CC BY
108
УДК 004.032.26 ББК Е70*739.1*718
E. ZAVGORODNYAYA
A COMPARISON OF SPATIOTEMPORAL FEATURE EXTRACTION ALGORITHMS FOR ERROR-RELATED POTENTIAL DETECTION IN BRAIN-COMPUTER INTERFACES
Key words: Error-related Potentials, Brain-Computer Interfaces, electroencephalography, spatiotemporal feature extraction algorithms, electroencephalographical data. The detection of Error-related Potentials (ErrPs) from the human brain can be a valuable addition to Brain-Computer Interfaces (BCIs) that are based on electroencephalography (EEG) or magnetoencephalography (MEG). With this capability, the performance of BCIs can namely be increased significantly [23,45]. However, detecting ErrPs on single-trial level is still challenging [23] and it is not clear which methods for data preprocessing and feature extraction give the best performance. This study sets out to fill this gap in our knownledge by comparing different state-of-the-art methods for data preprocessing and single-trial detection of ErrPs: feature extraction using prior spatiotemporal knowledge [23], automatic detection of spatiotemporal patterns [8], cross-correlation based features [44], PCA based features [16] and CSP based features [59]. Those methods were compared on two main criteria: the accuracy as a function of training trials and the robustness to intra- and inter-subject variability. Based on these two criteria, we conclude that a PCA based method is superior to the methods that have been used most commonly for ErrP detection so far.
ЕВ. ЗАВГОРОДНЯЯ СРАВНЕНИЕ АЛГОРИТМОВ ВЫДЕЛЕНИЯ ПРОСТРАНСТВЕННО-ВРЕМЕННЫХ ПОТЕНЦИАЛОВ МОЗГА, СВЯЗАННЫХ С ОШИБОЧНЫМ ДЕЙСТВИЕМ НЕЙРОКОМПЬЮТЕРНОГО ИНТЕРФЕЙСА ИЛИ ПОЛЬЗОВАТЕЛЯ ЭТОГО ИНТЕРФЕЙСА
Ключевые слова: пространственно-временные потенциалы мозга, электроэнцефалография, алгоритмы, нейрокомпьютерный интерфейс, пользовательский интерфейс, электроэнцефалографические данные.
Выявление потенциалов мозга, возникающих при совершении ошибочных действий ней-рокомпьютерных интерфейсов либо пользователями этих интерфейсов (ErrPs), может являться ценным дополнением для улучшения работы нейронных компьютеров, основанных на электроэнцефалографии (ЭЭГ) или магнитоэнцефалографии (MEG). Это позволит значительно повысить эффективность нейрокомпьютерных интерфейсов [23, 45]. Однако обнаружение ErrPs для разных пользователей остается затруднительным [23], к тому же не ясно, какие методы обработки данных и выделения ErrPs достигают наилучшего результата. Данное исследование - попытка решить эту проблему путём сравнения самых современных методов предварительной обработки данных и выявления ErrPs для разных пользователей нейрокомпьютерных интерфейсов. Первый метод основан на том, что пространство и время ErrPs строго определены [23]. За основу второго метода взято построение алгоритма автоматического обнаружения пространственно-временных характеристик ErrP [8]. Третий метод использует кросс-корреляции ErrPs [44]. Четвертый и пятый методы связаны с анализом главных компонентов (PCA) [16] и обобщением пространственных моделей (CSP) [59] ErrPs. Все эти методы сравниваются по двум главным критериям: точность достижения желаемых результатов и совместимость для разных пользователей. На основе этих двух критериев можно заключить, что метод, основанный на анализе главных компонентов (РСА), превосходит другие методы, чаще всего используемые для обнаружения ErrPs. Статья содержит общую информацию о нейрокомпьютерных интерфейсах (BCI), а именно: обнаружение сигнала ErrPs, предварительная обработка данных, выделение пространственно-временных характеристик ErrP и их классификация.
Introduction. A brain-computer interface (BCI) is a system that provides communication between a human and a computer by translating brain activity into con trol commands. This can be very important for severely disabled people. Therefore, many
research groups are working on applications of BCI on this domain [42, 46, 51, 63]. Moreover, healthy people can also be expected to start using BCI systems to serve as an additional man-machine interaction channel [17].
Different methods can be used to measure brain activity, including electroencephalography (EEG), magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI). Our research will focus on EEG recordings, which are recordings of small potential differences over the scalp. Hans Berger was the first to discover that electrical signals can be recorded from the human brain and developed electroen-cephalography (EEG), making him the first person who could record human brain activity in 1924. He inserted silver wires under the scalps of his patients, which were later replaced by silver foils attached to the patients' head by rubber bandages, and connected the electrodes to a Lippmann capillary electrometer. Nevertheless, this method did not lead to success immediately: only after the capillary electrometer was replaced by Siemens double-coil recording galvanometer, which displayed electric voltages as small as one ten thousandth of a volt, results were achieved.
