TMCnet News
Comparative Study of Data Classification Methods Between EEG and ECoG Used to BCI [Sensors & Transducers (Canada)](Sensors & Transducers (Canada) Via Acquire Media NewsEdge) Abstract: Effective decoding of the source signal is a key to improve Brain-computer interfaces (BCI) performances. Two groups of motor imagery (MI) data based on electroencephalograms (EEG) and electrocorticograms (ECoG) which provided by International Brain-Computer Interface Competition organization are analyzed, and concluded that ECoG signals processing is more suitable for model-driven approaches. Temporal-frequency features were extracted by model-driven method instead of data-driven method and compared, and classified by support vector machine (SVM). The results show 6 % improvement of motor imagery experiment classification accuracy on ECoG data, compared with of data-driven method. Copyright © 2014 IFSA Publishing, S.L. Keywords: Brain-Computer Interface (BCI), Electrocardiogram (ECoG), Motor Imagery (MI), Electroencephalogram (EEG), Support Vector Machines (SVM). (ProQuest: ... denotes formulae omitted.) 1. Introduction In recent years, BCI technology has aroused extensive concern and gained rapid development. EEG signal is used to realize human-computer interaction with no-motion, a new means of communication which does not depend on normal peripheral nerve and muscle output channel is provided to patients that have nerve and muscle damage [1]. Therefore, BCI technology study has important social value and bright application future. At present, BCI system has variety methods of acquisition signal, EEG is commonly used. ECoG, Magneto encephalography (MEG), Near-infrared Spectroscopy (NIRS), Positron Emission Tomography (PET), functional Magnetic Resonance Imaging (fMRI) and et al are used in research and clinical application. EEG signal is cerebral cortical potential recorded extracranial scalp surface with the help of metal electrode (Ag/AgCl) and conductive adhesive. There existing signal distortion, making positioning accuracy of the signal source poor and describing signal model is more difficult for the conductivity differences of scalp and non-homogeneous brain tissue. ECoG signals are neuronal activity reflection by electrodes arrays embedded in epidural under the brain skull, collecting electrode is more closer to signal source comparing the EEG mode, amplitude of the received signal is high and artifacts is fewer, and signal to noise ratio is relatively high, spatial resolution ratio is up to millisecond [2]. In addition, ECoG signals do not need to cross brain tissue (meninges, blood, skull and scalp) with the low-pass filter functional, ECoG signal frequency band is wider than EEG signal frequency band. Therefore, ECoG signals model is easier to establish than EEG signal model. Motor imagery BCI system is established based on Event-Related Desynchronization (ERD) of p rhythm and ß rhythm. The principle is that, 8-12 Hz brain electrical activity is occur in primary sensory cortex and motor cortex when a sober man with not sensory input and motor output, the signal to visual cortex called a rhythm, and called p rhythm to somatosensory or motion cortex. These p rhythms are usually related to ß rhythm of 14-25 Hz. p rhythms and ß rhythms are reduction when in motion or warming-up, especially in the moved limb opposite area, this phenomenon is called ERD. Even more interesting, ERD phenomenon can also be occur only in MI and no real action [3], it does not in any way dependent on the brain's normal output channels, so independent BCI system can be supported. In order to seek fast and effective method of extracting p rhythm and ß rhythm signal feature, MI BCI systems signal processing method is studied based on EEG, EcoG, and to increase BCI systems classification accuracy in this study. 2. MI Experiment A healthy woman, sitting comfortable and relaxed, is test object in MI experiment based on EEG. Experimental tasks are which imaging hand left and right move to control the cursor left and right move around based on randomly hint on computer screen. All data recorded by 3 electrodes and sampling frequency is 128 Hz. Each record has a length of 9 s; the first 2 s is preparatory stage; the left arrow or the right arrow hints would appear in the third second, motor imagery is carried according to hints. Experiments is completed in one day, there is 7 rounds, 40 experiments (with several minutes intervals between two adjacent experiments) in each round and 280 samples in total. Samples were divided into two groups in average; one is training samples of known category, the other is test samples of unknown category. A man, suffered from Epilepsy, is test object in MI experiment based on ECoG. To position the lesion, 8X8 ECoG electrode arrays (64 channels) is implanted in right cerebral motor cortex, number and location of electrode array are determined by clinical need, without considering BCI technology needs. The man is sitting comfortable and imaging left hand little finger or tongue motion based on symbol or character hint on computer screen. Each experiment sustained 3 s and sampling frequency is 1000 Hz. Data are recorded 0.5 s after appeared hint in order to remove ECoG artifacts, 278 training samples and 100 tested samples are collected eventually. 