实时BCI脑电信号分类的应用

� EEG Signal Classification for Real-Time Brain-Computer Interface Applications: A Review A. Khorshidtalab, M.J.E. Salami Department of Mechatronics Engineering, International Islamic University Malaysisa, Gombak, Malaysia E-mail: aida.khorshid@student.iium.edu.my; momoh@iiu.edu.my Abstract-Brain-computer interface (BCI) is linking the brain activity to computer, which allows a person to control devices directly with his brain waves and without any use of his muscles. Recent advances in real-time signal processing have made BCI a feasible alternative for controlling robot and for communication as well. Controlling devices using BCI is a crucial aid for people suffering from severe disabilities and more than that, BCIs can replace human to control robots working in dangerous or uncongenial situations. Effective BCIs demand for accurate and real-time EEG signals processing. This paper is to review the current state of research and to compare the performance of different algorithms for real-time classification of BCI- based electroencephalogram signals. Keywords -Brain Computer Interface, EEG, real- time signal processing. I. INTRODUCTION Brain-Computer Interface (BCI) asks user for brain signals instead of any muscular activities. This system enables people with severe motor disabilities to send command to electronic devices by help of their brain waves. Signals should be identified, processed, and classified to specific command. Feature extraction and classification methods are playing the main role in any BCI systems; since any misclassification and error may cause a wrong command. In the past few years, many research groups focused their work on classifying EEG records to desired mental task classes [1][2][3]. Several algorithms have been investigated by purpose of increasing the classification rate and accuracy of evoked potential- based BCIs. Despite the improvements that have been achieved in this area, on-line BCI still poses some challenges [4]. In this paper, we review the performances of different models for classification of BCI-based electroencephalogram signals regarding their real-time applications. The rest of paper is organized as follows. Section II reviews EEG signal classification and artifacts. Section III discusses different classification and feature extraction methods. Section IV presents parameters that can affect processing of mental tasks. Section V compares between widely applied classifications algorithms, and section VI concludes our study.� II. EEG SIGNAL CLASSIFICATION AND ARTIFACTS Recorded brain electrical waveforms are associated with electrical potentials which are not originated in brain. The sources of these electrical potentials are eye blinking, eye movement, activity of heart and muscles in general. They also can be from EEG equipments or recording systems. These interference waveforms known as artifacts can often cause serious misclassifications. Hence, developing a practical real-time system to recognize and eliminate artifacts is essential. 2011 4th International Conference on Mechatronics (ICOM), 17-19 May 2011, Kuala Lumpur, Malaysia 978-1-61284-437-4/11/$26.00 ©2011 IEEE � A method proposed by Bogacz et. al [5] is blink artifacts recognition using artificial neural network. The input they applied for the Neural Network was different coefficients which express characteristic properties of blinking artifacts based on their knowledge about the artifacts recognition. The major property of their method is using a large training dataset. Three classification algorithms including K- Nearest Neighbors (KNN), Radial Basis Function (RBF) networks – an artificial neural network with radial basis function as its activation function- and back propagation networks were examined using this method. The result is satisfying; however, the use of a large training dataset can be an inconvenient parameter for real-time brain computer interface. By assuming that, the EEG signal recorded from skull is a linear combination of EEG and EOG (signals produce by eye movement or movement of eyelid).Subtracting EOG which is recorded concurrently and separately from recorded EEG, can recover the EEG signals produce by brain. This method also has been used for eliminating BCG (recoil of human body due to momentum of the blood the heart is currently pumping). Nevertheless, subtracting the average heart beat waveform can cause error in classification since the recovered signal is not matched with the original brain signal. Average Artifact Subtraction (AAS) is based on subtracting an average artifact template which is generated considering the repetitive pattern of artifact. This method asks for high sampling frequency and is just capable of eliminating repetitive artifact patterns [6][7]. Another method is artifact rejection based on peak elimination which eliminates the corrupted sample from all EEG signals if the energy of the signal exceeds the considered threshold value for blinking. Result from this method is satisfactory while it is simple to apply. Thus, peak elimination method can be considered as an initial option for a real-time brain computer interface. Rejection method can be unacceptable when the amount of data lost is much, blinking happens too frequently or the available data is limited [1]. Independent Component Analysis (ICA) has proved that can deal with artifact since it needs neither assumption properties of the source signal nor the use of reference signal. It is able to perform better than standard methods of artifact elimination and reconstruct the signal in a promising way [8]. Regression technique tends to be simple and robust with any number of EEG channel. Moreover, for limited number of EEG channel high probably performs better than ICA [9]. Rejection method needs less computation compare to other approaches like regression approach; thus, it is a favorable alternative compare to other methods [1].