{"title": "Improving Transfer Rates in Brain Computer Interfacing: A Case Study", "book": "Advances in Neural Information Processing Systems", "page_first": 1131, "page_last": 1138, "abstract": null, "full_text": "Improving Transfer Rates in Brain Computer\n\nInterfacing: A Case Study\n\nPeter Meinicke, Matthias Kaper, Florian Hoppe, Manfred Heumann and Helge Ritter\n\nUniversity of Bielefeld\n\nBielefeld, Germany\n\n{pmeinick, mkaper, fhoppe, helge} @techfak.uni-bielefeld.de\n\nAbstract\n\nIn this paper we present results of a study on brain computer interfacing.\nWe adopted an approach of Farwell & Donchin [4], which we tried to\nimprove in several aspects. The main objective was to improve the trans-\nfer rates based on of\ufb02ine analysis of EEG-data but within a more realistic\nsetup closer to an online realization than in the original studies. The ob-\njective was achieved along two different tracks: on the one hand we used\nstate-of-the-art machine learning techniques for signal classi\ufb01cation and\non the other hand we augmented the data space by using more electrodes\nfor the interface. For the classi\ufb01cation task we utilized SVMs and, as mo-\ntivated by recent \ufb01ndings on the learning of discriminative densities, we\naccumulated the values of the classi\ufb01cation function in order to combine\nseveral classi\ufb01cations, which \ufb01nally lead to signi\ufb01cantly improved rates\nas compared with techniques applied in the original work. In combina-\ntion with the data space augmentation, we achieved competitive transfer\nrates at an average of 50.5 bits/min and with a maximum of 84.7 bits/min.\n\n1 Introduction\n\nSome neurological diseases result in the so-called locked-in syndrome. People suffering\nfrom this syndrom lost control over their muscles, and therefore are unable to communicate.\nConsequently, their brain-signals should be used for communication. Besides the clinical\napplication, developing such a brain-computer interface (BCI) is in itself an exciting goal\nas indicated by a growing research interest in this \ufb01eld.\n\nSeveral EEG-based techniques have been proposed for realization of BCIs (see [6, 12], for\nan overview). There are at least four distinguishable basic approaches, each with its own\nadvantages and shortcomings:\n\n1. In the \ufb01rst approach, participants are trained to control their EEG frequency pat-\ntern for binary decisions. Whether speci\ufb01c frequencies (the \nrhythms)\nin the power range are heightened or not results in upward or downward cursor\nmovements. A further version extended this basic approach for 2D-movements.\nTransfer rates of 20-25 bits/min were reported [12].\n\nand\n\n2. Imaginations of movements, resulting in the \u201cBereitschaftspotential\u201d over sensori-\nmotor cortex areas, are used to transmit information in the device of Pfurtscheller\n\n\u0001\n\fFigure 1: Stimulusmatrix with one column highlighted.\n\net al. [8], which is in use by a tetraplegic patient. Blankertz et al. [2] applied\nsophisticated methods for data-analysis to this approach and reached fast transfer\nrates of 23 bits/min when classifying brain signals preceding overt muscle activity.\n\n3. The thought translation device by Birbaumer et al. [5, 1] is based on slow cortical\npotentials, i.e. large shifts in the EEG-signal. They trained people in a biofeedback\nscenario to control this component. It is rather slow (<6 bits/min) and requires\nintensively trained participants but is in practical use.\n\n4. Farwell & Donchin [4, 3, 10] developed a BCI-System by utilizing speci\ufb01c posi-\ntive de\ufb02ections (P300) in EEG-signals accompanying rare events (as discussed in\ndetail below). It is moderately fast (up to 12 bits/min) and needs no practice of the\nparticipant, but requires visual attention.