3. Data Processing Methods At first, feature is extracted to the collected motor imagery data, and then classified by classifier; movement intention of participant can be inferred from. Feature extraction methods can be divided into data-driven and model-driven. Data-drive is used for data with many structures or models, and the model is difficult to define, law can be found from large amounts of data; model drive is used for that data characteristics is obvious and the model is determined. Common Spatial Pattem (CSP) [4-5] is adopted in feature extracting to movement-related EEG signal based on ERD, which is data-driven method; Linear Discriminant Analysis (LDA) and SVM [6-7] are used for Classifiers. Identification accuracy is used as an evaluation index of the algorithm. 3.1. Feature Extraction Pre-treatment which is band pass filtering is carried on band of p rhythm and ß rhythm before feature extraction. In addition, MI tasks of participant are carried 3 s after EEG MI started; only data in the last 6s can be used; sampling frequency of data based on ECoG is 1000 Hz, sampling factor is 10. Each experiment data are divided into training samples (known categories) and testing samples (unknown categories), training dataset is separated to twosample using known categories of training sample, sample matrix of training dataset is obtained. Two methods are adopted in feature extracting of EEG signal and ECoG signal, one is data driven method, p rhythm and ß rhythm feature are extracted and analyzed respectively from source signal data; the other is model driven method, feature is extracted by identified transcendental knowledge. Data driven process is which, at first, the data is filtered by CSP method, Continuous Wavelet Transform (CWT) is followed, and T weighting method is used to extract feature. Data feature is extracted only by CSP method in model driven method. CSP is one of spatial filtering methods in dealing with EEG signals of BCI system, and apply to extract features of p rhythm and ß rhythm, the treatment processing would be better for definite signal with high spatial resolution. The goal adopted CSP is to make two kind samples had the most differences, namely, two corresponding empty domain filters are constructed, and make the variance of one type signal after projection is the biggest and the variance of the other type signal after projection is smallest. CWT and T weighted methods are belonging to data driven; their goal is to search law from large amounts of data. 3.2. Design of SVM Classifier SVM is developed from linearly separable optimal classification [8], whose objective is to find an optimal classification; optimal classification can separate two samples and classified interval y is biggest. That is, to sample dataset (Xi^,),i= 1,* * * ,nA^RJ, y e{-1,+1}, meet ... (1) And make classified interval y=2/ltol been biggest, where, to is projection vector of classification, b is offset. In practical, SVM optimization problem has two forms, C-SVM and v-SVM; v-SVM is adopted to design classifier in this study. The expression is ...(2 where v(0<v<l) is the proportion upper bound of the wrong samples to the total number of samples and the proportion lower bound of the support vector numbers to the total number of samples. When Ç=0, classified interval between two types of data happens to be 2/9/lcol. At first, classifier is trained; Libsvm Toolbox is adopted in specific algorithms. The training dataset feature in formula (2) x is obtained using feature extraction, column vectors y in formula (2) is obtained according to known categories of training dataset, to and b can be got after training by SVM, at last, using discriminant function ... (3) The test dataset feature x is substituted to judge the test dataset classification results. Where Ns is the number of support vector, &(x,x,) is kernel function, Radial Basis Function (RBF) is used in this study. 