Research on BCIs began much later, when personal computers started to become reality, in the 1970s at the university of California, Los Angeles. Early studies showed that it is possible to see changes in monkeys' neural activities with delivery of a food pellet [25]. Similar work in the 1970s established that monkeys could quickly learn to voluntarily control the firing rates of individual and multiple neurons in the primary motor cortex if they were rewarded for generating appropriate patterns of neural activity [61].
Following years of animal experimentation, the first neuroprosthetic devices implanted in humans appeared in the mid-1990s [43]. This study showed a detection of movement-related event-related potentials (ERP's) with people who cannot perform physical movements. Nowadays it is possible to build practical BCIs with EEG [13, 50, 63, 68]. Subjects are, for example, able to control prostheses, wheelchairs [13, 50, 63, 68] or even a miniature robot in an indoor environment with several rooms, corridors and doorways [48], sometimes after only a couple of training days.
EEG has several advantages compared to MEG or fMRI. It is easy to use, relatively cheap, does not require a big machine, has a high temporal resolution and does not carry any risks for subjects. Unfortunately, EEG has downsides as well: it has a low spatial resolution and the electrodes generally require some preparation. Additionally, EEG based BCIs share two problems with most other BCIs: artifacts and poor signal to noise ratio.
Due to these disadvantages, the single-trial classification of EEG data is difficult. Even well trained subjects can generally not reach very high accuracies [24]. This is a major challenge for EEG-based BCI. In practice, this means a trade-off has to be made between speed and accuracy: we either have to accept a lower accuracy while maintaining a high speed, or accept a slower communication rate to increase the accuracy. An excellent example of this trade-off and ways to deal with it is the P300 speller. The amount of transferred information per time unit can be used to choose the optimal setting in this trade-off [62].
One way to increase this amount of transferred information, is to automatically detect errors in the classification of brain signals by the BCI. It has long been known that when a person notices an error, error related activity can be measured in the brain [60]. This activity is thought to originate from the Anterior Cingulate Cortex [8, 12, 62] and in case of an EEG recording, it is known as the error-related potential (ErrP, also
known as ERN, error related negativity). It has been shown that the ErrP is a quite strong and stable response. Therefore, it is possible to do single-trial ErrP classification [7, 18, 24, 53, 62] and a BCI can potentially use this signal to detect its own errors [15, 22, 66]. Furthermore, it has been shown that ErrP detection can significantly improve the performance of BCIs [8, 62].
Nevertheless, ErrP detection is still far from perfect. Due to the poor signal-to-noise ratio (SNR), the accuracy of the classification has much room for improvement. In this thesis we will discuss several important questions that have to be answered to improve ErrP detection. The first question is whether different filtering techniques such as CAR, sLAP and bLAP can help to increase classification performance. The second question is how many trials we need for training a classifier to achieve reasonable results. The third question is what is the best way to deal with inter- and intrasubject variability [2, 6, 9, 30, 31, 39, 64]. Finally, an important question is which methods for event-related potential (ERP) classification are best suited for ErrPs and how they compare. In this thesis, we will attempt to fill these gaps in our knowledge and find out how we can obtain the best results for ErrP detection by comparing a number of state-of-the-art techniques for ERP detection.
A BCI is an artificial intelligence system that can recognize a certain set of patterns in brain signals and translate them into control commands. It also means that we explicitly do not want to use signals that can be measured from axons, from muscle movement. A BCI can be seen as a system with five consecutive stages: signal acquisition, preprocessing, feature extraction, classification and feedback via a control interface [36]. First of all, the signal acquisition stage captures the brain signal. Then the preprocessing stage removes artifacts from the data and prepares the signal for further analysis. Following this, the feature extraction stage selects features that are relevant for discrimination and removes noise. Afterward, the extracted features are classified by a classifier. The classification performance depends on the quality of the features and in some cases feature selection and classification can be handled by one algorithm. Finally, the control interface stage translates the classified signal into meaningful commands for any connected device, such as a wheelchair or a virtual keyboard. We will discuss the different stages one by one in this chapter.
Electroencephalography based brain computer interfaces
1. Electroencephalography (EEG). The electrical activity is recorded with EEG by means of electrodes placed on the scalp [34]. EEG measures the potential differences between electrodes that correspond to a current flowing through the skin [4]. These currents are caused by neural activity. Only a tiny fraction of the electric currents through the neurons can be measured with EEG, because neural activity causes only small currents and the skull is not a good conductor. Therefore, we can only get a good signal when a large amount of neurons is firing synchronously. Moreover, the currents caused by the neurons should not be uniform in all directions, as the total current would sum up to zero. As a result, EEG measures mainly the activity of pyramidal neurons, because they form a large fraction of all neurons and they are all aligned in the same direction.