\n\nFor BCIs, it is very desirable to have fast transfer rates. In our own studies, we therefore\ntried to accelerate the fourth approach by using state-of-the-art machine learning techniques\nand fusing data from different electrodes for data-analysis. For that purpose we utilized the\nbasic setup of Farwell & Donchin (referred to as F&D) [4] who used the well-studied\nP300-Component to create a BCI-system. They presented a 6\n6-matrix (see Fig. 1), \ufb01lled\nwith letters and digits, and highlighted all rows and columns sequentially in random or-\nder. People were instructed to focus on one symbol in the matrix, and mentally count its\nhighlightings. From EEG-research it is known, that counting a rare speci\ufb01c event (oddball-\nstimulus) in a series of background stimuli evokes a P300 for the oddball stimulus. Hence,\n6-matrix should result in a P300, a character-\nhighlighting the attended symbol in the 6\nistic positive de\ufb02ection with a latency of around 300ms in the EEG-signal. It is therefore\npossible to infer the selected symbol by detecting the P300 in EEG-signals. Under suitable\ncircumstances, most brains expose a P300. Thus, no training of the participants is nec-\nessary. For identi\ufb01cation of the right column and row associated with a P300, Farwell &\nDonchin used the model-based techniques Area and Peak picking (both described in section\n2) to detect the P300. In addition, as a data-driven approach, they used Stepwise Discrimi-\nnant Analysis (SWDA). Using SWDA in a later study [3] resulted in transfer rates between\n4.8 and 7.8 symbols per minute at an accuracy of 80% with a temporal distance of 125ms\nbetween two highlightings.\n\nIn our work reported here we could improve several aspects of the F&D-approach by utiliz-\ning very recent machine learning techniques and a larger number of EEG-electrodes. First\nof all, we could increase the transfer rate by using Support Vector Machines (SVM) [11] for\nclassi\ufb01cation. Inspired by a recent approach to learning of discriminative densities [7] we\nutilized the values of the SVM classi\ufb01cation function as a measure of con\ufb01dence which we\naccumulate over certain classi\ufb01cations in order to speed up the transfer rate. In addition,\nwe enhanced classi\ufb01cation rates by augmenting the data-space. While Farwell & Donchin\nemployed only data from a single electrode for classi\ufb01cation, we used the data from 10\nelectrodes simultaneously.\n\n\n\n\f2 Methods\n\nIn the following we describe the techniques used for acquisition, preprocessing and analysis\nof the EEG-data.\n\nData acquisition. All results of this paper stem from of\ufb02ine analyses of data acquired\nduring EEG-experiments. The experimental setup was the following: participants were\nseated in front of a computer screen presenting the matrix (see Fig. 1) and user instruc-\ntions. EEG-data were recorded with 10 Ag/AgCl electrodes at positions of the extended\ninternational 10-20 system (Fz, Cz, Pz, C3, C4, P3, P4, Oz, OL, OR 1) sampled at 200Hz\nand low-pass \ufb01ltered at 30Hz. The participants had to perform a certain number of trials.\nFor the duration of a trial, they were instructed\n\nto focus their attention on a target symbol speci\ufb01ed by the program,\nto mentally count the highlightings of the target symbol, and\nto avoid any body movement (especially eye moves and blinks).\n\nEach trial is subdivided into a certain number of subtrials. During each subtrial, 12 stimuli\nare presented, i.e.\nthe 6 rows and the 6 columns are highlighted in random order. For\ndifferent BCI-setups, the time between stimulus onsets, the interstimulus interval (ISI),\nwas either 150, 300 or 500ms, while a highlighting always lasts 150ms. To each stimulus\ncorrespondes an epoch, a time frame of 600ms after stimulus onset 2During this interval a\nP300 should be evoked if the stimulus contains the target symbol.\n\nThere is no pause between subtrials, but between trials. During the pause, the participants\nhad time to focus on the next target symbol, before they initiated the next trial. The target\nsymbol was chosen randomly from the available set of symbols and was presented by the\nprogram in order to create a data set of labelled EEG-signals for the subsequent of\ufb02ine\nanalysis.\n\nData preprocessing. To compensate for slow drifts of the DC potential, in a \ufb01rst step the\nlinear trend of the raw data in each electrode over the duration of a trial was eliminated. In\na second step, the data was normalized to zero mean and unit standard deviation. This was\nseparately done for each electrode taking the data of all trials into account.