumar and R.N. Pal., Tracking of bandpass signals using center-frequency adaptive filters, IEEE Trans. Acoust. Speech signal processing., vol 38, pp.1710-1721, 1990. [17] S. J. Roberts and W. D. Penny, Real-time brain computer interfacing: A preliminary study using bayesian learning, Medical & biological engineering & computing., vol. 38, no. 1, 2000.

[18] S.Andrews, R.Palaniappan, and N.Kamel., Extracting Single Trial Visual Evoked Potentials using Selective Eigen-Rate Principal Components, In Proceeding of World Academy of Science, Engineeering. and Tech, Vol 7,2005. [19] S.Andrews, T.Andrew and C.K.Loo, Inquisitive Impact
of Different Electrode Configuration in Representing Single Trial P300 Visual Event Related Potentials,,College Science in India.,Vol 2007,2007.