A complete EEG set-up consists of electrodes, amplifiers, an A/D converter, and a recording device. First of all, the electrodes record the signal from the brain activity, next the amplifiers enhance the signal, so an A/D converter can read it, then the recording device stores and displays the data. There are two types of electrodes: 'dry' electrodes that do not use gels and 'wet' electrodes. 'Wet' electrodes are used with an EEG
gel. The EEG gel creates a conductive path between the skin and each electrode that reduces the impedance and thus improves the signal. Impedance depends on an interface layer, electrode surface area and temperature [1]. To be able to find the right position of each electrode on the head, two reference points are used: the nasion point that is located at the same level as the eyes and the inion that lies in the bony lump at the base of the skull. Traditionally, EEG has been used to measure oscillatory activity in the brain. There are five common frequency bands: delta, theta, alpha, beta, and gamma. First of all, delta waves lie below 4 Hz. They are usually observed in adults in a deep sleep and unusually in babies. It is easy to confuse delta waves with artifact signals, because of the similarity in signals between delta waves and waves that are caused by the large muscles of the neck or jaws [40]. In contrast with delta waves, theta waves lie within the 4 to 7 Hz range. A large amount of theta frequencies can be seen in young children, older children, and adults in drowsy, meditative or sleep states [40]. Furthermore, theta band has been associated with cognitive processes such as mental calculation [21], maze task demands [11] and conscious awareness [38]. By comparison with theta waves, alpha rhythms lie between 8 and 12 Hz range. Alpha activity increases with the eyes closed and the body relaxed and decreases when the eyes opened. These rhythms can be related with the working memory load [37]. In addition, sources of the alpha rhythms are found in parieto-occipital sulcus and sensorimotor areas. Mu rhythms may be found in the same range as alpha rhythms, although there are important physiological difference between both.
Subject
Figure 1. Non-invasive brain-computer interface (BCI). The EEG signal is recorded by means of electrodes placed on the scalp. Some features are extracted from the EEG signal and send to a classifier, whose response is translated into actions (The figure is reproduced from the [23])
Figure 2. An overview of the electrode placement on an EEG recording cap
In contrast to alpha rhythms, mu rhythms are strongly connected to motor activities and sometimes correlate with beta rhythms [55, 57]. Beta rhythms lie in 12 to 30 Hz range. Sources of the beta rhythms are found in (sensori-)motor cortex and associated with motor activities. Besides, these rhythms are topographically modulated by movements and somatosensori activation. Finally, gamma band oscillations occur from 30 to 100 Hz. The presence of gamma waves in the brain activity of a healthy adult is related to certain motor functions or perceptions [41]. However, this classification of EEG waves is only based on ongoing signals in certain frequency bands. A diffrent way to look at EEG is to investigate the transient signals after a certain stimulus, perhaps even trying to match a signal with a brain function. For example, event related potentials (ERPs) are signals generated by a population of neurons in response to a perceptual, cognitive or motor event. The different types of ERPs will be described in the following section, as the Error related Potential (ErrP) is one of them.
2. EEG Signal Types of Interest for BCI. The main task of BCIs is to make mental control possible. As described before, this is possible by monitoring cerebral activity with EEG and separate the interesting activity from the abundance of noise. Over the past thirty years, different types of activity have already been distinguished. We can use this knowledge for BCI systems. Visual evoked potentials, the P300, slow cortical potentials, and sensorimotor rhythms are some of the most important examples of these signals. We will discuss them one by one, concluding with the ERP of our interest, the Error Related Potential (ErrP).
Visual Evoked Potentials (VEPs) and P300. VEPs are brain activities recorded from the visual cortex of a human or other animal brain following presentation of a visual stimulus [58]. They can be evoked by using flash stimuli or using graphic patterns such as a checkerboard lattice, a gate, or a random-dot map [52]. In addition, this brain activity is relatively easy to detect, because the amplitude of VEPs increase as the stimulus is moved closer to the central visual field [67].
Figure 3. Example of a checker-board stimulus
VEPs depend on the area of the on-screen stimulus. For example, if only a half of the checker-board is displayed on the screen and the person is looking at the
center of the screen, then VEPs will only be induced in half of the visual cortex. Thus, VEPs are very tightly coupled with the visual input. Therefore, this gives the main advantage of VEPs: they do not require long training sessions. The only one requirement for a subject is to stare on digits or letters displayed on a screen or concentrate at one point of a screen.
VEPs can be classified as transient VEPs (TVEPs) and as steady-state VEPs (SSVEPs). TVEPs occur in the frequency band below 6 Hz, while SSVEPs lie in much higher frequency range [27, 58]. TVEPs can be elicited by any changes in the visual field such as flashing lights, a checkerboard lattice that changes the squares from black to white and from white to black abruptly [52]. Compared to TVEPs, SSVEPs are elicited by the same visual stimulus, but at a frequency higher than 6 Hz. In addition, TVEPs are more susceptible to artifacts caused by blinks, eye movements [54] and electromyographic noise contamination than SSVEPs [29].
