Online detection of error-related potentials boosts the performance of mental typewriters
© Schmidt et al; licensee BioMed Central Ltd. 2012
Received: 14 October 2011
Accepted: 15 February 2012
Published: 15 February 2012
Increasing the communication speed of brain-computer interfaces (BCIs) is a major aim of current BCI-research. The idea to automatically detect error-related potentials (ErrPs) in order to veto erroneous decisions of a BCI has been existing for more than one decade, but this approach was so far little investigated in online mode.
In our study with eleven participants, an ErrP detection mechanism was implemented in an electroencephalography (EEG) based gaze-independent visual speller.
Single-trial ErrPs were detected with a mean accuracy of 89.1% (AUC 0.90). The spelling speed was increased on average by 49.0% using ErrP detection. The improvement in spelling speed due to error detection was largest for participants with low spelling accuracy.
The performance of BCIs can be increased by using an automatic error detection mechanism. The benefit for patients with motor disorders is potentially high since they often have rather low spelling accuracies compared to healthy people.
KeywordsBrain-computer interface Electroencephalography ERP-Speller Error-related potentials Information transfer rate
Brain-computer interfaces (BCIs) establish a direct communication link between the human brain and an electronic device [1, 2]. The intent of the user is 'decoded' from her/his brain signals, e.g. from electroencephalography (EEG) or magnetoencephalography (MEG), and transformed into control commands for an external device. A great amount of research focuses on restoring sensory-motor functionality or communication ability in people who suffer from motor disorders, such as amyotrophic lateral sclerosis (ALS) . For ALS patients, BCI is a promising technology , because it can restore their ability to communicate wishes and needs and to interact with their environment, e.g. by controlling a spelling application [5, 6], a PC-cursor , or a wheelchair .
In EEG-based BCIs, many approaches capitalize on event-related potentials (ERPs) that arise as a response to sensory stimulation. An often targeted ERP component is the P300, a positive deflection at central and parietal electrode sites about 300 ms after the presentation of a stimulus that the user is attending to. The P300 and other ERP components have been successfully used as features in BCI spelling applications in order to identify the characters the user intends to write. The classic spelling application is the so-called P300-speller introduced by Farwell and Donchin , which is denoted here more specifically as Matrix Speller. It consists of a 6 × 6 matrix of characters. Each row and column is intensified (flashed) briefly in a random order, while the user is directing her/his gaze to the target character. Since detecting the P300 in single trials is intricate, the intensification sequence is repeated several times. By optimizing the number of sequence repetitions, the duration of the flashes, as well as the classification methods, a spelling speed of up to 5.8 characters per minute has been reported .
Compared to alternative technologies such as eye-trackers or EOG-based systems, where users communicate with up to 10 words per minute , this spelling speed is rather low. Therefore, currently, the clinical application of BCI spellers is mainly of interest in cases of severe oculomotor impairment. It has been shown however, that the spelling accuracy of the Matrix Speller also depends on the user's capability to direct her/his eye gaze to the desired target character. The accuracy drops critically low when the user is required to fixate a dot in the center of the matrix with her/his eyes [12, 13].
Recently, some novel approaches for visual spellers have been proposed to overcome this restriction [14–16]. Our study builds on the so-called Center Speller [15, 17], but the method could similarly be applied to other spellers. The Center Speller is a visual ERP-speller, which uses a two-step selection process: first, six groups of five characters are presented one by one in a fast sequence in the center of the screen. The user is attending to the target group, i.e. is waiting for its appearance. In the second step, the characters of the previously selected group are presented in the same way. In both steps, the six choices are coupled to simple geometric shapes of unique colors in order to facilitate the allocation of attention in fast stimulus sequences (see  and method section for a more detailed description).