\n\nClassi\ufb01cation of Epochs. Test- and trainingsets were created by choosing the data ac-\ncording to one symbol as testset, and the data of the other symbols as trainingset in a\ncrossvalidation scheme.\n\nThe task of classifying a subtrial for the identi\ufb01cation of a target symbol has to be distin-\nguished from the classi\ufb01cation of a single epoch for detection of a signal, correlated with\noddball-stimuli, which we brie\ufb02y refer to as a \u201cP300 component\u201d in a simpli\ufb01ed manner\nin the following. In case of using a subtrial to select a symbol, two P300 components have\nto be detected within epochs: one corresponding to a row-, another to a column-stimulus.\nThe detection algorithm works on the data of an epoch and has to compute a score which\nre\ufb02ects the presence of a P300 within that epoch. Therefore, 12 epochs have to be evaluated\nfor the selection of one target symbol.\n\nFor the P300-detection, we utilized two model-based methods which had been proposed by\nF&D, and one completely data-driven method based on Support Vector Machines (SVMs)\n[11]. For training of the classi\ufb01ers, we built up a sets of epochs containing an equal number\nof positive and negative examples, i.e. epochs with and without a P300 component.\n\n1OL denotes the position halfway between O1 and T5, and OR between O2 and T6 respectively.\n2With an ISI shorter than 450ms, there is a time overlap of consecutive epochs.\n\n\n\n\n\ftime course\n\ntrial\n\nmodel\u2212based methods\n\nsubtrial 1\n\nsubtrial 2\n\nsubtrial 3\n\nstimulus\nonsets\n\nepoch of 600ms\n\nFigure 2: Trials, subtrials and epochs in the course of time (left). Model-based methods for\nanalysis. Area calculates surface in the P300-window, Peak picking calculates differences\nbetween peaks.\n\nThe \ufb01rst model-based method uses as its score as shown in Fig. 2 the area in the P300-\n\nwindow (\u201cArea method\u201d, \u0002\u0001 ), while the second model-based method uses the difference\npicking method\u201d,\u0002\u0003 ). Hyperparameters of the model-based methods were the boundaries\n\nbetween the lowest point before, and the highest point within the P300-window (\u201cPeak\n\nof the P300-window. They were selected regarding the average of epochs containing the\nP300 by taking the boundaries of the largest area.\n\nFor the completely data-driven approach, SVMs were optimized to distinguish between the\ntwo classes (w/o P300) implied by the training set. As compared with many traditional\nclassi\ufb01ers, such as the SWDA method used by F&D, SVMs can realize Bayes-consistent\nclassi\ufb01ers under very general conditions without requiring any speci\ufb01c assumptions about\nthe underlying data distributions and decision boundaries. Thereby convergence to the\nBayes optimum can be achieved by a suitable choice of hyperparameters.\n\nWhen using SVMs, it is not clear what measure to take as the score of an epoch. The\nproblem is that the SVM has \ufb01rst of all been designed to assign binary class labels to its\ninput without any measure of con\ufb01dence on the resulting decision.\n\nHowever, a recent approach to learning of discriminative densities [7] suggests an interpre-\ntation of the usual discrimination function for SVMs with positive kernels in terms of scaled\ndensity differences. This \ufb01nding provides us with a well-motivated score of an epoch: with\n\npositive/negative for epochs with/without target stimulus the SVM-score is computed as\n\n\u0004 as the data vector of an epoch and \u0005\u0007\u0006\t\b\u000b\n\r\f\u000f\u000e\u0010\f\u0012\u0011 as the corresponding class label which is\nwhere&\nexample. The mixing weights #\nweight/\n\n\u0004\u0002\u001e\n\u001c-,\n! were estimated by quadratic optimization for an SVM\nfor the soft-margin penalties by 0 -fold crossvalidation) evaluated at the 1 -th data\n\nin our case is a Gaussian Kernel function with bandwidth.\n\nobjective with linear soft-margin penalties where we used the SMO-algorithm [9].\n\n\u0014\u0013\u0016\u0015\u0018\u0017\u001a\u0019\u001a\u001b\u001d\u001c\n\n\u0019 \u001f\"!\n\n!$#%!'