An example of a combination of VEPs and the P300 evoked potential is the P300 speller. The P300 evoked potential is an ERP elicited by infrequent, task-relevant stimuli (and has therefore a cognitive component, instead of only a visual component). It comprises a matrix of letters, numbers, or other symbols or commands [19, 49, 65]. First of all, when the EEG is monitored, the rows and columns of this matrix are randomly flashed. The subject's task is to choose one of the symbols and count how many times the row or column with the desired symbol was flashed. Next, the BCI tries to determine the target symbol. Due to the low SNR, the rows and the columns have to be flashed several times before the classifier is certain enough the user wants to make a certain choice. However, this repetition decreases the number of choices per minute [19].
Slow Cortical Potentials (SCPs). Slow cortical potentials are negative or positive polarizations of the electroencephalogram (EEG) or magnetic field changes in the magnetoencephalogram (MEG) that last from 300 ms to several seconds. Negative SCPs correlate with increased neuronal activity, while positive SCPs coincide with decreased activity in individual cells [5]. SCPs can be used for controlling a moving cursor or selecting targets presented on a screen by both healthy users and paralyzed patients [32]. However, SCPs have several disadvantages. First of all, they require long training sessions that lead to concentration problems of the subject. Secondly, SCPs depend on the psychological and physical state of a subject, which decreases performance due to intrasubject variability. Thirdly, they are very sensitive to artifacts.
Oscillatory Activity. Oscillatory activity can also be used for BCI purposes. Sen-sorimotor rhythms, for example, lie in the mu and beta range and originate in sensori-motor cortex. The causes and effects of these rhythms are not fully understood. However, phenomenologically, a strong amplitude of sensorimotor rhythms is produced by a person when the corresponding sensory-motor areas are idle, e.g. during states of immobility. Therefore, sensorimotor rhythms decrease in amplitude when corresponding sensory or motor areas are activated, e.g. during motor tasks and even during imaginary movements [33]. Sensorimotor rhythms can be used to control BCIs, because people can generate these modulations at will [10, 56]. BCIs based on sensorimotor rhythms can operate in synchronous or asynchronous mode (working with specific repeated trial times or initiated at will, respectively). Recent studies based on sensorimo-tor rhythms have shown that it is possible to predict human movements before they actually occur [3]. In addition, other oscillatory activity has also been used for BCIs, for example, the alpha rhythm in the covert attention paradigm [14].
Error Related Potentials (ErrPs). Finally, the signal that is exploited in this thesis: the Error Related Potential (ErrP). It is the ERP that is found after the brain notices an error, either consciously or unconsciously (however, the ErrP is less pronounced when the subject is not consciously aware of it) [23]. The ErrP has two main components. The first main component is the Error Related Negativity (ERN), which peaks around 100 ms after the error. This error can, for example, be the wrong decision in a multiple choice task, even without feedback [23]. A similar signal can also be observed after negative feedback and is known as the feedback ERN (fERN) [26]. After the negative peak, the second main component of the ErrP can be observed: the error-related positivity (Pe). This peak occurs around 300-400 ms. In addition, it has been proposed that the ERN and the Pe are not generated in the same brain region, but respectively in the anterior cingulate cortex and posterior cingulate cortex [18, 23].
The ErrP can be used in a BCI to improve the classification performance by automatically detecting errors [15, 23]. One of the advantages is that the ErrP is a strong response. Furthermore, in a trial-based BCI application, the ErrP will generally be timelocked to the feedback. For example, it has been shown that the performance of the P300 speller can be improved by detecting ErrPs [15].
3. Artifacts in BCIs. Artifacts are all signals that are not caused by brain activity. Moreover, they can, just as noise, reduce the performance of BCI-based systems significantly. There are two major categories of artifacts: physiological and non-physiological artifacts.
Physiological artifacts usually occur due to muscular activity. This is known as electromyography (EMG) in general, while ocular and heart activity in particular are known as electrooculography (EOG) and electrocardiography (ECG), respectively [20]. First of all, EMG artifacts are usually produced when patients are talking, chewing and swallowing and generate large disturbances in EEG signals. Secondly, EOG artifacts are produced when patients are blinking or moving their eyes. Blinking generally generates high-frequency patterns while eye movements generate low-frequency patterns. Thirdly, ECG introduces a rhythmic signal into brain activity, due to the heart beat. These artifacts can be removed from the signal, by recording EOG, EMG and ECG activity. However, this removal (for example by subtracting a prototype eyeblink from the signal) is never perfect and physiological artifacts are an important problem in EEG recordings. An extra complication is that these artifacts are sometimes coupled with the stimuli; for example, a person blinking every time a new stimulus is presented. Therefore, subjects are always asked to keep movement to a minimum and blink at moments that are not used for the experiment.
