Professional Documents
Culture Documents
Manuscript Number:
I am enclosing a submission to the Biomedical Signal Processing and Control entitled, "Signal
processing methods in motor-imagery BCI: A review". The manuscript is 15 pages long and
includes 7 figures. There are no related manuscripts that have been submitted for simultaneous
consideration to the same or to another journal.
I am enclosing also a certificate of professional proofreading and editing (see next page).
Sincerely,
DECLARATION
We wish to confirm that there are no known conflicts of interest associated with this
publication and there has been no significant financial support for this work that could have
influenced its outcome.
We confirm that the manuscript has been read and approved by all named authors and that
there are no other persons who satisfied the criteria for authorship but are not listed.
We further confirm that the order of authors listed in the manuscript has been approved by all of
us.
We confirm that we have given due consideration to the protection of intellectual property
associated with this work and that there are no impediments to publication, including the
timing of publication, with respect to intellectual property. In so doing we confirm that we
have followed the regulations of our institutions concerning intellectual property.
We understand that the Corresponding Author is the sole contact for the Editorial process
(including Editorial Manager and direct communications with the office). He is
responsible for communicating with the other authors about progress, submissions of
revisions and final approval of proofs. We confirm that we have provided a current, correct
email address which is accessible by the Corresponding Author and which has been
configured to accept email from (Dariusz Zapala d.zapala@gmail.com).
Piotr Wierzgała
Dariusz Zapała
Grzegorz Wójcik
Lublin, 18.01.2016
*Manuscript
Click here to view linked References
1
Institute of Computer Science, Maria Curie-Sklodowska University, Lublin, Poland
2
Department of Experimental Psychology, The John Paul II Catholic University of Lublin, Poland
Corresponding author:
Abstract
Brain-computer interfaces (BCI), sensorimotor rhythms, motor imagery, signal processing, EEG
Introduction
If compared with other types of brain-computer interfaces based on relatively automatic and
passive physiological responses (visual potentials SSVEP and P300 component), MI BCI is more
versatile and practicable, yet it is also less efficient and requires long-lasting individual training and
calibration sessions (Friedrich et al., 2009). Nevertheless, there are many examples of possible
applications of SMR-BCI for various purposes, such as communication (Müller & Blankertz, 2006),
control of prosthetic devices (Pfurtscheller et al., 2000), control of rehabilitation equipment (Huang
et al., 2012) and navigation in 2D and 3D space (Leeb et al., 2007).
Control by means of changes in electrical activity of the brain is enabled by the employed
procedures of EEG signal processing. The specific operations to which the signal is subjected are
designed to identify the characteristic changes in the continuous EEG recording (the so-called
Features Extraction) and to match the identified fragments with specific mental activities (the so-
called Features Classification) (McFarland & Krusienski, 2012).
In the case of motor-imagery based brain-computer interfaces, the changes identified in
EEG recording are associated with movements imagined by the user. Such changes may involve
e.g. a decrease in the power of sensorimotor rhythms on the electrodes placed centrally.
Subsequently, the identified changes are classified to match one of the states representing specific
intentions of the user (e.g. the thought about the right or left hand movement).
Feature extraction and classification in MI BCI is carried out with the use of numerous
algorithms, which differ in terms of effectiveness and are examined by many researchers (Lotte et
al., 2007). The procedure of multi-stage data processing in MI BCI is presented in diagram 2. The
choice of an adequate signal processing procedure influences interface effectiveness. The features
of changes in brain activity associated with motor imagery may, to a degree, differ in each
individual, relative to current internal and external factors. Given the above, particularly effective
procedures are those, which apply input data filtering for various factors such as space, time and
frequency.
-----------
Fig. 2
-----------
Fig. 2. Stages of data processing in MI BCI. Elaborated by the authors.