3. Results Fig. 1 and Fig. 2 show the time-frequency figures of the first samples in the two sets experimental data after CWT. As can be seen from Fig. 1 and Fig. 2, EEG data have clear time-frequency structure after CWT, ECoG data only shown features in frequency domain, the features in time domain is scattered, with no obvious divisibility. There is little significance in time-frequency transforming to ECoG data by CWT. Therefore, CWT and T weighting method of datadriven is used in feature extraction of EEG signals, CSP method of model-driven is only used in feature extraction of ECoG signals. Fig. 3 shows the features of EEG signals extracted by data-driven, Fig. 4 shows the features of EEoG signals extracted by model-driven (interception data in one second). It can be seen clearly that energy difference of two features signals in different directions, that is, the separability of extracted features is good, it is better for designing classifier from Fig. 3 and Fig. 4. 4.1. Classification Results Obtained in Different Band Filtering Signal preprocessing goals for BCI systems based on ERD MI is hope to filters out components reflected EEG power changes on the band of p rhythm and ß rhythm. Filter on different bands is carried to two sets data, with better feature extraction methods respectively. Namely, data driven method is adopted in feature extracting for EEG signal, and model driven method is adopted in feature extracting for EEoG signal, then classified by v-SVM, the classified results is shown in Table 1. It can be seen from Table 1 that effects on final classification accuracy in different frequency band filtering, ones that classification accuracy is 89.3 % and 86 % on frequency band of p rhythm (8-12 Hz), 75.7 % and 81 % on frequency band of ß rhythm( 14-25 Hz). And, classification accuracy is 78.6 % and 78 % respectively on high band of p rhythm (10-12 Hz), the highest classification accuracy is 89.3 % and 92 % respectively, the bandwidth is 2.5-25 Hz, band width of p rhythm and ß rhythm is contained. 4.2. Classification Results Obtained in Different Processing Methods The best classification results of test set are obtained by using different signal processing methods for two sets of experimental data, (Table. 2). 5. Discussion 5.1. Discussions on Signal Processing Method The best classification results in Table 2 are obtained by using different data processing methods; classification results obtained by classifier v-SVM as an example, classification accuracy is relevant to feature extraction methods under the same condition of preprocessing filter bandwidth and classifier design method about two sets data. Experimental results based on ECoG signal data show that the best classification results is obtained only by using CSP for feature extraction, experimental results based on EEG signal data show that classification accuracy is only 81.4 % only by using CSP for feature extraction, the difference is large by comparing the best classification accuracy. Possible causes are as follows: 1) Formation mechanism of EEG signal and ECoG signal is different. EEG signal and ECoG signal have fundamental differences due to difference methods in signal collecting. Therefore collecting area of EEG signal is large, other unforeseen information exist in extracted movement-related signal, and making EEG signals had many structures or models. In addition, EEG signal can occur drift in crossing multiple brain tissues, and some sportsrelated artifacts (such as EMG) is accompanied, signal-to-noise ration (SNR) and spatial resolution are low [9-10]. For these reasons, EEG signal is difficult to describe by determined transcendental knowledge, and EEG model is difficult to determine. CSP is a spatial filtering method, the extraction features effect of p rhythm and ß rhythm is ideal from EEG signals with the high spatial resolution; features extraction of EEG signals solely by CSP is not enough and further processing by which using datadriven method such as CWT and T weighting is need, feature law is extracted from the data. ECoG signals are gathered only in cerebral motor area and electrode are close to signal source, SNR and spatial resolution are high, the model is determined, CSP can be used only in feature extracting. 2) Time - frequency property of EEG signal and ECoG signal is different. EEG signals collected area is more dispersed than ECoG signals', the collected time of the former is longer than the latter, their time resolution is better, EEG signals is better than ECoG signals in time axis. It can be seen that SVM classification accuracy is higher than LDA to same characteristics from Table 1. For one thing, SVM has certain advantages and better robustness in overcoming "dimension disaster" and "over-fitting"; for another, the classifier merits will directly affect the system classification accuracy. 