&\n\n\u001e)(+*\n\n(1)\n\n(selected as the\n\nCombination of subtrials. Because EEG-data possess a very poor signal-to-noise ratio\n(SNR), identi\ufb01cation of the target symbol from a single subtrial is usually not reliable\nenough to achieve a reasonable classi\ufb01cation rate. Therefore, several subtrials have to be\ncombined for classi\ufb01cation, slowing down the transfer rate. Thus, an important goal is to\ndecrease the amount of subtrials which have to be combined for a satisfactory classi\ufb01cation\nrate.\n\n\u0005\n\u001c\n\u0004\n\u000e\n\u0004\n!\n\u000e\n\u0004\n!\n\u001e\n\fAn important constraint for the development of the speci\ufb01c of\ufb02ine-analysis programs was\nto realize a testing scheme which should be as close as possible to a corresponding online\nevaluation. Therefore, we tested a method for certain\n-combinations of subtrials in the\nfollowing way: different series of\nsuccessive subtrials were taken out of a test set and the\ncorresponding single classi\ufb01cations were combined as explained below. Thereby, the test\nseries contained only subtrials belonging to identical symbols and these were combined in\ntheir original temporal order3.\nIn contrast, Farwell & Donchin randomly chose samples from a test set, built from subtrials\ntaken from different trials and belonging to different symbols. With this procedure, they\nbroke up the time course of the recorded data and did not distinguish between different\nsymbols, i.e. different positions in the matrix on the screen.\n\nBased on the data of\n\nthe target symbol, i.e. to classify a trial. Therefore, in a \ufb01rst step, the single scores 4\nof the epoch\u0004\n\u0013\u0015\u0014\b\u0016\u0018\u0017\u0019\u0013\u001b\u001a\n\nsubtrials, one has to choose a row and a column in order to identify\n\u0001\u0003\u0002\u0005\u0004\n-th subtrial\n\u001e . Then, the target row was chosen\n\n\u0019\f\u000b\u000e\n\n\u0001\u0007\u0002\b\u0004\n\n\u0001\u0003\u0002\b\u0004\n\ncorresponding to the stimulus associated to the1 -th row of the\n! with1\n\n\u0001\u0003\u0002\u0005\u0004\nwere summed up to the total score\nas\n. Equivalent steps were performed to choose the target\ncolumn. Based on these decisions the target symbol was \ufb01nally selected in accordance to\nthe presented matrix.\n\n\f\u000f\u000e\u001d\u001c\u001e\u001c\u001d\u001c\u0014\u000e \u001f\n\n\u0006\u0010\u000f\u0012\u0011\n\n\u0001\u0007\u0002\b\u0004\n\n3 Experimental Results\n\nBefore going into details, we outline our investigations about improving the usability of the\nF&D-BCI. First, the different methods were compared to classify the data of the Pz elec-\ntrode, which was originally used by Farwell & Donchin. Second, further single electrodes\nwere taken as input source. This revealed information about interesting scalp positions to\nrecord a P300 and on the other hand indicated which channels may contain a useful signal.\nThird, the SVM classi\ufb01cation rate with respect to epochs was improved by increasing the\ndata-space. Therefore, the input vector for the classi\ufb01er was extended by combining data\nfrom the same epoch but from different electrodes. These tests indicated that the best clas-\nsi\ufb01cation rates could be achieved using as detection method an SVM with all ten electrodes\nas input sources.\n\nSince the results of the \ufb01rst three steps were established based on the data of one initial\nexperiment with only one participant, we evaluated the generality of these techniques by\ntesting different subjects and BCI parameters. Finally, the BCI performance in terms of\nattainable communication rates is estimated from these analyses.\n\nMethod comparison using the Pz electrode as input source. All four methods were\napplied to the data of one initial experiment with an ISI of 500ms and 3 subtrials per trial.\nFigure 3 presents the classi\ufb01cation rates of up to 10 subtrials.\n\nThe SVM method achieved best performance, its epoch classi\ufb01cation rate was 76.3%\n(SD=1.0) in a 10-fold crossvalidation with about 380 subtrials samples in the training sets,\nand about 40 in the test sets. Of each subtrial in the training set, 4 epochs (2 with, 2 without\na P300) were taken as training samples, whereas all 12 epochs of the subtrials of the test\nset were classi\ufb01ed. For each training set, hyperparameters were selected by another 3-fold\ncrossvalidation on this set.