In case filtering turns out to be unreliable or too much effort, the easiest approach is to reject all trials in which eye movement or blinks are detected. This can be done automatically and is particularly simple when EOG data has been recorded with separate electrodes at the eyes. This is known as automatic artifact rejection1. When this process is not automated, all trials have to be examined by humans. This method is known as visual artifact rejection. However, this is a very slow and tiring method and is not recommended.
1 In the Fieldtrip implementation, automatic artifact rejection is done by calculating the power of the signal over time. Then, every time point is z-normalized (mean subtracted and divided by standard deviation) and, finally, averaging z-values over channels allows evidence for an artifact to accumulate. More information can be found at http://fieldtrip.fcdonders.nl/tutorial/automatic artifact rejection.
Non-physiological artifacts are either environmental noise or problems with the equipment. The prime example is 50 Hz line noise in recordings.
This type of artifacts should be predictable and can therefore be avoided or corrected, for example with a 50 Hz band-stop filter.
4. Data Preprocessing. The purpose of data processing is to transform the measured brain activity in such a way that the signal-to-noise ratio is maximized. The most common types of preprocessing are artifact rejection, spectral filtering and spatial filtering. Firstly, artifact detection finds confounding signals, such as eye and muscle artifacts, and removes them from the data Secondly, spectral filtering is used to remove noise signals, such as slow drifts and enhance the frequency bands of interest. Finally, spatial filtering combines signals from multiple electrodes to focus on activity at a particular location in the brain. In this subsection we consider different possible spatial filtering, such as Common Average Reference (CAR) and Laplace filters.
Common Average Reference (CAR). If XGMNxT with N being the number of channels and T the number of sampled time points is the matrix representation of one trial of EEG data, CAR is a spatial filtering technique for X that uses all elec-trades as the reference and is calculated according to the formula:
1 N
X CAR = X i - 1 £ X j (1)
N j=1
where, I and j is the electrode number and N is the number of electrodes.
CAR is effectively a high-pass spatial filter and can enhance focal activity from local sources and reduce widely distributed activity, including that from distant sources [47].
Laplacian. The Laplacian is a spatial filter and we will investigate two types: a small Laplacian (sLAP) and a big Lapla-cian (bLAP). A sLAP filter is obtained by rereferencing an electrode to the mean of its four neighboring electrodes. bLAP is described by rereferencing an electrode to the mean of its four nextnearest neighboring electrodes. They can be calculated as:
1 N
X1 = X1 —y Xj (2)
4 j= ' }
where I is the electrode number and j is the electrode number from a neighborhood of electrode I (the neighboring electrodes are illustrated in figure 5).
Being high-pass spatial filters, sLAP and bLAP enhance focal activity from local sources and reduce widely distributed activity, including that from distant sources [47]. They have a much higher spatial frequency than CAR, however.
5. Feature Extraction. Usually, the data from an EEG recording is too high-dimensional to be directly handled by a classifier. Therefore, a method to reduce the dimensionality and extract the useful features is required. If the extracted features are
CAR
Figure 4. Electrode locations used for a CAR. The figure is reproduced from [47]
Small Laplacian
Large Laplacian
Figure 5. Electrodes locations used in the small Laplacian and the big Laplacian. The figure is reproduced from [47]
carefully chosen, it is expected that the feature set will still contain the relevant information from the input data, but with a much smaller dimension. In this thesis we will consider some of the methods that can be used for feature extraction: Using prior spatiotemporal knowledge [22]; Automatic detection of spatiotemporal patterns [8]; Cross-correlation; Principal Component Analysis (PCA); Common Spatial Pattern (CSP).
6. Classification Algorithms. The goal of the classification step in a BCI system is to interpret the intention of the user on the basis of a feature vector that characterizes his or her brain activity. The most commonly used classifiers are Linear Discriminant Analysis (LDA), Support Vector Machine Vector (SVM) and Logistic Regression.
One of the most often used classification algorithms is the Support Vector Machine (SVM). First of all, it performs well on many problems and is a comparatively fast method. Secondly, it is robust to the curse-of-dimensionality and insensitive to overtraining. Thirdly, it gives very good results for BCI applications [28, 35]. For these reasons, we will also use the SVM for our experiments.
In addition, the regularization parameter of the SVM was automatically fixed using cross-validation.
Bibliography
1. U. Ali Bulent. Improvement of EEG signal acquisition: An electrical aspect for state of the art of front end. Computational Intelligence and Neuroscience, 2010, 2010.
2. B. Allison, T. Luth, D. Valbuena, A. Teymourian, I. Volosyak, and A. Graser. Bci demographics: How many (and what kinds of) people can use an ssvep bci? Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 18(2):107-116, 2010.
3. O. Bai, V. Rathi, P. Lin, D. Huang, H. Battapady, D.Y. Fei, L. Schneider, E. Houdayer, X. Chen, and M. Hallett. Prediction of human voluntary movement before it occurs. Clinical Neurophysiology, 122(2):364-372, 2011.