As mentioned above, a bottleneck of current state-of-the-art BCIs is the low information throughput. For the Center Speller, a previous study showed an average spelling speed of about 1.5 characters/minute at 10 sequence repetitions (i.e. each of the six groups/characters is presented 10 times) . Several approaches have been explored in order to increase communication speed. One possibility is to reduce the number of repetitions, at the risk of decreasing spelling accuracy and fatigue of the participant. An optimal balance between the number of repetitions and accuracy can be achieved by means of a dynamic stopping method that statistically evaluates the confidence of the classification after each intensification sequence. If the classifier is confident about the selection, the presentation sequence is stopped [18–20]. Another factor affecting communication speed is experimental overhead. In the Center Speller, the selection process for each character begins with a countdown before the sequence presentation starts. Furthermore, it contains a few animations and presentation of the selected character (feedback). Spelling speed can be increased by reducing the durations of countdown, feedback and animations. As with reduction of repetitions, a potential drawback in reducing the overhead is that a too-short spelling process could be exhausting to the user because it may require more attention.
A different to increasing the spelling speed is the detection of error-related potentials (ErrPs). ErrPs are a certain type of ERPs that are present in the EEG signals when the user is aware of erroneous behavior. ErrPs probably arise in the anterior cingulate cortex, a brain area involved in processing of emotion and attention, and are thus found over central and prefrontal electrode positions . They are characterized by an early negative voltage deflection over fronto-central regions, referred to as error-negativity (NE) or error-related negativity, followed by a positive deflection over parietal regions, referred to as error-positivity (P E ) . The characteristics of the ErrPs vary, depending on the situation in which the erroneous behavior was perceived. In errors during a choice reaction task, where the subjects respond to a stimulus by pressing a button, erroneous button presses yield ErrPs that are sometimes referred to as "response ErrPs". The N E appears after 80 ms, the larger P E follows around 200-500 ms relative to the button press [23, 24]. When users perform wrong in a reinforcement learning task and receive a feedback indicating the wrong action, the observed main component is the N E around 250 ms after the stimulus and this is referred to as "feedback ErrP" . When users observe erroneous behavior of other persons, the so-called "observation ErrP" appears to be similar to the feedback ErrP . In BCI experiments the situation is different. Errors are usually neither caused by the user's action nor by another person the user is observing but by the misclassification of the BCI. Interestingly, in this case ErrPs also arise, with an N E component after 270 ms and a larger P E component 350-450 ms after the appearance of the BCI's feedback [26–31]. Ferrez and Millán  coined the term "interaction ErrP" for this type of ErrP.
Few studies have been conducted so far on the detection of interaction ErrPs. ErrP detection has been used to detect error trials offline in EEG-data of motor imagery experiments , in EEG-data of button press experiments with artificially induced errors , in MEG-data of covert attention experiments , as well as in EEG-data of Matrix Speller experiments . Dal Seno et. al  used online ErrP detection in pseudo-online Matrix Speller experiments with five healthy participants, and later in online Matrix Speller experiments with three participants . Spüler  showed successful online ErrP detection with the Matrix Speller in 12 healthy participants (29.5% increase of bit rate) and 4 patients with motor disorders (35.6% increase of bit rate).
The aim of the present study was to investigate, whether the communication rate of gaze-independent BCIs can be increased using online detection of ErrPs. To this end, an error detection mechanism was implemented in the Center Speller. If an error potential was detected by the ErrP classifier upon presentation of the classified symbol, the selection was vetoed and the trial was restarted. The communication rate in characters/minute of this modified speller was then compared to the communication rate of the Center Speller without error detection. Moreover, two different ErrP classifiers were compared; one classifier was trained on Center Speller data and another one was trained on data of a calibration experiment and was then applied in the Center Speller experiment. In the Methods section, both classifiers are introduced and the experimental protocol is explicated. In the Results section, we report on the neurophysiological data and on the impact of error potential detection on communication rate.
After having presented all stimulus sequences of a level, the feedback indicating the selected group or character is presented for 1s. In case of a wrong selection by the classifier, an ErrP is elicited and can be detected by another classifier. If this classifier detects an ErrP, a red 'X' appears over the feedback to indicate that the selection by the group or character is vetoed and the stimulus sequence is immediately repeated.
For the case that a wrong group was selected but no ErrP could be detected, the character level provides a backdoor indicated by an accent character". By selecting the backdoor, one returns to the group level without spelling a character. If a wrong selection occurred at the character level, a correction can be made via the less-than symbol ' < ', which serves as a backspace.
ErrP calibration speller
In order to train an ErrP classifier, a sufficient number of trials needs to be collected. However, the Center Speller yields a spelling speed of about two characters per minute, i.e. four samples of feedback evaluation per minute. In other words, to obtain a moderate training set for the ErrP detection, participants need to engage in a long spelling session. Depending on the desired size of the training set, this could exceed the reasonable duration of an experiment.