The article is an attempt to recapitulate the findings of reports presenting the effectiveness of
signal processing algorithms in MI BCI. It also takes into account details related to the method of
collecting and preparing the data subjected to analyses. The review is based on 103 articles,
published from 1997–2015 in peer-reviewed journals. The reports were identified using the feature
of keywords (e.g. motor imagery BCI, sensorimotor rhythms BCI, Graz BCI, Wandsworth BCI),
which were entered into search engines indexing scientific publications, i.e. Google Scholar and
ScienceDirect. The contents of the articles were analysed in order to identify structural components
of brain-computer interfaces, which significantly impact the effectiveness of the systems. In
addition to data preparation and processing methods, the analyses also focused on devices and
equipment used in conducting the study, their configuration and the data collecting procedure or the
applied set of data. The findings provided grounds for conclusions with regard to the dominating
trends in research of signal processing in MI BCI. The observed trends were also examined for their
consistency with developments and methodological requirements postulated in the literature.
The operation of each EEG-BCI system starts with recording the bioelectric activity of the
brain with the use of electrodes. From the point of view of brain-computer interfaces and their
operation, the critically important features include the system’s tolerance to distortions in EEG
recording (artefacts) and the ergonomics of using the device. Unlike the EEG systems used e.g. in
performing medical diagnoses or psychophysiological experiments, the equipment employed for the
specific needs of BCI must be easy to operate and maintain, must enable long-lasting operation
without a decrease in the accuracy of the recording and must be relatively cheap. This is because
brain-computer interfaces are mainly intended for individuals with disabilities using such devices at
home, and this is associated with a greater risk of artefacts. To meet the related requirements, EEG-
BCI systems generally utilize a lower number of electrodes than other is the case for
electroencephalographs used at present in clinical trials and primary research. By using less than 20
downleads, it is possible to reduce the costs of the device and enable rapid application (Portelli et
al., 2011).
The number of channels used in currently conducted studies is significantly varied (Fig. 3).
No evidence has been found for the use of records from the so-called high-density EEG nets, which
comprise over 128 channels. High-density recording sets ranging from 64 to 128 electrodes
constitute the smallest number among the identified groups (5% of cases). The reviewed articles
suggest that the most frequently used number of channels ranges from 9 to 16 (28% of cases).
-----------
Fig. 3
-----------
Fig 3. Number of electrodes in the analysed record.
Sensorimotor rhythm-based brain-computer interfaces are classified among active BCIs, so,
in their operation, they take advantage of the changes in the recording of brain activity in response
to the mental task, which is being performed. A number of different tasks are executed by users of
MI BCI to control the device. These include moving the cursor on the screen without any specific
instruction defining the contents of imagery (Friedrich et al., 2009), imagined movement vs. lack of
movement (Neuper et al., 2003), imagined right hand vs. left hand movement (Scherer, Müller-Putz
& Pfurtscheller, 2009) and imagined movement of the foot or tongue (Leeb et al., 2007). While
evaluating specific types of mental tasks applied in MI BCI, it is necessary to take into account the
number of decisions (choices), which can be made using a given procedure. The systems involving
imagined movement vs. relaxing and right vs. left hand movement enable a choice to be made
between two conditions (0 and 1), which is sufficient e.g. for virtual keyboard control but not for
movement in three-dimensional space (McFarland & Krusienski, 2012)
In the reviewed articles, the effectiveness of motor imagery recognition in BCI was most
frequently tested with the use of a signal representing two states (72.1% cases), imagined right and
left hand movement being the predominant procedure. Recordings, which represent three states,
such as the imagined movement of both hands and the tongue, are considerably less common
(Table 1).
Types of imagery Percentage of studies
left hand, right hand 45.09%
foot, left hand, right hand, tongue 12.27%
right foot, right hand 11.66%
foot, left hand, right hand 9.20%
left index finger, right index finger 3.68%
feet, left hand, right hand, tongue 2.45%
Other 13.65%
The majority of the reviewed articles present results acquired with the use of more than one
signal processing algorithm. As a rule, they compare a number of combinations of extractors,
selectors and classifiers tested with the same datasets. On the whole, the identified reports contained
findings of 326 experiments, yet only 13% of the total number were related to studies in which
algorithms were tested online, i.e. during real-time interface control. In the remaining cases, signal
processing was conducted based on datasets of EEG recordings performed during motor imagery
tasks. The majority of the latter group used the signals available in the framework of the so-called
BCI Competitions (Brunner et al., 2008; Leeb et al., 2008). This is a project, which has provided a
large group of researchers with access to EEG datasets to facilitate comparison of various signal
processing techniques and their effectiveness. In the review, 63% of the results were based on data
from BCI Competitions (Fig. 4).
