000334
10.3205/000334
urn:nbn:de:0183-0003340
Research Article
Significance of EEG-electrode combinations while calculating filters with common spatial patterns
Signifikanz von EEG-Elektroden-Kombinationen bei der Berechnung von Filtern mittels Common Spatial Patterns
Wetzel
Wetzel
Dominik
D
Max-Pechstein-Str. 75, 08056 Zwickau, GermanyUniversity of Applied Sciences Zwickau, Faculty of Physical Engineering/Computer Sciences, Zwickau, Germany
dominik.wetzel@fh-zwickau.de
author
Jacobs
Jacobs
Paul-Philipp
PP
University Leipzig, Department of Diagnostic and Interventional Radiology, Leipzig, Germany
paul-philipp.jacobs@medizin.uni-leipzig.de
author
Winkler
Winkler
Dirk
D
University Leipzig, Department of Neurosurgery, Leipzig, Germany
dirk.winkler@medizin.uni-leipzig.de
author
Grunert
Grunert
Ronny
R
University Leipzig, Department of Neurosurgery, Leipzig, Germany
Fraunhofer Institute for Machine Tools and Forming Technology, Fraunhofer Plastics Technology Center Oberlausitz, Zittau, Germany
ronny.grunert@medizin.uni-leipzig.de
author
German Medical Science GMS Publishing House
Düsseldorf
610
algorithms
electroencephalography
statistics
Algorithmen
Elektroenzephalographie
Statistik
Medical Informatics
20230628
20240215
20240925
engl
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License.
Dieser Artikel ist ein Open-Access-Artikel und steht unter den Lizenzbedingungen der Creative Commons Attribution 4.0 License (Namensnennung).
1612-3174
22
GMS German Medical Science
GMS Ger Med Sci
08
Ziel: Common Spatial Patterns (CSP) ist eine gängige Filtertechnik, die für die Vorverarbeitung von Elektroenzephalographie-Signalen (EEG) zur Klassifizierung gedachter Bewegungen verwendet wird. Besonders in Fällen, in denen nur wenige Daten verfügbar sind, ist es wichtig, die Anzahl der Merkmale zu reduzieren. In dieser Forschungsarbeit werden verschiedene Ansätze zur Reduzierung der Anzahl der für die CSP-B<![CDATA[e</PlainText></TextGroup>rechnung verwendeten Elektroden untersucht.</Pgraph><Pgraph><Mark1>Methoden:</Mark1> Frei verfügbare EEG-Datensätze werden für die Evaluierung genutzt. Dazu wird eine einfache Klassifizierungspipeline bestehend aus CSP und linearer Diskriminanzanalyse zur Klassifizierung genutzt. Es wird eine Basislinie unter Nutzung aller Elektroden berechnet und diese mit den Ergebnissen der verschiedenen Ansätze verglichen.</Pgraph><Pgraph><Mark1>Ergebnisse:</Mark1> Der vielversprechendste Ansatz besteht darin, die Fähigkeit von CSP zu nutzen, Informationen über den Ursprung des erstellten Filters zu liefern. Es wird ein Algorithmus vorgeschlagen, der diese Fähigkeit ausnutzt und so wichtige Elektroden aus dem CSP extrahiert.</Pgraph><Pgraph>Die Ergebnisse zeigen, dass die Verwendung von probandenspezifischen Elektrodenpositionen eine positive Auswirkung auf die Genauigkeit der Klassifizierungsaufgabe hat. Außerdem wird gezeigt, dass Elektrodenkombinationen, die in einem Versuch gut funktionieren, nicht unbedingt auch in einem anderen Versuch desselben Probanden gut abschneiden. Zusätzlich zu den mit Hilfe des entwickelten Algorithmus berechneten Kombinationen werden 26 weitere Elektrodenkombinationen vorgeschlagen. Diese können bei der Auswahl gut funktionierender Elektrodenkombinationen berücksichtigt werden. In dieser Arbeit konnten wir damit eine Verbesserung der Genauigkeit von über 10% erzielen.</Pgraph><Pgraph><Mark1>Schlussfolgerung:</Mark1> Die sorgfältige Auswahl der richtigen Elektrodenkombination kann die Genauigkeit bei der Klassifizierung von einer gedachten Bewegung verbessern.</Pgraph></Abstract>