In order to obtain error potentials of the type interaction-ErrP, artificial errors were induced. In these cases, the element located on the side opposite to the selected one was selected instead. Choosing the symbol at the opposite side of the screen guaranteed that the participant did not misperceive the interaction error as her/his own error (own errors occurred when the participant hit the key while the arrow was not pointing on the target, i.e. too early or too late).
Twelve participants (7 males and 5 females), aged 23-31 years (μ = 26), participated in the study. One participant was excluded due to lack of BCI-control. The spelling accuracy for this participant was below 50% which made it impossible to spell and the experiment was aborted. All but one participant were right-handed and all had normal or corrected-to-normal visual acuity. Normal color vision in all but one participant (iac) was confirmed using the Ishihara color vision test . All participants gave written consent and the study was performed in accordance with the Declaration of Helsinki.
EEG was recorded using a Brain Products (Munich, Germany) actiCAP active electrode system with 64 electrodes and Brain Amp amplifiers sampling at a rate of 1000 Hz. The electrodes were placed according to the international 10-10 system at positions Fp1,2, AF3,4,7,8, Fz, F1-10, FCz, FC1-6, FT7,8, T7,8, Cz, C1-6, TP7,8, CPz, CP1-6, Pz, P1-10, POz, PO3,4,7,8, Oz and O1,2. One electrode was placed under the right eye and labeled as EOGvu. Active electrodes were referenced to a nose electrode, using forehead ground. Impedances were kept below 15 kΩ. EEG signals were hardware filtered at 0.016-250 Hz. The stimuli were presented on a 24" TFT screen with a resolution of 1920 × 1200px 2 and a refresh rate of 60 Hz. Participants were seated at 70 cm distance from the screen. To correct the EEG markers for the TFT latency, a photo diode (g.TRIGbox; g.tec medical engineering, Graz, Austria) was attached to the lower left corner of the screen for the first six experiments, registering the exact stimulus onset. The median TFT latencies over the six experiments range from 69 ms to 71 ms and for offline analysis, the mean value of 69.8 ms was added to the EEG marker times of all experiments. The Center Speller and the Calibration Speller were implemented in Python http://www.python.org using the open-source-framework Pyff  with VisionEgg  and Pygame http://www.pygame.org. Remote-controlling of the experiments, online classification as well as offline analysis was done with an inhouse toolbox using MATLAB (The MathWorks, Natick, MA, USA). The Center Speller is freely available in the Pyff repository (see http://bbci.de/pyff).
For online classification, the EEG data was downsampled to 100 Hz and baseline corrected for the 200 ms prestimulus interval (both, for the speller stimuli and for the feedbacks, which form the ErrP stimuli).
For offline analysis, the data was lowpass filtered using a Chebyshev filter with 42 Hz passband and 49 Hz stopband and then downsampled to 250 Hz. The continuous signal was divided into epochs ranging from -200 ms to 1200 ms relative to the onset of the stimulus and epochs were baseline corrected for the 200 ms prestimulus interval. Artifacts were detected to account for eye blinks, eye movements, muscular activity and malfunctioning hardware. Trials and channels containing such artifacts were rejected for visual ERP analysis, but not for classification purposes. The artifact detection was done using a variance criterion, i.e. channels and trials with too low or too high voltage variance were labeled as contaminated by artifacts, as well as using a min-max criterion, i.e. all trials in which the difference between maximum and minimum voltage exceeds 75 μV were labeled as contaminated by artifacts. In the ErrP Calibration Speller experiments, all trials containing errors made by the participant were also rejected.
The signed square of the point-biserial correlation coefficient sgn r 2 was used for ERP analysis as a measure for the discriminability of two classes (target vs. non-target or error vs. non-error) .
The performance of the spelling classifier (i.e. classification accuracy) is given as the percentage of correctly selected symbols (per level). The performance of the ErrP classifiers is given in terms of the receiver-operating-characteristic (ROC). An ROC curve can be depicted as a plot of false alarms against hits. Since a good classifier allows for a high hit rate at few false alarms, the area under the ROC curve (AUC) is a commonly used quantification of classifier performance. Furthermore, the accuracy of the ErrP detection is split into hits (an error trial classified as error) and false alarms (a correct trial classified as error), or their respective rates for some analyses. The spelling speed is given in terms of the number of characters that were spelled per minute, abbreviated char/min.