� III. CLASSIFICATION AND FEATURE EXTRACTION METHODS Classification algorithm can directly affect the BCI behavior. Therefore, any improvements have a significant impact on the real-time brain computer interface systems. In order to obtain excellent classification result, effective methods of feature extraction is necessary [10]. To make decision about the classification method, it is essential to know what the features are, what is their application and in which way they may help classification. Feature extraction can be burden for BCI systems and make the classification process complicated and computationally costly. There are cases that some features are redundant or not enough discriminate to available data. Therefore, feature reduction helps for better result as classifier learn a robust solution and achieve a better performance. They have been introduced some classification models for EEG signals with capability of robust and accurate classification of raw EEG signal without feature extraction in prior step [11]. There are quite a number of different features set for different designs and applications of BCI systems such as Band Power (BP), Power Spectral Density (PSD) values, Auto Regressive (AR) and Adaptive Autoregressive (ARR) parameters, amplitude of EEG signals, Time –frequency feature,� CerebroSpinal Fluid (CSF) and fractal dimension. Applying combination of features for EEG signals is a new trend of BCI research. Several experiments which show that a combined set of feature produce a better classification rate than using feature independently [12]. Combination of adaptive band pass filters and adaptive autoregressive (ARR) for classification of right and left motor imaginary signal was a successful experiment [13]. � A better classification result is also reported by combining adaptive radial-basis function (ARBF) and AAR especially with PCA and LDA as classifiers [13]. For classification and analyzing EEG signals different method have been proposed ,namely, neural networks [14], statistical methods [15], autoregressive model [16][17], mixture of densities approach [18], independent component analysis [19], time-frequency analysis [20], Bayes quadratic, Hidden Markov Model[3], and Linear Discriminant Analysis(LDA)[3]. Pfurtscheller et. al [21] applies neural network for on-line classification of specific frequency bands for separating of left and right motor imagery. Using off- line adaptive autoregressive model a considerable improvement is achieved. To apply this method artifact has to be controlled because of its sensitivity to noise. Two different neural networks topologies are compared by Haselsteiner and Pfurtscheller [22] for classification of single trial EEG data. Multi Layer Perceptron (MLP) and FIR MLP which is a MLP that the static weight of it, is replaced with finite impulse response filters. Their experiment shows that FIR MLPs has less error and performs better. They also compare feed forward neural network, Bayesian quadratic, Bayesian network, Fisher linear classifier and Hidden Markov Models for EEG signal classification. Since these classifiers cover both linear and nonlinear methods, a comparison between them can lead us to a quite comprehensive conclusion [23]. In general Bayesian quadratic classifier performed better than Bayesian network. However, it has been reported that in one case Bayesian network gives better results. In case of comparing Bayesian network with Neural Network, Fisher linear classifier and HMM, it also delivered a better result. Although it is necessary to mention that in term of execution time and time to be trained for classifying EEG signals, Bayesian quadratic classifiers and Fisher linear classifier requested for minimum time while the most time consuming classifier is Bayesian network [24]. One of the remarkable proposed online learning algorithms is hybrid recursive least squares estimator. This structure is a dynamic structure that adds neuron autonomously and optimal brain surgeon technique is the inspiration of this structure. This algorithm is able to classify EEG signals in a truly on-line manner. Self-Organizing Fuzzy Neural Network (SOFNN) is another hybrid system that has the advantage of improving some aspects of neural network performance using fuzzy logic theory. Besides, this structure is self organizing structure and does not need a specified topology to perform a particular task. Another advantage of this method is compatibility of SOFNN and non-stationary nature of EEG signal. It also nicely deals with complexities, nonlinearities and its large dimension. This technique is even more suitable for EEG applications as it is able to perform quite well with limited data or expert knowledge [25]. Ensemble methods and combining several classifiers in different ways is a recent, successful trend in this area of research. Boosting, voting, stacking, bagging and subspace have showed better performances in comparison to applying a single classifier [26]. Boosting is applying several classification algorithms in cascade where each of them takes care of errors carried out by the earlier ones. AdaBoost uses kernel to construct linear classification boundaries in higher dimensional space. Adaboost calls a given base learning algorithm which can be any algorithm that performs slightly better than random guessing in series of rounds. For each rounds there is a weight. Initially, all the weights are equal but in each round a higher weight will be assigned to incorrectly classified examples. Therefore, the base learner has to focus on the hard examples. In this way, a weak learning algorithm is boosted into an accurate and strong learning algorithm. This method has been examined successfully by decision tree as a base learner [11][27].Yet, this method is computationally heavy for real-time BCI systems. Bagging is less complex and faster than boosting; although, boosting performs better processing noisy data [11]. Using several classifiers, while each one allocates the input feature vector to a class and chooses the final class based on the majority votes is called voting method. Majority voting can be simple or weighted voting, when there are some classifiers more qualified than the others. Therefore, heavily voting them may improve the overall performance of classification. � Stacking consists of several classifiers in its level- 0 plus another classifier in level-1. Classifying the input feature vector is by each of level-0 classifier and making the final decision is by level-1 classifier or meta-classifier. It has been reported that stack generalization performs significantly better than SVM in term of accuracy and also better than majority voting [28]. Several papers state that classification error and variance is less by combination of classifiers. Variance can vary with time, subjects, and sessions and is one of the most important causes of misclassification and error that is avoidable with