5.2. Discussions on Frequency Bandwidth of Preprocessing It can be seen that from Table 1, frequency band range of original signal in processing BCI signal based on MI must include band of p rhythm (8-12 Hz), otherwise, classification accuracy declined significantly. Classification accuracy reaches the highest on p rhythm frequency band for EEG signal; classification accuracy is only 75.7 % on ß rhythm frequency band for EEG signal, classifier classification accuracy does not increase with the band widened including band range p rhythm and ß ihythm, and which show that frequency band of EEG signal is narrow, does not reflect the role of ß rhythm; classifier classification accuracy is higher with frequency band widened of EcoG signal, in addition, classification accuracy reached 81 % on ß rhythm frequency band for ECoG signal, and which showed that ECoG signal can embody the role of ß rhythm, which is also in line with feature of ECoG signal is wide, effects of ß rhythm on the final classification accuracy is less than p rhythm's. These show clearly that p rhythm provides the main information on the classification, MI classification is carried only extracting feature of p rhythm for BCI systems of MI based on EEG signals, which is illustrated in literature [11-12]. 6. Conclusions 1) Data collected of EEG signals is simpler than ECoG signals, yet signal quality of EEG signals is poorer than ECoG signals', data-driven method is used in feature extracting; 2) ECoG signals has good feature, it is easy to deal with, feature extraction can take advantage of model-driven method, and feature separability is good; 3) A suitable kernel function can make classification accuracy of SVM greater than LDA method. References [1] . G. Schalk, J. Dennis, H. Thilo et al., BCI2000: A General-Purpose Brain-Computer Interface (BCI) System, IEEE Transactions Biomedical Engineering, Vol. 51, Issue 6,2004, pp. 1034-1043. [2] . J. A. Wilson, E. A. Felton, P. C. Garell, et al., ECoG Factors Underlying Multimodal Control of a BrainComputer Interface, IEEE Transactions Rehabilitation Engineering, Vol. 14, Issue 2, 2006, pp. 246-250. [3] . G. Pfurtscheller, Lopes da Silva F. H., Event-related EEG/MEG Synchronization and Desynchronization: Basic Principles, Clincal Neurophysiology, 1999. [4] . Suk, Heung-Il, Lee, Seong-Whan, A novel Bayesian framework for discriminative feature extraction in brain-computer interfaces, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 35, Issue 2,2013, pp. 286-299. [5] . Homer, Mark L., Nurmikko, Arto V., Donoghue, John P., et al., Sensors and decoding for intracortical brain computer interfaces, Annual Review of Biomedical Engineering, Vol. 15, Issue 7, 2013, pp. 383-405. [6] . S. Lemm, B. Blankkertz, G. Curio, K.-R. Müller, Spatio-spectral Filters for Improved Classification of Single Trial EEG, IEEE Trans. Biomed. Eng., Vol. 52, Issue 9,2005, pp. 1541-1548. [7] . G. Domhege, B. Blankertz, G. Curio, and K. Müller, Boosting Bit Rates in Noninvasive EEG Single-Trial Classifications by Feature Combination and Multiclass Paradigms, Transactions Biomedical Engineering, Vol. 51, Issue 6, 2004, pp. 993-1002. [8] . N. Cristianini, J. Shawe-Taylor, An Introduction to Support Vector Machines and Other Kernel-based learning Methods, Cambridge University Press, 2000. [9] . W. J. Freeman, M. D. Holmes, B. C. Burke, S. Vanhatolo, Spatial Spectra of Scalp EEG and EMG from Awake Humans, Clincal Neurophysiology, 2003. [10] . R. Srinivasan, P. L. Nunez, R. B. Silberstein, Spatial Filtering and Neocoritical Dynamics: Estimates of EEG coherence, IEEE Transactions Rehabilitation Engineering, Vol. 45, Issue 4, 1998, pp. 814-826. [11] . D. J. McFarland, J. R. Wolpaw, Sensorimotor Rhythm-Based Brain-Computer Interface (BCI): Feature Selection by Regression Improves Performance, IEEE Transactions Rehabilitation Engineering, Vol. 13, Issue 3,2005, pp. 372-379. [12] . Yang, Bang-Hua, He, Mei-Yan, Liu, Li, et al., EEG classification based on batch incremental SVM in brain computer interfaces, Journal of Zhejiang University (Engineering Science), Vol. 47, Issue 8, 2013, pp. 1431-1436. Yu Ge, Xinfeng Ge, College of Mechanical & Electrical Engineering, Xuchang University, No. 88 of Bayiroad-Xuchang Henan, 461000, China Tel: +86-374-2968700, fax: +86-374-2968708 E-mail: [email protected] Received: 27 June 2014 /Accepted: 29 August 2014 /Published: 30 September 2014 (c) 2014 IFSA Publishing, S.L. |