\n\n3For a higher number of subtrial combinations, subtrials from different trials had to be combined.\nHowever, real-world-application of this BCI don\u2019t require such combinations with respect to the\n\ufb01nally achieved transfer rates reported in section 3.\n4The method index is omitted in the following.\n\n\n\n\n\u001c\n\u0004\n!\n\u0006\n\u001e\n!\n\u0006\n\t\n\n!\n\n\u001c\n\u0004\n!\n\u0006\n!\n\n\u0019\n\fFigure 3: (left) Method comparison on the Pz electrode: The three techniques were applied\nto the data of the initial experiment. (right) Classi\ufb01cation rates for different number of\nelectrodes.\n\nPeak picking\n\nSVM\n\n)\n\n%\n\nt\n\n(\ne\na\nr\nn\no\n\ni\nt\n\na\nc\ni\nf\ni\ns\ns\na\nc\n\nl\n\n100\n\n90\n\n80\n\n70\n\n60\n\n50\n\n40\n\n30\n\n20\n\n10\n\n0\n\n)\n\n%\n\nt\n\n(\ne\na\nr\nn\no\ni\nt\na\nc\ni\nf\ni\ns\ns\na\nc\n\nl\n\n100\n\n90\n\n80\n\n70\n\n60\n\n50\n\n40\n\n30\n\n20\n\n10\n\n0\n\n6\n\n12 18 24 30 36 42 48 54 60 66 72 78 84 90\n\n6\n\n12 18 24 30 36 42 48 54 60 66 72 78 84 90\n\ntime (s)\n\nP3\nFz\n\nP4\nCz\n\nOL\nPz\n\nOR\nC3\n\nOZ\nC4\n\ntime (s)\n\nFigure 4: Electrode comparison on the data of the initial experiment.\n\nDifferent electrodes as input source. The method comparison tests were repeated for\neach electrode. The results of the Peak picking and SVM method are shown in Figure 3.\n\nThe SVM is able to extract useful information from all ten electrodes, whereas the Peak\npicking performance varies for different scalp positions. Especially, the electrodes over the\nvisual cortex areas OZ, OR and OL are useless for the model-based techniques, as the same\ncharacteristics are revealed by tests with the Area method.\n\nHigher-dimensional data-space. While Farwell & Donchin used only one electrode for\ndata-analysis, we extended the data-space by using larger numbers of electrodes. We calcu-\nlated classi\ufb01cation rates for Pz alone, three, seven, and ten electrodes. A signal correlated\nwith oddball-stimuli was classi\ufb01ed at rates of 76.8%, 76.8%, 90.9%, and 94.5%, respec-\ntively for the different data-spaces of 120, 360, 840, and 1200 dimensions. These rates were\ncalculated with 850 positive and 850 negative epoch samples and a 3-fold crossvalidation.\nThis classi\ufb01ed signal might be more than solely the traditional P300 component. Apply-\ning data-space augmentation for classi\ufb01cation to infer symbols in the matrix results in the\nclassi\ufb01cation rates depicted in Figure 3 (right) for an ISI of 500ms. Using ten electrodes\nsimultaneously, combined in one data vector, outperforms lower-dimensional data-spaces.\n\n\fFigure 5: Mean-classi\ufb01cation rates (left) and transfer rates (right) for different ISIs. Error\nbars range from best to worst results. Note that a subtrial takes a speci\ufb01c amount of time.\nTherefore, the time dependend transfer rates are decreasing with the number of subtrials.\n\nReducing the ISI and using more participants. The improved classi\ufb01cation rates en-\ncouraged further experiments. To accelerate the system, we reduced the ISI to 300ms and\n150ms. Additionally, to generalize the results, we recruited four participants. Means, best\nand worst classi\ufb01cation rates are presented in Figure 5, as well as average and best transfer\nrates. The latter were calculated according to\n\n\u0002\u0001\n\n\u0019\u0004\u0003\u0006\u0005\n\n\u001f\b\u0007\n\n\u001c\n\t\f\u000b\n\n\u0016\u000e\r\u0010\u000f\n\n(\u0012\u0011\n\n\t\f\u000b\n\n\u0016\u0013\n\n\t\f\u000b\n\n\u0016\u0014\n\nis the number of choices (36 here), \u0011\n\ntime required for classi\ufb01cation.\n\nwhere \u000f\n\nthe probability for classi\ufb01cation, and \u0003\n\nthe\n\nUsing an ISI of 300ms results in slower transfer rates than using an ISI of 150ms. The\nlatter ISI results on the average in classifying a symbol after 5.4s with an accuracy of 80%\n(disregarding delays between trials). The poorest performer needs 9s to reach this criterion,\nthe best performer achieves an accuracy of 95.2% already after 3.6s. The transfer rates, with\na maximum of 84.7 bits/min and an average of 50.5 bits/min outperform the EEG-based\nBCI-systems we know.