<Abstract language="en" linked="yes"><Pgraph><Mark1>Objective:</Mark1> Common spatial pattern (CSP) is a common filter technique used for pre-processing of electroencephalography (EEG) signals for imaginary movement classification tasks. It is crucial to reduce the amount of features especially in cases where few data is available. Therefore, different approaches to reduce the amount of electrodes used for CSP calculation are tried in this research.</Pgraph><Pgraph><Mark1>Methods:</Mark1> Freely available EEG datasets are used for the evaluation. To evaluate the approaches a simple classification pipeline consisting mainly of the CSP calculation and linear discriminant analysis for classification is used. A baseline over all electrodes is calculated and compared against the results of the approaches.</Pgraph><Pgraph><Mark1>Results:</Mark1> The most promising approach is to use the ability of CSP to provide information about the origin of the created filter. An algorithm that extracts the important electrodes from the CSP utilizing these information is proposed.</Pgraph><Pgraph>The results show that using subject specific electrode positions has a positive impact on accuracy for the classification task. Further, it is shown that good performing electrode combinations in one session are not necessarily good performing electrodes in another session of the same subject. In addition to the combinations calculated using the developed algorithm, 26 additional electrode combinations are proposed. These can be taken into account when selecting well-performing electrode combinations. In this research we could achieve an accuracy improvement of over 10%.</Pgraph><Pgraph><Mark1>Conclusions:</Mark1> Carefully selecting the correct electrode combination can improve accuracy for classifying an imaginary movement task.</Pgraph></Abstract>
<TextBlock linked="yes" name="Introduction">
<MainHeadline>Introduction</MainHeadline><Pgraph>Common spatial pattern (CSP) is an algorithm which decomposes a signal into spatial patterns that are extracted from multiple classes. These are used to calculate spatial filters that maximize the ratio of the variance of one class to another <TextLink reference="1"></TextLink>, <TextLink reference="2"></TextLink>. It can be utilized to decompose an electroencephalography (EEG) signal into components that separate different classes <TextLink reference="1"></TextLink>, <TextLink reference="2"></TextLink>, which in turn can be used to control a brain-computer interface (BCI). BCIs can be used to control different types of hardware and software, such as wheel chairs <TextLink reference="3"></TextLink> or an avatar in a virtual reality environment <TextLink reference="4"></TextLink>. Integration of BCIs into the rehabilitation therapy of stroke patients is an ongoing research subject <TextLink reference="5"></TextLink>, <TextLink reference="6"></TextLink>.</Pgraph><Pgraph>Many variants of CSP exist and are still subject to research <TextLink reference="7"></TextLink>, <TextLink reference="8"></TextLink>, <TextLink reference="9"></TextLink>. As CSP is used to create signal-specific filters; an interesting characteristic of CSP is the ability to reveal information about the origin of the created filter. This can be utilized to select important features or electrodes for a specific subject <TextLink reference="10"></TextLink>, <TextLink reference="11"></TextLink>. Wang et al. calculate the event-related desynchronization and readiness potential using all electrodes and select only those electrodes that have maximum value based on the computed CSPs <TextLink reference="10"></TextLink>. Another approach is omitting channels where the average of the CSP mixing matrix coefficients is less than a certain threshold and evaluate the remaining signals with the help of a neural network <TextLink reference="11"></TextLink>. However, the authors did not state the impact on the electrode reduction and did not compare it with other approaches.</Pgraph><Pgraph>In this study, our objective is to identify improved electrode combinations for classifying an imaginary movement (IM) task through the use of CSP. We aim to compare these combinations with CSP calculations incorporating all available electrodes.</Pgraph></TextBlock>
<TextBlock linked="yes" name="Material and methods">