4. S. Baillet, J.C. Mosher, and R.M. Leahy. Electromagnetic brain mapping. Signal Processing Magazine, IEEE, 18(6):14-30, 2001.
5. N. Birbaumer, T. Elbert, AG Canavan, B. Rockstroh, et al. Slow potentials of the cerebral cortex and behavior. Physiological reviews, 70(1):1,1990.
6. B. Blankertz, G. Dornhege, M. Krauledat, K.R. Müller, and G. Curio. The non-invasive berlin brain-computer interface: Fast acquisition of effective performance in untrained subjects. NeuroImage, 37(2):539-550, 2007.
7. B. Blankertz, G. Dornhege, C. Schafer, R. Krepki, J. Kohlmorgen, K.R. Muller, V. Kunzmann, F. Losch, and G. Curio. Boosting bit rates and error detection for the classification of fast-paced motor commands based on single-trial EEG analysis. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):127-131, 2003.
8. B. Blankertz, S. Lemm, M. Treder, S. Haufe, and K.R. Müller. Single-trial analysis and classification of ERP components-a tutorial. NeuroImage, 56(2):814-825, 2011.
9. B. Blankertz, C. Sanelli, S. Halder, E.M. Hammer, A. Kübler, K.R. Müller, G. Curio, and T. Dickhaus. Predicting BCI performance to study BCI illiteracy. 2009.
10. B. Blankertz, C. Sannelli, S. Halder, E.M. Hammer, A. Kübler, K.R. Müller, G. Curio, and T. Dickhaus. Neurophysiological predictor of SMR-based BCI performance. NeuroImage, 51(4):1303-1309, 2010.
11. Jeremy B. Caplan, Joseph R. Madsen, Sridhar Raghavachari, and Michael J. Kahana. Distinct patterns of brain oscillations underlie two basic parameters of human maze learning. Journal of Neurophysiology, 86(1):368-380, 2001.
12. C.S. Carter, T.S. Braver, D.M. Barch, M.M. Botvinick, D. Noll, and J.D. Cohen. Anterior cingu-late cortex, error detection, and the online monitoring of performance. Science, 280(5364):747, 1998.
13. F. Cincotti, D. Mattia, F. Aloise, S. Bufalari, G. Schalk, G. Oriolo, A. Cherubini, M.G. Mar-ciani, and F. Babiloni. Non-invasive brain-computer interface system: towards its application as assis-tive technology. Brain research bulletin, 75(6):796-803, 2008.
14. T.M. Cover. Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition. Electronic Computers, IEEE Transactions on, (3):326-334, 1965.
15. B. Dal Seno, M. Matteucci, and L. Mainardi. Online detection of p300 and error potentials in a BCI speller. Computational intelligence and neuroscience, 2010:11, 2010.
16. Koel Das, Sergey Osechinskiy, and Zoran Nenadic. A classwise pca-based recognition of neural data for brain-computer interfaces. InEngineering in Medicine and Biology Society, 2007. EMBS 2007. 29th Annual International Conference of the IEEE, pages 6519-6522. IEEE, 2007.
17. G. Dornhege. Toward brain-computer interfacing. The MIT Press, 2007.
18. M. Falkenstein, J. Hoormann, S. Christ, and J. Hohnsbein. ERP components on reaction errors and their functional significance: a tutorial. Biological psychology, 51(2-3):87-107, 2000.
19. L.A. Farwell and E. Donchin. Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalography and clinical Neurophysiology, 70(6):510-523, 1988.
20. M. Fatourechi, A. Bashashati, R.K. Ward, and G.E. Birch. EMG and EOG artifacts in brain computer interface systems: A survey. Clinical Neurophysiology, 118(3):480-494, 2007.
21. Thalia Fernandez, Thalia Harmony, Mario Rodriguez, Jorge Bernal, Juan Silva, Alfonso Reyes, and Erzsebet Marosi. EEG activation patterns during the performance of tasks involving different components of mental calculation. Electroencephalography and clinical Neurophysiology, 94(3):175-182, 1995.
22. P.W. Ferrez and J. del R. Millan. Error-related EEG potentials generated during simulated brain-computer interaction. Biomedical Engineering, IEEE Transactions on, 55(3):923-929, 2008.
23. P.W. Ferrez and J.R. Millan. Error-related EEG potentials in braincomputer interfaces. Toward brain-computer interfacing, page 291, 2007.
24. P. Ferrez and J.R. Millan. You are wrong automatic detection of interaction errors from brain waves. 2005.
25. E.E. Fetz. Operant conditioning of cortical unit activity. Science, 163(3870):955, 1969.
26. Michael J. Frank, Brion S. Woroch, and Tim Curran. Error-related negativity predicts reinforcement learning and conflict biases. Neuron, 47(4):495-501, 2005.
27. X. Gao, D. Xu, M. Cheng, and S. Gao. A BCI-based environmental controller for the motion-disabled. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):137-140, 2003.