All classifiers, that is, the spelling classifier (used for detecting the target symbol) and the two ErrP classifiers (used for detecting ErrPs upon presentation of the feedback), used spatio-temporal features for a linear discriminant analysis with shrinkage of the covariance matrix (see e.g. ). In the spatial domain, all electrodes except for Fp1,2, AF3,4,7,8 and EOGvu were considered (57 channels) for online classification. In the temporal domain, a heuristic method  was used, searching for peaks in the sgn r 2 between targets and non-targets in the 100-700 ms post-stimulus interval (for the spelling), and between errors and non-errors in the 150-900 ms post-feedback interval (for the ErrP detection), respectively. The heuristic method initially determined 5 temporal intervals, but the number of intervals and the exact temporal position of them could be adjusted by the experimenter before the online operation was started. Finally, the voltages of all selected electrodes were averaged within the selected intervals, constituting a feature vector of length d = nIvals nElectrodes (d = 285 in case of nIvals = 5 intervals and nElectrodes = 57 electrodes).
In an offline analysis, the spatial distribution of the class-discriminative information for ErrP detection was investigated by training one classifier individually for each electrode channel . For each channel, four time intervals were chosen automatically by selecting peaks in the sgn r 2 values. Voltages were then averaged within the respective intervals resulting in four dimensional features. Training and test sets were chosen in the same two ways as in the online experiments, relating to classifier A and to classifier B. We refer to these classifiers as type-A and type-B classifiers.
To investigate whether artifacts from eye blinks or raised eyebrows could explain the classification results, classification was repeated offline for frontal electrodes (Fp2, F9, F10, EOGvu) that were most susceptible to ocular artifacts.
Results of the online study.
of Classifier A
of Classifier A
of Classifier B
of Classifier B
(Accuracy l1, l2/Speed)
Spelling speed improvement
Spatial distribution of discriminative information
ErrP detection performance was lower for frontal electrodes (Fp2, F9, F10, EOGvu) than for all other electrodes (type-A: t = 2.89, p < 0.05; type-B: t = 4.42, p < 0.01). This is in line with Figure 9, where classification performance is at a minimum for frontal channels. These results suggests that ocular artifacts are unlikely to substantially contribute to successful error detection.
Single-trial ErrPs were detected with a mean accuracy of 89.1% (AUC 0.90). The online detection rate was similar to the cross-validation results in offline studies, where 82%  and 80%  have been reported. ErrP detection using a classifier trained on the online data increased the mean spelling speed by 49.0% compared to the case without ErrP detection. A similar rate of improvement was obtained by  with ErrP detection in Matrix Speller experiments (29.5% increase of the bit rate). This illustrates that ERP spellers can be enhanced significantly by detecting and vetoing erroneous decisions of the BCI based on error potentials. Furthermore, the gain in communication speed was relatively higher for participants with a medium or low BCI performance (say, > 10% errors) than for participants with a high BCI performance.
In some cases, ErrP detection could impede spelling speed instead of accelerating it. False alarms prolonged the spelling process because a correct selection was vetoed and had to be repeated. In particular, in cases where participants produced few errors (due to high spelling accuracy), the potential of improvement due to error detection was limited and could easily be outweighed by the detrimental effect of false alarms. This shows that the balance between hits and false alarms of the ErrP classifier has a crucial influence on the overall spelling performance in terms of speed. By moving the decision boundary of the ErrP classifier (ErrP bias), this balance can be controlled by the experimenter. Hence, the trade-off that maximizes communication speed is not only a function of the number of repetitions (which affects the spelling accuracy and thus the speed), but is also affected by the placement of the ErrP bias. Due to the recursive nature of the speller paradigm (the ErrP detector can potentially veto every trial and lead to an infinite loop), finding the optimal trade-off is an intricate problem that will be addressed in future theoretical work. One approach to be mentioned here could be to use estimates of the spelling accuracy for different repetition numbers, together with estimates of the hit rate and false alarm rate of the ErrP classifier for different bias values (both could be obtained e.g. from the calibration blocks). Knowing the duration of a Center Speller trial, one could simulate the spelling process for different values of repetition number and ErrP bias in a Monte-Carlo fashion and chose the combination of the parameters that maximizes the speed.