<MainHeadline>Material and methods</MainHeadline><Pgraph>In this work two different approaches are proposed: </Pgraph><Pgraph><UnorderedList><ListItem level="1">Approach 1: Check every possible combination on one dataset, take the best combinations and try them on the other datasets (to see if some generalization is possible).</ListItem><ListItem level="1">Approach 2: Infer possible combinations for each subject in each dataset using the patterns from the base CSP calculation.</ListItem></UnorderedList></Pgraph><Pgraph>Another approach would be to try every possible combination for each subject in each dataset and choose the best one. Due to the high computational expense this approach is unfeasible and therefore was discarded.</Pgraph><Pgraph>To acquire the needed data we utilize the library “Mother of all BCI Benchmarks” (MOABB) <TextLink reference="12"></TextLink> as it contains interfaces to different freely available BCI datasets. Only datasets that fulfill the following conditions were considered for the evaluation:</Pgraph><Pgraph><UnorderedList><ListItem level="1">All sessions in the dataset are recorded with at least 90 Hz sampling rate.</ListItem><ListItem level="1">The paradigm of the dataset contains at least a left-hand imagination and a right-hand imagination class.</ListItem><ListItem level="1">The electrodes used for data acquisition contain at least the electrodes used in <TextLink reference="13"></TextLink> (see Figure 1 <ImgLink imgNo="1" imgType="figure"/>).</ListItem></UnorderedList></Pgraph><Pgraph>Four datasets meet these conditions: Yi2014 with n=10 <TextLink reference="14"></TextLink>, BCI Competition IV Dataset 2a (BNCI2014) with n=9 and 2 sessions per subject <TextLink reference="15"></TextLink>, Cho2017 with n=49 <TextLink reference="16"></TextLink> and PhysioNet with n=109 <TextLink reference="17"></TextLink>.</Pgraph><Pgraph>We use the EEG data from the datasets with the left-h<TextGroup><PlainText>and imagi</PlainText></TextGroup>nation and right-hand imagination classes (which consist of single finger or full hand movements). S<TextGroup><PlainText>ubsequ</PlainText></TextGroup>ently we utilized a simple pipeline as shown in Figure 2 <ImgLink imgNo="2" imgType="figure"/> containing: 1) a preprocessing step, where the data is reduced to the specified electrodes and bandpass filtered between 5 and 45 Hz, 2) predefined cross-validation splits, 3) calculation of the CSP filter for each split and filtering the signal with it using a CSP implementation from MNE-Python <TextLink reference="18"></TextLink>, 4) a classification step with linear discriminant analysis (LDA) and 5) the combination of the accuracies to a mean accuracy. </Pgraph><Pgraph><Mark1>Baseline:</Mark1> We perform a baseline calculation with the pipeline utilizing all electrodes for each subject from the dataset (BFull) as well as the pipeline with 16 electrodes that are positioned as shown in Figure 1 <ImgLink imgNo="1" imgType="figure"/> (B16).</Pgraph><Pgraph><Mark1>Approach 1:</Mark1> We consider only the 16 electrodes lead<TextGroup><PlainText>ing to an amount</PlainText></TextGroup> of possible combinations of <ImgLink imgNo="1" imgType="inlineFigure"/>. We choose to only use combinations that contain at least 8 electrodes to have enough <TextGroup><PlainText>features for LDA</PlainText></TextGroup> after the CSP filter, which results in <ImgLink imgNo="2" imgType="inlineFigure"/> combinations. Further we decided to use Yi2014 as base dataset as it is fast to calculate for each combination. We run the pipeline for all combinations on each subject of Yi2014. The percentage change between the 16 electrode baseline accuracy (of Yi2014) and each combination’s accuracy is calculated afterwards. Last, we assess whether the expectation µ is significant better over all subjects for each combination using a t-test with the following hypothesis:</Pgraph><Pgraph>(1) H<Subscript>0</Subscript>:µ≤0, H<Subscript>a</Subscript>:µ>0 and α=5%</Pgraph><Pgraph>Significant combinations are used for processing the other datasets. </Pgraph><Pgraph><Mark1>Approach 2:</Mark1> We developed an algorithm to infer more useful electrodes by using the CSP calculation from the baselines. Therefore, we use the pattern p from the first 