----------
Fig. 4
----------
Fig. 4. Data used in studies: A – Online and offline experiments; B – Data from BCI
Competitions and the remaining sources of data.
Regardless of the recording acquisition method, the studies differed regarding the number of
subjects (Fig. 5). Most frequently, their number was in the range of N = 2–10 (77% of cases). A
significant group represented findings of case studies, i.e. involving one subject only (18%). Few
analyses were carried out for study groups of more than 10 subjects (5% of cases).
-----------
Fig. 5
-----------
Fig. 5. Sizes of populations in the analysed datasets.
The most frequently applied practice at the stage of data pre-processing involves band-pass
filtering of the signal. As a result, the signal retains only those frequency bands, which represent the
electrical activity of the brain connected with motor imagery. Only in a few cases (1.5%) was the
signal filtered individually for each participant of the study. Subject-specific frequency bands were
identified based on initial examinations. Such an approach made it possible to adjust the interface to
each subject’s brain activity. In most cases, all the recorded data were filtered in the same frequency
bands. MI BCIs based on EEG typically use the following waves: μ (8–12 Hz) and β (18–30 Hz)
(Pfurtscheller & McFarland, 2012) In most cases, the data designated for analyses were not
subjected to the procedure of removing artefacts from the signal, yet, in some variants, the
recording was purified manually.
-----------
Fig. 6
-----------
Fig. 6. Percentage of studies in which artefacts were removed manually.
During a BCI session, the recorded and pre-processed signal is subjected to further
processing in order to identify the features representing the activity, which will be used for the
interface-enabled control. Hence, information, which is of significance for the system operation, is
extracted, and the remaining data are rejected (Nicolas-Alonso & Gomez-Gil, 2012). In the case of
MI BCI, the search focuses on changes in the recording, which reflect a mental representation of a
specific movement.
MI BCIs may operate in synchronous and asynchronous ways. A user of synchronous BCI
performs imagery tasks in strictly defined moments and for a specified duration of time. In the case
of asynchronous BCIs, the moment and duration of the mental representation are unknown, and the
analysis of signal for the possible occurrence of imagery is a continuous process (Townsend,
Graimann & Pfurtscheller, 2004).
As a rule, the features extracted from the signal include specific amplitudes or values of
signal strength, which occurred in consecutive trials involving control with the use of the interface
(Krusienski, McFarland & Principe, 2012). The number of features depends on the method of signal
conversion, the number of trials and the number of channels used to perform the recording. In the
case of an interface whose operation is based on the strength of a signal in consecutive frequencies
ranging from 8 to 12 Hz (4 values) during a session of 600 trials recorded by a 32-channel EEG
system, the total number of extracted features will amount to 76,800 (4 x 600 x 32).
A number of signal conversion methods are employed at the stage of features extraction
(Lotte et al., 2007). Features within the EEG signal are most frequently determined with the variants
of Common Spatial Patterns method (48.23% cases), which, in terms of popularity, is followed by
the Band Power (13.62% cases).
Effective learning and efficient performance of the classifier are hindered by the excessive
number of features identified at the stage of extraction. Due to this, a data selection stage is used in
some interfaces in order to identify the effective features, which contribute the most information to
the classification process. The majority of studies did not use any algorithms for feature selection
(16.56% cases).
Signal classification
The stage of features extraction is followed by conversion into the response of the device or
computer application. This stage involves matching the previously extracted and selected features
with the state defined by the relevant type of brain-computer interface (Wolpaw & Wolpaw, 2012).
The state is a model representation of brain activity used in BCI enabled control. In MI BCI,
the states may represent e.g. the imagined movement of one hand, the tongue or one foot.
Depending on the number of imagery tasks utilized by a given interface, there is a varied number of
model states. The process of EEG signal conversion into BCI system response is a sequence of
algorithms, which are designed to ultimately match the input data with one of the states recognized
by the interface. Each BCI is subjected to the stages of training and testing. The states representing
model imagery tasks are defined at the training stage. BCI receives a time series representing the
activity of the brain during imagery and information about the type of imagery (state), which has
been performed. As a result of acquiring many pairs of such data, BCI can learn, to an extent, to
match EEG data with appropriate states. During the trials carried out at the stage of testing, the
interface is provided only with the data representing the activity of the brain (typically during the
performance of imagery tasks). The interface is to correctly identify the imagery or its absence in
the received input data. Based on the number of correctly identified imagery tasks, it is possible to
calculate the effectiveness of the interface.