28. Deon Garrett, David A. Peterson, Charles W. Anderson, and Michael H. Thaut. Comparison of linear, nonlinear, and feature selection methods for EEG signal classification. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):141-144, 2003.
29. M. Gray, A.H. Kemp, R.B. Silberstein, and P.J. Nathan. Cortical neurophysiology of anticipatory anxiety: an investigation utilizing steady state probe topography (SSPT). NeuroImage, 20(2):975-986, 2003.
30. C. Guger, S. Daban, E. Sellers, C. Holzner, G. Krausz, R. Carabalona, F. Gramatica, and G. Edlinger. How many people are able to control a P300-based brain-computer interface (BCI)? Neuroscience letters, 462(1):94-98, 2009.
31. C. Guger, G. Edlinger, W. Harkam, I. Niedermayer, and G. Pfurtscheller. How many people are able to operate an eeg-based brain-computer interface (bci)? Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):145-147, 2003.
32. T. Hinterberger, S. Schmidt, N. Neumann, J. Mellinger, B. Blankertz, G. Curio, and N. Birbaumer. Brain-computer communication and slow cortical potentials. Biomedical Engineering, IEEE Transactions on, 51(6):1011-1018, 2004.
33. G.L. Holmes. Basic principles, clinical applications, and related fields. Journal of Epilepsy, 10(1):47, 1997.
34. H.H. Jasper. The ten-twenty electrode system of the international federation. Electro-encephalography and clinical neurophysiology, 10(2):371-375, 1958.
35. Matthias Kaper, Peter Meinicke, Ulf Grossekathoefer, Thomas Lingner, and Helge Ritter. BCI competition 2003-data set iib: Support vector machines for the p300 speller paradigm. Biomedical Engineering, IEEE Transactions on, 51(6):1073-1076, 2004.
36. M.B. Khalid, N.I. Rao, I. Rizwan-i Haque, S. Munir, and F. Tahir. Towards a brain computer interface using wavelet transform with averaged and time segmented adapted wavelets. InComputer, Control and Communication, 2009. IC4 2009. 2nd International Conference on, pages 1-4. IEEE, 2009.
37. W. Klimesch. EEG-alpha rhythms and memory processes. International Journal of Psychophysiology, 26(1-3):319-340, 1997.
38. W. Klimesch, M. Doppelmayr, A. Yonelinas, NEA Kroll, M Lazzara, D Röhm, and W. Gruber. Theta synchronization during episodic retrieval: neural correlates of conscious awareness. Cognitive Brain Research, 12(1):33-38, 2001.
39. M. Krauledat, M. Tangermann, B. Blankertz, and K.R. Müller. Towards zero training for brain-computer interfacing. PLoS One, 3(8):e2967, 2008.
40. Andrea Kubler, Boris Kotchoubey, Jochen Kaiser, Jonathan R Wolpaw, and Niels Birbau-mer. Brain-computer communication: unlocking the locked in. Psychological bulletin, 127(3):358-375, 2001.
41. Kwang-Hyuk Lee, Leanne M Williams, Michael Breakspear, Evian Gordon, et al. Synchronous gamma activity: a review and contribution to an integrative neuroscience model of schizophrenia. Brain Research Reviews, 41(1):57-78, 2003.
42. R. Leeb, D. Friedman, G.R. Müller-Putz, R. Scherer, M. Slater, and G. Pfurtscheller. Self-paced (asynchronous) BCI control of a wheelchair in virtual environments: a case study with a tetrap-legic. Computational intelligence and neuroscience, 2007:7, 2007.
43. Simon P Levine, Jane E Huggins, Spencer L BeMent, Ramesh K Kushwaha, Lori A Schuh, Mitchell M Rohde, Erasmo A Passaro, Donald A Ross, Kost V Elisevich, and Brien J Smith. A direct brain interface based on event-related potentials. Rehabilitation Engineering, IEEE Transactions on, 8(2):180-185, 2000.
44. Y. Li, J. Wu, J. Yang, et al. Developing a logistic regression model with cross-correlation for motor imagery signal recognition. InComplex Medical Engineering (CME), 2011 IEEE/ICME International Conference on, pages 502-507. IEEE, 2011.
45. Alberto Llera, MAJ Van Gerven, Vicenc Gomez, O. Jensen, and H.J. Kappen. On the use of interaction error potentials for adaptive brain computer interfaces. Neural Networks, 24(10):1120-1127, 2011.
46. D.J. McFarland, D.J. Krusienski, W.A. Sarnacki, and J.R. Wolpaw. Emulation of computer mouse control with a noninvasive brain-computer interface. Journal of neural engineering, 5:101, 2008.
47. D.J. McFarland, L.M. McCane, S.V. David, and J.R. Wolpaw. Spatial filter selection for EEG-based communication. Electroencephalography and clinical Neurophysiology, 103(3):386-394, 1997.