help of combination of classifiers. IV. EFFECTIVE PARAMETERS EEG signal is one of the most complicated and sensitive bio signals that can be affected with many different parameters. In this section, some of the considerable parameter in term of EEG signal classification is discussed. One of the fundamental parameters is to control mu wave amplitude. Mu wave is an EEG wave associated with motor activities. In most recent researches, classification algorithm has been tried on mu and beta waves. Mu wave has the same frequency as alpha wave. It has been stated in literature that people are usually able to learn how to change the amplitude of this wave through suitable mental efforts. Wolpaw and his colleagues [29] trained normal subjects to make them capable of controlling cursor. Mu wave control does not require subjects to have any other motor control. The Graz Brain- Computer Interface group trained people to control the amplitude of their ERS/ERD patterns. To learn how to control their brain waves and therefore the cursor, up to few sessions were needed for some subjects. Even with help of few sessions training not all subjects could control the cursor accurately. Two out of six subjects were not able to perform the cursor control task [30]. The possibility of controlling BCI accurately with help of visual feedback for inexperienced subject shows that seven of ten subjects were able to choose the correct target. Surprisingly, six of ten subjects did not do any mistakes which show the potential for practical applications [31]. It has been mentioned in the many papers that not all subjects were able obtain control via their brain signals [5][27]. About one third of BCI users could not achieve BCI communication at all [32]. One of the comforts of classification in on-line mode is that, subject is able to modify their EEG signals based on given feedback. Classification sharpness is expected with more number of on-line experiments along with feedback [21].When users see the changes in their performances they are able to adjust their imagination and the amplitude of their mu waves. Use of visual feedback has proved to be significantly helpful for an accurate performance [21][31]. V. COMPARISION A review of literature shows that among all different methods that have been applied for feature extraction and classification, some of them attracted more attention and are widely used under different conditions. Varieties of different preprocessing and feature extraction technique for the same algorithms of classification have been used. A comparison between these commonly employed classifiers gives us a clear vision about their average performances and accuracy in different situations, when the other parameters like the feature extraction methods, pre- processing methods, and source EEG signal records vary. In this section five tables represent the accuracy of five major classification methods, namely, LDA, MLP, HMM, SVM, and Bayes quadratic. LDA: Linear Discriminant Analysis is considered to be a stable classifier because of its low complexity and not being affected by small variation. It is the most popular algorithm applied for BCI application. LDA is simple to use and does not need complicated computation and more importantly, tend to be a right choice for real-time classification. Moreover, in many trials for BCI systems has been successful. Fig. 1, shows the classification accuracy rate for different analysis of LDA that vary having preprocessing (p), not having preprocessing (n), and having feature extraction (f). � Fig. 1 Classification accuracy for LDA SVM: Support Vector Machine is a statistical learning theory with margin principal. Based on the choice of Kernel function it can be among linear or nonlinear classifiers. Compare to LDA it is quite insensitive to overtraining due to margin maximization. In many BCI systems SVM exhibited very small errors. The most commonly used SVM is Gaussian SVM. SVM reaches the full accuracy with STFT (Short-Time Fourier Transform) and Wigner- Ville TF kernels [33]. Fig. 2, shows the classification accuracy rate for different experiments of SVM that vary having preprocessing (p), not having preprocessing (np), having feature extraction (f), not having feature extraction, or using Gaussian SVM (G). Fig. 2, Classification accuracy for SVM HMM: Hidden Markov Models has been mostly applied for speech recognition. For the classification of time series, HMM is successful while for BCI system it has not been widely employed and is not as satisfactory as the other algorithms. Fig. 3, shows the classification accuracy rate for different trials of HMM that vary having preprocessing (p), not having preprocessing (np), having feature extraction (f) or not having feature extraction and applying only raw EEG records. Fig. 3, Classification accuracy for HMM MLP: MultiLayer Perceptron is the most popular Neural Networks that has been used for BCI. Neural Network has been generally successful among new computing tools which have been employed to process EEG signals in BCI systems. Basically for Neural Networks topology and its architecture should be carefully design for an optimum performance. MLP has high complexity and small variations can cause unexpected result. Fig. 4, shows the classification accuracy rate for different experiments of MLP that vary having pre-processing (p), not having pre-processing (np), having feature extraction (f) or not having feature extraction. Fig. 4, Classification accuracy for MLP Bayes quadratic: It is based on computing the likelihood of each class considering the feature vector and selecting the most likely one. Bayes quadratic assigns the EEG signal to the most probable possible class. This classifier is not so popular for BCI. However, it has been accurate enough for classification of motor imagery compare to other not so widespread methods. Fig. 5, shows the � classification accuracy rate for different experiments of Bayes quadratic algorithm that vary having preprocessing (p), not having preprocessing (np), having feature extraction (f) or not having feature extraction. Fig. 5, Classification accuracy for Bayes quadratic VI. CONCLUSION In this paper, we review different methods and algorithms for EEG signal classification with considering principal parameters and conditions that affect real-time BCI system. Among neural networks models, SOFNN shows a better agreement with EEG signals nature. Therefore,