Classification algorithms most frequently use one of two approaches: classification by
discrimination or classification by regression (McFarland et al., 2005). In classification by
discrimination, the algorithm divides input data into independent groups, each of which corresponds
to one of the states used in the interface-enabled control. In classification by regression, the
algorithm, taking into account the input data, calculates the value, which can be arranged in order
relative to the values computed for other input data. Determining the order this way enables
allocation of the input data to one of the states or translation of the brain activity into e.g. cursor
movement on a computer screen (McFarland & Krusienski, 2012).
The regression and discrimination based classification models comprise a number of signal
processing algorithms. These include such techniques as generative, linear, non-linear and
combined techniques, i.e. using features of the previous methods (Nicolas-Alonso & Gomez-Gil,
2012). The basic property of all the classifiers is the effectiveness of classification, i.e. the
efficiency with which they can translate the identified features into the desired response of the
device (McFarland & Krusienski, 2012). The effectiveness of classification is an indicator enabling
comparison of various classification methods (Lotte et al., 2007).
Cross validation is used to ensure the appropriate testing of classification effectiveness. In
the course of the cross validation procedure, the set of all data acquired during the examination is
subjected to multiple divisions into two subsets: the training and testing subset. The classifier is
taught to allocate input data to one of the states, based on the data in the training subset. The effects
of the learning process are verified with the use of the testing subset. The mean effectiveness of
classification for the test data in all of the executed divisions is a measure of classifier effectiveness.
The reviewed publications most frequently used signal classification methods based on such
algorithms as the Support Vector Machine (SVM) (39.57% of cases) and Linear Discriminant
Analysis (LDA) (31.90% of cases). The remaining techniques were reported by significantly fewer
studies.
The acceptable BCI classification effectiveness is, as a rule, defined at a minimum level of
60% (Guger et al., 2003) or 70% (Brunner et al., 2010). If a user achieves a score below this value,
it can be assumed that he/she will not be able to effectively operate the specific interface. The
accuracy of the majority of the tested algorithms exceeds both of these values (Fig. 7) Importantly,
the majority of the reported findings relate to comparisons conducted offline, which means that the
real-time accuracy of control performed by users by means of a given variant of BCI is unknown.
----------
Fig. 7
----------
Fig. 7. The reported accuracy of classification in the tested algorithms.
Conclusions
Since the last decade of the 20th century and in the beginning of the 21st century there has
been growing interest in issues related to brain-computer interfaces (Huggins & Wolpaw, 2014). As
pointed out by researchers, studies into BCI mainly focus on improving the effectiveness of signal
processing algorithms (Hwang et al., 2013). Nevertheless, the effectiveness of information transfer
aided by brain-computer interfaces is insufficient for widespread practical application of the
technique (Ahn & Jun, 2015). Because of the easy application and price, the most common today
are the encephalography based BCIs, and, in this group, those in which control is based on motor
imagery (Hwang et al., 2013). This explains the large number of articles focusing on classification
of signal containing motor imagery. Yet, it is unclear to what extent findings and conclusions from
these studies can be utilized in practice for engineering more effective SMR-BCI systems.
In accordance with the “golden standard” proposed for BCI-related research by McFarland
and Krusienski (2012), the most valuable results showing possible uses of a given algorithm in an
existing interface are acquired by means of both offline and online studies. This means that the new
method is subjected to testing with the use of generally accessible models and readily available sets
of records, as well as during sessions where users control the interface in real time. This approach
makes it possible to compare the findings of a specific study with the results acquired with other
techniques and to answer the question of whether a given method is effective if utilized in a real-life
situation. Obviously, this approach is much more challenging and labour-intensive, as reflected in
the significantly lower number of publications containing accounts of online experiments (12.25%
of cases). The majority of the remaining studies contain analyses carried out with the use of datasets
from BCI Competitions. This can be recognized as an advantage since the results of various
experiments based on the same datasets are easier to compare; on the other hand, the possibility of
translating the results into BCI operation in real-life settings is significantly limited.