8 CSP components [the amount of components calculated with CSP is the amount of electrodes, but we only consider the first 8 as the latter are less significant] and select the electrodes that fulfill the following condition:</Pgraph><Pgraph>(2) |e<Subscript>p</Subscript>–<ImgLink imgNo="3" imgType="inlineFigure"/>|>θ·σ(p)</Pgraph><Pgraph>where e<Subscript>p</Subscript> is the value of the electrode in the pattern p, <ImgLink imgNo="3" imgType="inlineFigure"/> is the mean of all electrode values of the pattern, θ is a predefined threshold and σ(p) is the standard deviation of all electrode values of the pattern. In our experiments we set θ=1.5. Figure 3 <ImgLink imgNo="3" imgType="figure"/> shows a visualization of the algorithm.</Pgraph><Pgraph>We calculate the electrodes for each cross validation split. The found electrode combinations (containing more than 2 electrodes) are then used to calculate the accuracy for the subject and the combination with the highest accuracy is furthermore compared against the baselines. The complete pipeline is visualized in Figure 4 <ImgLink imgNo="4" imgType="figure"/>. </Pgraph></TextBlock>
<TextBlock linked="yes" name="Evaluation">
<MainHeadline>Evaluation</MainHeadline><Pgraph>To evaluate our approaches we perform Bayesian A/B tests <TextLink reference="19"></TextLink>, <TextLink reference="20"></TextLink>. We compare different variants against the full electrode baseline (BFull). The used variants are:</Pgraph><Pgraph><UnorderedList><ListItem level="1">The 16 electrode baseline (B16)</ListItem><ListItem level="1">The best electrode combinations from approach 1 (A1)</ListItem><ListItem level="1">The per subject best electrode combinations using our algorithm with 16 electrode base (Algo16)</ListItem><ListItem level="1">The per subject best electrode combinations using our algorithm with full electrode base (AlgoFull)</ListItem><ListItem level="1">The per subject best electrode combinations of the combinations from approach 1 (PSA1)</ListItem><ListItem level="1">The per subject best electrode combinations of the combinations from approach 1 and our algorithm (both bases) (Comb)</ListItem></UnorderedList></Pgraph><Pgraph>We use the beta distribution for our prior and posterior distribution as we have dichotomic data. </Pgraph><Pgraph>Further we assume a priori that all variants perform better than average so we set α=6 and β=4. However the prior has a marginal effect due to the amount of samples provided. Afterwards we perform a Monte-Carlo simulation with n=10,000 for our variants and calculate the fraction of each variant against the full electrode baseline. By calculating the mean of these fractions we get a percentual improvement against BFull.</Pgraph></TextBlock>
<TextBlock linked="yes" name="Results">
<MainHeadline>Results</MainHeadline><Pgraph>After testing all combinations on Yi2014, 26 remain as significant over all 10 subjects from Yi2014. Applying those combinations separately on the other datasets did not lead to improvements of the accuracy compared to BFull (see Table 1 <ImgLink imgNo="1" imgType="table"/> and A1 in Table 2 <ImgLink imgNo="2" imgType="table"/>), whereas using only the best of all 26 combinations for each subject individually increases the accuracy significantly (see PSA1 in Table 2 <ImgLink imgNo="2" imgType="table"/>). Using the proposed algorithm improves accuracy compared to BFull, but it is not as effective as PSA1 (see Algo16 and AlgoFull in Table 2 <ImgLink imgNo="2" imgType="table"/>). Interestingly combining the different approaches leads to an improvement of more than 10% (see Comb in Table 2 <ImgLink imgNo="2" imgType="table"/>).</Pgraph><Pgraph>These findings are also visualized in Figure 5 <ImgLink imgNo="5" imgType="figure"/>, which shows the computed beta distributions of where the samples are taken from for the percentual improvement calculation.</Pgraph><Pgraph>As BNCI2014 contains 2 sessions per subject, we test whether a good performing combination in one sessi<TextGroup><PlainText>on also improves</PlainText></TextGroup> the accuracy for the other session. Figure 6 <ImgLink imgNo="6" imgType="figure"/> shows the result for subject 1 and 3. It is obvious that a combination which works well for one session may not perform equally well in another session. </Pgraph></TextBlock>