48. J. Millan, J. Mourio, et al. Asynchronous BCI and local neural classifiers: an overview of the adaptive brain interface project. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):159-161, 2003.
49. E.M. Mugler, C.A. Ruf, S. Halder, M. Bensch, and A. Kubler. Design and implementation of a P300-based brain-computer interface for controlling an internet browser. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 18(6):599-609, 2010.
50. G.R. Muller-Putz and G. Pfurtscheller. Control of an electrical prosthesis with an ssvep-based bci. Biomedical Engineering, IEEE Transactions on, 55(1):361-364, 2008.
51. F. Nijboer, E.W. Sellers, J. Mellinger, M.A. Jordan, T. Matuz, A. Furdea, S. Halder, U. Mochty, D.J. Krusienski, T.M. Vaughan, et al. A P300-based brain-computer interface for people with amyotrophic lateral sclerosis. Clinical neurophysiology, 119(8):1909-1916, 2008.
52. J.V. Odom, M. Bach, C. Barber, M. Brigell, M.F. Marmor, A.P. Tormene, and G.E. Holder. Visual evoked potentials standard. Documenta ophthalmologica, 108(2):115-123, 2004.
53. L.C. Parra, C.D. Spence, A.D. Gerson, and P. Sajda. Response error correction-a demonstration of improved human-machine performance using real-time eeg monitoring. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 11(2):173-177, 2003.
54. W.M. Perlstein, M.A. Cole, M. Larson, K. Kelly, P. Seignourel, and A. Keil. Steady-state visual evoked potentials reveal frontally-mediated working memory activity in humans. Neuroscience letters, 342(3):191-195, 2003.
55. G. Pfurtscheller, C. Brunner, A. Schlogl, and FH Lopes da Silva. Mu rhythm (de) synchronization and EEG single-trial classification of different motor imagery tasks. NeuroImage, 31(1):153-159, 2006.
56. G. Pfurtscheller and C. Neuper. Motor imagery and direct brain-computer communication. Proceedings of the IEEE, 89(7):1123-1134, 2001.
57. J.A. Pineda. The functional significance of mu rhythms: Translating. Brain Research Reviews, 50(1):57-68, 2005.
58. D. Regan. Human brain electrophysiology: Evoked potentials and evoked magnetic fields in science and medicine. 1989.
59. Wojciech Samek, Carmen Vidaurre, Klaus-Robert Müller, and Motoaki Kawanabe. Stationary common spatial patterns for brain-computer interfacing. Journal of Neural Engineering, 9(2):026013, 2012.
60. Marten K Scheffers, Michael GH Coles, Peter Bernstein, William J Gehring, and Emanuel Donchin. Event-related brain potentials and error-related processing: An analysis of incorrect responses to go and no-go stimuli. Psychophysiology, 33(1):42-53, 1996.
61. E.M. Schmidt, J.S. McIntosh, L. Durelli, and M.J. Bak. Fine control of operantly conditioned firing patterns of cortical neurons. Experimental neurology, 61(2):349-369,1978.
62. N.M. Schmidt, B. Blankertz, and M.S. Treder. Online detection of error-related potentials boosts the performance of mental typewriters. BMC neuroscience, 13(1):19, 2012.
63. E.W. Sellers, T.M. Vaughan, J.R. Wolpaw, et al. A brain-computer interface for long-term independent home use. Amyotrophic lateral sclerosis, 11(5):449, 2010.
64. P. Shenoy, M. Krauledat, B. Blankertz, R.P.N. Rao, and K.R. Müller. Towards adaptive classification for BCI. Journal of Neural Engineering, 3:R13, 2006.
65. K. Tanaka, K. Matsunaga, and H.O. Wang. Electroencephalogram-based control of an electric wheelchair. Robotics, IEEE Transactions on, 21(4):762-766, 2005.
66. G. Visconti, B. Dal Seno, M. Matteucci, and L. Mainardi. Automatic recognition of error potentials in a P300-based brain-computer interface. In Proceedings of the 4th International Brain-Computer Interface Workshop & Training Course, pages 238-243, 2008.
67. Y. Wang, R. Wang, X. Gao, B. Hong, and S. Gao. A practical vep-based brain-computer interface. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 14(2):234-240, 2006.
68. J.R. Wolpaw and D.J. McFarland. Control of a two-dimensional movement signal by a noninvasive brain-computer interface in humans. Proceedings of the National Academy of Sciences of the United States of America, 101(51):17849, 2004.
ZAVGORODNYAYA ELENA - master of Science in Physics and Astronomy, Radboud University Nijmegen, Netherlands, Rotterdam (oliverres@mail.ru).
ЗАВГОРОДНЯЯ ЕЛЕНА ВЛАДИМИРОВНА - магистр физики и астрономии, Университет Неймегена имени святого Радбота, Нидерланды, Роттердам.