The fact that the relevant analyses are based on an insufficient number of observations is a
drawback from the viewpoint of a practical applicability of the findings related to algorithms of
signal processing in MI BCI. Proportionally, a significant number of experiments are based on
datasets acquired from one subject, and analyses based on groups of more than 10 subjects are a
rarity. It is a well-established fact that subject-specific differences significantly impact the
effectiveness of BCI control (Blankertz et al., 2010; Hammer et al., 2012). According to some
researchers, a considerable number of users, from several to a few dozen per cent, face significant
difficulties trying to effectively convey information in this way (Allison & Neuper, 2010). This
phenomenon is sometimes called BCI illiteracy or aphasia (Kübler & Muller, 2007). Additionally,
there are differences among subjects reflecting the varied pace and effectiveness of learning to
control brain-computer interfaces (McFarland et al., 2005). Differences of this kind occurring in the
population may impact the findings of analyses related to small study groups.
The applicability of reported findings may also be adversely affected by errors occurring at
the stage of data preparation and during the presentation of results. Although MI BCIs are designed
for real-time operation, some studies are based on records manually purged of artefacts (9.81% of
cases) or with frequency bands individually determined for the participants of the study (1.5% of
cases). The procedure of removing fragments of a signal contaminated with non-representative
brain activity e.g. caused by eye or muscle movements is widely used in EEG based studies (Jung et
al., 2000). The method, however, cannot be introduced during the real-time process of brain-
computer interface control with minimum operation by people other than the user. The effectiveness
of the algorithm in recognizing signal patterns in an initially purified recording may lead to inflating
the final result and may fail to represent the actual conditions of BCI operation.
In order to ensure comparability of results showing the effectiveness of signal processing
achieved by various methods, it is necessary to adopt a similar way of reporting results.
Researchers, however, use varied ways of presenting the findings; for instance, they report the mean
error in signal classification, the mean percentage of correctly classified samples or results related
to the most successfully or least effectively classified fragments of data. The matter is additionally
confounded by the fact that MI BCIs differ in term of the number of imagery tasks performed by the
user to control the system (e.g. imagined right hand and left hand movement or imagined movement
of the hand and foot), or they vary in terms of the length of a single trial. Hence, the same result
related to various types of BCIs does not represent the same effectiveness. Due to this, some
researchers have postulated a method of converting the results into the number of bits of
information, which can be transmitted by the relevant BCI per a specified unit of time e.g. minute
(McFarland, Sarnacki & Wolpaw, 2003). This method of converting the results is not commonly
used in the reviewed articles.
Additionally, the articles at times contain insufficient information related to the method of
recording and preparing the data for analyses. Information concerning reference or impedance
during the process of recording constitutes a standard in reporting results of psychophysiological
examinations and conveys the quality and key features of the signal. This principle is not observed
in the case of some publications. Other practices, which seem inadequate, include the presentation
of results exclusively in a graphic form without providing numerical values, as well as
interpretation of differences between algorithms without subjecting them to statistical testing.
Improvements in these standards may contribute to an increased scientific value and higher
applicability of research findings related to data processing in MI BCI.
In order to unify the method of reporting the most important aspects of research, the authors
suggest that articles should contain a chart of particulars (See Appendix 1). The chart would enable
quick assessment of experimental findings and would facilitate access to data important from the
viewpoint of future replication studies or practical application. Moreover, the concise presentation
of data related to equipment and analysis would leave more space in the article for information
about the procedure and mathematical basics underlying the applied signal processing method. The
proposed chart is divided into four parts: 1. Settings, 2. Study group, 3. Procedure and 4. Results.
The first and second parts contain information on the data acquisition method. The third part relates
to details of the tested algorithms. The final section presents the findings. The shape and contents of
the chart can be freely modified. The tool is, in fact, a point of reference for the concise presentation
of information regarding the method of analysing data in a BCI experiment.
Acknowledgments
This work was supported in part by the National Science Centre (Poland) grants No.
2013/08/T/HS6/00472
References
Ahn, M., Jun, S. C. (2015). Performance variation in motor imagery Brain-Computer Interface: A
brief review. Journal of Neuroscience Methods, 243, 103‒110.