<TextBlock linked="yes" name="Discussion">
<MainHeadline>Discussion</MainHeadline><Pgraph>Our results show that using only specific electrodes has the potential to increase the accuracy of an IM task compared to using all electrodes (as also shown in <TextLink reference="21"></TextLink>). Reducing the amount to a predefined number of electrodes (from all to 16) does not lead to better results. We also showed that even for the same subject the calculated electrode combination cannot be transferred to another session. In summary, it can be seen that it is important to calculate a well-performing electrode combination for each session and subject so that the overall amount of electrodes applied to the subject cannot be reduced, but the evaluation speed and accuracy of an IM task can. The proposed algorithm for determining good performing electrode combinations can be used even in a clinical setting, as it is fast to calculate and EEG tasks usually need calibration anyway. Furthermore, we proposed 2<TextGroup><PlainText>6 e</PlainText></TextGroup>lectrode combinations that can be considered additionally for better performance. In summary, sophisticated selection of the appropriate electrode combination can enhance the accuracy of an IM task. In our research we achieved an improvement of over 10%.</Pgraph><Pgraph>To further improve performance the algorithm to detect electrode combinations could be revised or changed as there are subjects where it performs worse. It was assumed that this is due to a low amount of channels. However, there is no correlation between the number of channels and the percentage of change. Another step could be to determine another set of well-performing electrode combinations using the brute-force calculation (approach 1) on another dataset and compare it to the shown results. </Pgraph></TextBlock>
<TextBlock linked="yes" name="Notes">
<MainHeadline>Notes</MainHeadline><SubHeadline>Source code</SubHeadline><Pgraph>The source code for this research is available at <Hyperlink href="https://gitlab.com/domwet/csp-research.git">https://gitlab.com/domwet/csp-research.git</Hyperlink>. </Pgraph><SubHeadline>Datasets</SubHeadline><Pgraph>The underlying datasets can be accessed as follows:</Pgraph><Pgraph><UnorderedList><ListItem level="1">Yi2014 <TextLink reference="14"></TextLink>: <Hyperlink href="https://doi.org/10.7910/DVN/27306">https://doi.org/10.7910/DVN/27306</Hyperlink> <TextLink reference="22"></TextLink></ListItem><ListItem level="1">BNCI2014 <TextLink reference="15"></TextLink>: <Hyperlink href="https://doi.org/10.21227/katb-zv89">https://doi.org/10.21227/katb-zv89</Hyperlink> <TextLink reference="23"></TextLink></ListItem><ListItem level="1">Cho2017 <TextLink reference="16"></TextLink>: <Hyperlink href="https://doi.org/10.5524/100295">https://doi.org/10.5524/100295</Hyperlink> <TextLink reference="24"></TextLink></ListItem><ListItem level="1">PhysioNet <TextLink reference="17"></TextLink>: <Hyperlink href="https://doi.org/10.13026/C28G6P">https://doi.org/10.13026/C28G6P</Hyperlink> <TextLink reference="25"></TextLink> </ListItem></UnorderedList></Pgraph><SubHeadline>Competing interests</SubHeadline><Pgraph>The authors declare that they have no competing interests.</Pgraph></TextBlock>
<References linked="yes">
<Reference refNo="1">
<RefAuthor>Ramoser H</RefAuthor>
<RefAuthor>Müller-Gerking J</RefAuthor>
<RefAuthor>Pfurtscheller G</RefAuthor>
<RefTitle>Optimal spatial filtering of single trial EEG during imagined hand movement</RefTitle>
<RefYear>2000</RefYear>
<RefJournal>IEEE Trans Rehabil Eng</RefJournal>
<RefPage>441-6</RefPage>
<RefTotal>Ramoser H, Müller-Gerking J, Pfurtscheller G. Optimal spatial filtering of single trial EEG during imagined hand movement. IEEE Trans Rehabil Eng. 2000 Dec;8(4):441-6.