A drawback of using ErrPs is the fact that one has to collect a substantial number of error trials in order to train the ErrP classifier. There are different possible routes to accomplish this. First, one may perform an experiment in two successive stages. In the first stage, spelling would be done without error correction. In the next stage, the trials would have to be labeled and used for training an ErrP classifier (just as we did for classifier B). Second, one could use a calibration phase to collect trials for the ErrP classifier (as we did for classifier A). Regarding the second case, our data show that an ErrP classifier trained in one paradigm (classifier A in the present study) can transfer to a similar paradigm, albeit with a reduced performance. However, the Calibration Speller is not applicable in a clinical context because it involves key presses. However, a calibration experiment that would be completely passive and thus applicable to patients, similar to the offline calibration phase of the spelling classifier, could be used to collect ErrP data for classifier training. It is true that the ErrPs from such a calibration experiment may have large differences to the ErrPs obtained in the Center Speller ('observation ErrPs' instead of 'interaction ErrPs'). Therefore, the applicability of such an approach remains to be investigated.
Ultimately, the utility of ErrP detection is dictated by whether a successful implementation in a clinical setting is feasible. ErrP detection could be relevant to patients because their BCI performance is more variable and often lower than the performance of healthy participants. However, this critically depends on whether error potentials can be detected reliably in patients. Regarding this question, the work of Spüler  is instructive. In a clinical study with four patients, ErrPs were classified with an accuracy of 71%. Using a Matrix speller, the bit rate was increased by 35.6% on average. If patients are in a progressed state of the locked-in syndrome, a possible approach for calibration of the ErrP classifier would be to have patients passively observe errors and train on the resulting observation ErrPs, as outlined above .
Concluding, we demonstrated a significant increase of communication speed of gaze-independent ERP spellers when error potentials are detected online. Since BCI performance is often low in patients and successful detection of ErrPs has been demonstrated in ALS patients , we believe that ErrP detection can complement conventional BCIs in a clinical application.
amyotrophic lateral sclerosis
area under the ROC curve
analysis of variance.
- Kübler A, Müller KR: An introduction to brain computer interfacing. Toward Brain-Computer Interfacing. Edited by: Dornhege G, del R Millán J, Hinterberger T, McFarland D, Müller KR. 2007, Cambridge, MA: MIT Press, 1-25.
- Kübler A, Kotchoubey B, Kaiser J, Wolpaw J, Birbaumer N: Brain-computer communication: unlocking the locked in. Psychol Bull. 2001, 127 (3): 358-375.PubMedView Article
- Kübler A, Nijboer F, Birbaumer N: Brain-Computer Interfaces for Communication and Motor Control - Perspectives on Clinical Applications. Toward Brain-Computer Interfacing. Edited by: Dornhege G, de R Millán J, Hinterberger T, McFarland D, Müller KR. 2007, Cambridge: MIT Press, 373-391.