Allison, B. Z., Faller, J., Neuper, C. (2012). BCIs that use steady-state visual evoked potentials or
slow cortical potentials. In J. R. Wolpaw, E. W. Wolpaw (Ed.), Brain-Computer Interfaces:
Principles and Practice. (pp. 241-249). New Jork: Oxford University Press.
Allison, B. Z., & Neuper, C. (2010). Could anyone use a BCI? In D. S. Nijholt, & A. Tan, (Ed.),
Brain-computer interfaces. (pp. 35‒54). Berlin: Springer.
Blankertz, B., Sannelli, C., Halder, S., Hammer, E. M., Kübler, A., Müller, K. R., Curio, G., &
Dickhaus, T. (2010). Neurophysiological predictor of SMR-based BCIperformance. Neuroimage,
51(4), 1303‒1309.
Brunner, C., Allison, B. Z., Krusienski, D. J., Kaiser, V., Müller-Putz, G. R., Pfurtscheller, G., &
Neuper, C. (2010). Improved signal processing approaches in an offline simulation of a hybrid
Brain-Computer Interface. Journal of Neuroscience Methods, 188(1), 165‒173.
Brunner, C., Leeb, R., Müller-Putz, G., Schlögl, A., & Pfurtscheller, G. (2008). BCI Competition
2008–Graz data set A. Institute for Knowledge Discovery (Laboratory of Brain-Computer
Interfaces), Graz University of Technology, 136-142.
Friedrich, E. V., McFarland, D. J., Neuper, C., Vaughan, T. M., Brunner, P., & Wolpaw, J. R. (2009).
A scanning protocol for a sensorimotor rhythm-based brain-computer interface. Biological
Psychology, 80(2), 169‒175.
Guger, C., Edlinger, G., Harkam, W., Niedermayer, I., & Pfurtscheller, G. (2003). 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.
Hammer, E. M., Halder, S., Blankertz, B., Sannelli, C., Dickhaus, T., Kleih, S., Müller, K. R., &
Kübler, A. (2012). Psychological predictors of SMR-BCI performance. Biological Psychology,
89(1), 80‒86.
Huggins, J. E., Wolpaw, J. R. (2014). Papers from the Fifth International Brain-Computer Interface
Meeting. Journal of Neural Engineering, 11(3), 030301.
Hwang, H. J., Kim, S., Choi, S., & Im, C. H. (2013). EEG-Based Brain-Computer Interfaces: A
Thorough Literature Survey. International Journal of Human-Computer Interaction, 29(12),
814‒826.
Jung, T. P., Makeig, S., Humphries, C., Lee, T. W., Mckeown, M. J., Iragui, V., & Sejnowski, T. J.
(2000). Removing electroencephalographic artifacts by blind source separation. Psychophysiology,
37(02), 163-178.
Krusienski, D. J., McFarland, D. J., & Principe, J. C., (2012). BCI signal processing: feature
extraction. In J. R. Wolpaw, E. W. Wolpaw (Ed.), Brain-Computer Interfaces: Principles and
Practice. (pp. 123-146). New Jork: Oxford University Press.
Leeb, R., Brunner, C., Müller-Putz, G. R., Schlögl, A., & Pfurtscheller, G. (2008). BCI Competition
2008–Graz data set B. Graz University of Technology, Austria.
Leeb, R., Lee, F., Keinrath, C., Scherer, R., Bischof, H., & Pfurtscheller, G. (2007). Brain-computer
communication: motivation, aim, and impact of exploring a virtual apartment. Neural Systems and
Rehabilitation Engineering, IEEE Transactions on, 15(4), 473‒482.
Lemm, S., Müller, K. R., & Curio, G. (2009). A generalized framework for quantifying the
dynamics of EEG event-related desynchronization. PLoS Computational Biology, 5(8),
e1000453‒e1000453.
Lotte, F., Congedo, M., Lécuyer, A., Lamarche, F., & Arnaldi, B. (2007). A review of classification
algorithms for EEG-based brain-computer interfaces. Journal of Neural Engineering, 4(2), R1-R13.
McFarland, D. J., & Krusienski, D. J. (2012). Brain-computer interfaces: something new under the
sun. In J. R. Wolpaw, E. W. Wolpaw (Ed.), Brain-Computer Interfaces: Principles and Practice. (pp.