\n\n4 Conclusion\n\nWith an application of the data-driven SVM-method to classi\ufb01cation of single-channel\nEEG-signals, we could improve transfer rates as compared with model-based techniques.\nFurthermore, by increasing the number of EEG-channels, even higher classi\ufb01cation and\ntransfer rates could be achieved. Accumulating the value of the classi\ufb01cation function as\nmeasure of con\ufb01dence proved to be practical to handle series of classi\ufb01cations in order to\nidentify a symbol. This resulted in high transfer rates with a maximum of 84.7 bits/min.\n\n5 Acknowledgements\n\nWe thank Thorsten Twellmann for supplying the SVM-algorithms and the Department of\nCognitive Psychology at the University of Bielefeld for providing the experimental envi-\nronment. This work was supported by Grant Ne 366/4-1 and the project SFB 360 from the\nGerman Research Council (Deutsche Forschungsgemeinschaft).\n\n\u0011\n(\n\u001c\n\f\n\n\u0011\n\u001e\n\u001c\n\f\n\n\u0011\n\u0005\n\u000f\n\n\f\n\u001e\n\u001e\n\u000e\n\fReferences\n\n[1] N. Birbaumer, N. Ghanayim, T. Hinterberger, I. Iversen, B. Kotchoubey, A. K\u00fcbler,\nJ. Perelmouter, E. Taub, and H. Flor. A spelling device for the paralysed. Nature,\n398:297\u2013298, 1999.\n\n[2] B. Blankertz, G. Curio, and K.-R. M\u00fcller. Classifying single trial eeg: Towards brain\ncomputer interfacing.\nIn T. G. Dietterich, S. Becker, and Z. Ghahramani, editors,\nAdvances in Neural Information Processing Systems 14, Cambridge, MA, 2002. MIT\nPress.\n\n[3] E. Donchin, K.M. Spencer, and R. Wijeshinghe. The mental prosthesis: Assessing the\nspeed of a p300-based brain-computer interface. IEEE Transactions on Rehabilitation\nEngineering, 8(2):174\u2013179, 2000.\n\n[4] L.A. Farwell and E. Donchin. Talking off the top of your head: toward a mental pros-\nthesis utilizing event-related brain potentials. Electroencephalography and clinical\nNeurophysiology, 70(S2):510\u2013523, 1988.\n\n[5] A. K\u00fcbler, B. Kotchoubey, T. Hinterberger, N. Ghanayim, J. Perelmouter, M. Schauer,\nC. Fritsch, E. Taub, and N. Birbaumer. The thought translation device: a neurophys-\niological approach to commincation in total motor paralysis. Experimental Brain\nResearch, 124:223\u2013232, 1999.\n\n[6] A. K\u00fcbler, B. Kotchoubey, J. Kaiser, J.R. Wolpaw, and N. Birbaumer. Brain-computer\ncommunication: Unlocking the locked in. Psychological Bulletin, 127(3):358\u2013375,\n2001.\n\n[7] P. Meinicke, T. Twellmann, and H. Ritter. Maximum contrast classi\ufb01ers. In Proc. of\n\nthe Int. Conf. on Arti\ufb01cial Neural Networks, Berlin, 2002. Springer. in press.\n\n[8] G. Pfurtscheller, C. Neuper, C. Guger, B. Obermaier, M. Pregenzer, H. Ramoser, and\nIEEE\n\nA. Schl\u00f6gl. Current trends in graz brain-computer interface (bci) research.\nTransactions On Rehabilitation Engineering, pages 216\u2013219, 2000.\n\n[9] J. Platt. Fast training of support vector machines using sequential minimal optimiza-\ntion. In B. Sch\u00f6lkopf, C. J. C. Burges, and A. J. Smola, editors, Advances in Kernel\nMethods \u2014 Support Vector Learning, pages 185\u2013208, Cambridge, MA, 1999. MIT\nPress.\n\n[10] J.B. Polikoff, H.T. Bunnell, and W.J. Borkowski. Toward a p300-based computer\ninterface. RESNA \u201995 Annual Conference and RESNAPRESS and Arlington Va., pages\n178\u2013180, 1995.\n\n[11] V. N. Vapnik. The Nature of Statistical Learning Theory. Springer, New York, 1995.\n[12] J.R. Wolpaw, N. Birbaumer, D.J. McFarland, G. Pfurtscheller, and T.M. Vaughan.\nBrain-computer interfaces for communication and control. Clinical Neurophysiology,\n113:767\u2013791, 2002.\n\n\f", "award": [], "sourceid": 2256, "authors": [{"given_name": "Peter", "family_name": "Meinicke", "institution": null}, {"given_name": "Matthias", "family_name": "Kaper", "institution": null}, {"given_name": "Florian", "family_name": "Hoppe", "institution": null}, {"given_name": "Manfred", "family_name": "Heumann", "institution": null}, {"given_name": "Helge", "family_name": "Ritter", "institution": null}]}