- Huggins JE, Wren PA, Gruis KL: What would brain-computer interface users want? Opinions and priorities of potential users with amyotrophic lateral sclerosis. Amyotroph Lateral Scler. 2011, 12: 318-324. 10.3109/17482968.2011.572978. [http://informahealthcare.com/doi/abs/10.3109/17482968.2011.5 72978]PubMed CentralPubMedView Article
- Nijboer F, Sellers EW, Mellinger J, Jordan MA, Matuz T, Furdea A, Halder S, Mochty U, Krusienski DJ, Vaughan TM, Wolpaw JR, Birbaumer N, Kübler A: A P300-based brain-computer interface for people with amyotrophic lateral sclerosis. Clin Neurophysiol. 2008, 119: 1909-1916. 10.1016/j.clinph.2008.03.034.PubMed CentralPubMedView Article
- Sellers EW, Vaughan TM, Wolpaw JR: A brain-computer interface for long-term independent home use. Amyotrophic Lateral Sclerosis. 2010, 5: 449-455. [http://informahealthcare.com/doi/abs/10.3109/174829610037774 70]View Article
- McFarland DJ, Krusienski DJ, Sarnacki WA, Wolpaw JR: Emulation of computer mouse control with a noninvasive brain-computer interface. J Neural Eng. 2008, 5: 101-110. 10.1088/1741-2560/5/2/001. [http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2757111]PubMed CentralPubMedView Article
- Leeb R, Friedman D, Müller-Putz G, Scherer R, Slater M, Pfurtscheller G: Self-Paced (Asynchronous) BCI Control of a Wheelchair in Virtual Environments: A Case Study with a Tetraplegic. Comput Intell Neurosci. 2007, 2007: 79642.PubMed Central
- Farwell L, Donchin E: Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalogr Clin Neurophysiol. 1988, 70: 510-523. 10.1016/0013-4694(88)90149-6.PubMedView Article
- Lenhardt A, Kaper M, Ritter H: An adaptive P300-based online brain-computer interface. IEEE Trans Neural Syst Rehabil Eng. 2008, 16: 121-130.PubMedView Article
- Majaranta P, MacKenzie S, Aula A, Räihä KJ: Effects of feedback and dwell time on eye typing speed and accuracy. Univ Access Inf Soc. 2006, 5 (2): 199-208. 10.1007/s10209-006-0034-z.View Article
- Treder MS, Blankertz B: (C)overt attention and visual speller design in an ERP-based brain-computer interface. Behav Brain Funct. 2010, 6: 28-10.1186/1744-9081-6-28. [http://www.behavioralandbrainfunctions.com/content/6/1/28]PubMed CentralPubMedView Article
- Brunner P, Joshi S, Briskin S, Wolpaw JR, Bischof H, Schalk G: Does the "P300" speller depend on eye gaze?. J Neural Eng. 2010, 7: 056013-10.1088/1741-2560/7/5/056013.PubMed CentralPubMedView Article
- Acqualagna L, Treder MS, Schreuder M, Blankertz B: A novel brain-computer interface based on the rapid serial visual presentation paradigm. Conf Proc IEEE Eng Med Biol Soc. 2010, 2010: 2686-2689. [http://dx.doi.org/10.1109/IEMBS.2010.5626548]PubMed
- Treder MS, Schmidt NM, Blankertz B: Gaze-independent brain-computer interfaces based on covert attention and feature attention. J Neural Eng. 2011, 8 (6): 066003-10.1088/1741-2560/8/6/066003. [Open Access], [http://dx.doi.org/10.1088/1741-2560/8/6/066003]PubMedView Article
- Liu Y, Zhou Z, Hu D: Gaze independent brain-computer speller with covert visual search tasks. Clin Neurophysiol. 2011, 122 (6): 1127-1136. 10.1016/j.clinph.2010.10.049.PubMedView Article
- Treder MS, Schmidt NM, Blankertz B: Towards gaze-independent visual brain-computer interfaces. Front Comput Neurosci. 2010, [Conference Abstract: Bernstein Conference on Computational Neuroscience 2010], [http://dx.doi.org/10.3389/conf.fncom.2010.51.00117]
- Schreuder M, Höhne J, Treder MS, Blankertz B, Tangermann M: Performance Optimization of ERP-Based BCIs Using Dynamic Stopping. Front Comput Neurosci. 2011, 2011: 4580-4583.