147‒163). New Jork: Oxford University Press.
McFarland, D. J., Sarnacki, W. A., & Wolpaw, J. R. (2003). Brain–computer interface (BCI)
operation: optimizing information transfer rates. Biological psychology, 63(3), 237-251.
McFarland, D. J., Sarnacki, W. A., Vaughan, T. M., & Wolpaw, J. R. (2005). Brain-computer
interface (BCI) operation: signal and noise during early training sessions. Clinical Neurophysiology,
116(1), 56‒62.
Müller, K., & Blankertz, B. (2006). Toward noninvasive brain-computer interfaces. Signal
Processing Magazine, IEEE, 23(5), 128‒126.
Neuper, C., Müller, G. R., Kübler, A., Birbaumer, N., & Pfurtscheller, G. (2003). Clinical application
of an EEG-based brain-computer interface: a case study in a patient with severe motor impairment.
Clinical Neurophysiology, 114(3), 399‒409.
Nicolas-Alonso, L. F., & Gomez-Gil, J. (2012). Brain-computer interfaces, a review. Sensors, 12(2),
1211‒1279.
Pfurtscheller, G., Brunner, C., Schlögl, A., & Lopes da Silva, F. H. (2006). Mu rhythm (de)
synchronization and EEG single-trial classification of different motor imagery tasks. Neuroimage,
31(1), 153‒159.
Pfurtscheller, G., & Lopes da Silva, F. H. (1999). Event-related EEG/MEG synchronization and
desynchronization: basic principles. Clinical Neurophysiology, 110(11), 1842‒1857.
Pfurtscheller, G., & McFarland, D. J. (2012). BCIs that use sensorimotor rhythms. In J. R. Wolpaw,
E. W. Wolpaw (Ed.), Brain-Computer Interfaces: Principles and Practice. (pp. 227‒240). New Jork:
Oxford University Press.
Pineda, J. A. (2005). The functional significance of mu rhythms: translating “seeing” and “hearing”
into “doing”. Brain Research Reviews, 50(1), 57‒68.
Portelli, A. J., Daly, I., Spencer, M., & Nasuto, S. J. (2011). Low cost brain-computer interface: first
results. . In: Müller-Putz, G. R., Scherer, R., Billinger, M., Kreilinger, A., Kaiser, V. and Neuper, C.
(Eds.) Proceedings of the 5th International Brain-Computer Interface Conference 2011. Verlag der
Technischen Universität Graz 2011, Graz University of Technology, Austria, pp. 320-323.
Scherer, R., Müller‐ Putz, G. R., & Pfurtscheller, G. (2009). Flexibility and practicality: Graz brain–
computer interface approach. International review of neurobiology, 86, 119-131.
Sellers, E. W., Arbel, Y., & Donchin, E. (2012). BCIs that use P300 event-realated potentials. In J.
R. Wolpaw, E. W. Wolpaw (Ed.), Brain-Computer Interfaces: Principles and Practice. (pp.
215‒226). New Jork: Oxford University Press.
Townsend, G., Graimann, B., & Pfurtscheller, G. (2004). Continuous EEG classification during
motor imagery-simulation of an asynchronous BCI. Neural Systems and Rehabilitation
Engineering, IEEE Transactions on, 12(2), 258-265.
Wolpaw, J. R., & Wolpaw, E. W. (2012). Brain-computer interfaces: something new under the sun.
In J. R. Wolpaw, E. W. Wolpaw (Ed.), Brain-Computer Interfaces: Principles and Practice. (pp.
3‒12). New Jork: Oxford University Press.
Appendix 1
Settings
Recorded Reference Impedance
channels [N]
Analysed Ground Type of electrodes
channels [N]
Amplifier model
Cap model
Healthy [N]
Procedure
Trials [N] Synchronous [Y/N]
Pre-procesing
Feature extraction
Feature selection
Feature
classification
Results
Accuracy (Avg) [%] Accuracy (SD) [%] ITR [bit/s]
Figure 1
Click here to download high resolution image
Figure 2
Click here to download high resolution image
Figure 3
Click here to download high resolution image
Figure 4
Click here to download high resolution image
Figure 5
Click here to download high resolution image
Figure 6
Click here to download high resolution image
Figure 7
Click here to download high resolution image