- Höhne J, Schreuder M, Blankertz B, Tangermann M: Two-dimensional auditory P300 Speller with predictive text system. Conf Proc IEEE Eng Med Biol Soc. 2010, 2010: 4185-4188. [http://dx.doi.org/10.1109/IEMBS.2010.5627379]PubMed
- Zhang H, Guan C, Wang C: Asynchronous P300-based brain-computer interfaces: a computational approach with statistical models. IEEE Trans Biomed Eng. 2008, 55: 1754-1763.PubMedView Article
- Holroyd C, Coles M: The neural basis of human error processing: reinforcement learning, dopamine, and the error-related negativity. Psychol Rev. 2002, 109: 679-709.PubMedView Article
- Falkenstein M, Hoormann J, Christ S, Hohnsbein J: ERP components on reaction errors and their functional significance: a tutorial. Biol Psychol. 2000, 51 (2-3): 87-107. 10.1016/S0301-0511(99)00031-9.PubMedView Article
- Blankertz B, Dornhege G, Schäfer C, Krepki R, Kohlmorgen J, Müller KR, Kunzmann V, Losch F, Curio G: Boosting bit rates and error detection for the classification of fast-paced motor commands based on single-trial eeg analysis. IEEE Trans Neural Syst Rehabil Eng. 2003, 11 (2): 127-131. 10.1109/TNSRE.2003.814456. [http://dx.doi.org/10.1109/TNSRE.2003.814456]PubMedView Article
- Parra L, Spence C, Gerson A, Sajda P: Response error correction - a demonstration of improved human-machine performance using real-time eeg monitoring. IEEE Trans Neural Syst Rehabil Eng. 2003, 11 (2): 173-177. 10.1109/TNSRE.2003.814446.PubMedView Article
- Schie HTV, Mars RB, Coles MG, Bekkering H: Modulation of activity in medial frontal and motor cortices during error observation. Nat Neurosci. 2004, 7 (5): 549-554. 10.1038/nn1239.PubMedView Article
- Ferrez P, Millán J: You are wrong! - Automatic detection of interaction errors from brain waves. 19th International Joint Conference on Artificial Intelligence. 2005, 1413-1418.
- Buttfield A, Ferrez P, Millán JDR: Towards a robust BCI: error potentials and online learning. IEEE Trans Neural Syst Rehabil Eng. 2006, 14: 164-168. 10.1109/TNSRE.2006.875555.PubMedView Article
- Ferrez P, Millán J: Error-related eeg potentials generated during simulated brain-computer interaction. IEEE Trans Biomed Eng. 2008, 55: 923-929.PubMedView Article
- Spüler M: Online Erkennung von Fehlerpotentialen zur Fehlerkorrektur bei einem P300 Brain-Computer Interface. Eberhard Karls Universität Tübingen. 2010, [Diploma thesis]
- Dal Seno B, Matteucci M, Mainardi L: Online Detection of P300 and Error Potentials in a BCI Speller. Comput Intell Neurosci. 2010, 2010: 307254-[http://dx.doi.org/10.1155/2010/307254]PubMed Central
- Combaz A, Chumerin N, Manyakov NV, Robben A, Suykens JAK, Van Hulle MM: Error-related potential recorded by EEG in the context of a p300 mind speller brain-computer interface. IEEE International Workshop on Machine Learning for Signal Processing, IEEE. 2010: 65-70. [http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=5589217]
- Schalk G, Wolpaw JR, McFarland DJ, Pfurtscheller G: EEG-based communication: presence of an error potential. Clin Neurophysiol. 2000, 111: 2138-2144. 10.1016/S1388-2457(00)00457-0.PubMedView Article
- Llera A, van Gerven M, Gómez V, Jensen O, Kappen H: On the use of interaction error potentials for adaptive brain computer interfaces. Neural Netw. 2011, 24 (10): 1120-1127. 10.1016/j.neunet.2011.05.006. [http://www.sciencedirect.com/science/article/pii/S0893608011 001481]PubMedView Article
- Dal Seno B, Matteucci M, Mainardi L: A genetic algorithm for automatic feature extraction in P300 detection. Proceedings of the International Joint Conference on Neural Networks (IJCNN'08). 2008, 3145-3152.
- Ishihara S: Tests for colour-blindness. 1919, Tokyo: Hongo Harukicho
- Venthur B, Scholler S, Williamson J, Dähne S, Treder MS, Kramarek MT, Müller KR, Blankertz B: Pyff - a pythonic framework for feedback applications and stimulus presentation in neuroscience. Front Neuroscience. 2010, 4: 179-[http://dx.doi.org/10.3389/fnins.2010.00179]View Article
- Straw AD: Vision egg: an open-source library for realtime visual stimulus generation. Front Neuroinformatics. 2008, 2: 4.PubMed CentralView Article
- Blankertz B, Lemm S, Treder MS, Haufe S, Müller KR: Single-trial analysis and classification of ERP components - a tutorial. NeuroImage. 2011, 56: 814-825. 10.1016/j.neuroimage.2010.06.048. [http://dx.doi.org/10.1016/j.neuroimage.2010.06.048]PubMedView Article
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.