DOI: 10.1109/86.895946</RefTotal>
<RefLink>https://doi.org/10.1109/86.895946</RefLink>
</Reference>
<Reference refNo="2">
<RefAuthor>Grosse-Wentrup M</RefAuthor>
<RefAuthor>Buss M</RefAuthor>
<RefTitle>Multiclass common spatial patterns and information theoretic feature extraction</RefTitle>
<RefYear>2008</RefYear>
<RefJournal>IEEE Trans Biomed Eng</RefJournal>
<RefPage>1991-2000</RefPage>
<RefTotal>Grosse-Wentrup M, Buss M. Multiclass common spatial patterns and information theoretic feature extraction. IEEE Trans Biomed Eng. 2008 Aug;55(8):1991-2000.
DOI: 10.1109/TBME.2008.921154</RefTotal>
<RefLink>https://doi.org/10.1109/TBME.2008.921154</RefLink>
</Reference>
<Reference refNo="3">
<RefAuthor>Galán F</RefAuthor>
<RefAuthor>Nuttin M</RefAuthor>
<RefAuthor>Lew E</RefAuthor>
<RefAuthor>Ferrez PW</RefAuthor>
<RefAuthor>Vanacker G</RefAuthor>
<RefAuthor>Philips J</RefAuthor>
<RefAuthor>Millán Jdel R</RefAuthor>
<RefTitle>A brain-actuated wheelchair: asynchronous and non-invasive Brain-computer interfaces for continuous control of robots</RefTitle>
<RefYear>2008</RefYear>
<RefJournal>Clin Neurophysiol</RefJournal>
<RefPage>2159-69</RefPage>
<RefTotal>Galán F, Nuttin M, Lew E, Ferrez PW, Vanacker G, Philips J, Millán Jdel R. A brain-actuated wheelchair: asynchronous and non-invasive Brain-computer interfaces for continuous control of robots. Clin Neurophysiol. 2008 Sep;119(9):2159-69.
DOI: 10.1016/j.clinph.2008.06.001</RefTotal>
<RefLink>https://doi.org/10.1016/j.clinph.2008.06.001</RefLink>
</Reference>
<Reference refNo="4">
<RefAuthor>Vourvopoulos A</RefAuthor>
<RefAuthor>Pardo OM</RefAuthor>
<RefAuthor>Lefebvre S</RefAuthor>
<RefAuthor>Neureither M</RefAuthor>
<RefAuthor>Saldana D</RefAuthor>
<RefAuthor>Jahng E</RefAuthor>
<RefAuthor>Liew SL</RefAuthor>
<RefTitle>Effects of a Brain-Computer Interface With Virtual Reality (VR) Neurofeedback: A Pilot Study in Chronic Stroke Patients</RefTitle>
<RefYear>2019</RefYear>
<RefJournal>Front Hum Neurosci</RefJournal>
<RefPage>210</RefPage>
<RefTotal>Vourvopoulos A, Pardo OM, Lefebvre S, Neureither M, Saldana D, Jahng E, Liew SL. Effects of a Brain-Computer Interface With Virtual Reality (VR) Neurofeedback: A Pilot Study in Chronic Stroke Patients. Front Hum Neurosci. 2019 Jun 19;13:210.
DOI: 10.3389/fnhum.2019.00210</RefTotal>
<RefLink>https://doi.org/10.3389/fnhum.2019.00210</RefLink>
</Reference>
<Reference refNo="5">
<RefAuthor>Soekadar SR</RefAuthor>
<RefAuthor>Birbaumer N</RefAuthor>
<RefAuthor>Slutzky MW</RefAuthor>
<RefAuthor>Cohen LG</RefAuthor>
<RefTitle>Brain-machine interfaces in neurorehabilitation of stroke</RefTitle>
<RefYear>2015</RefYear>
<RefJournal>Neurobiol Dis</RefJournal>
<RefPage>172-9</RefPage>
<RefTotal>Soekadar SR, Birbaumer N, Slutzky MW, Cohen LG. Brain-machine interfaces in neurorehabilitation of stroke. Neurobiol Dis. 2015 Nov;83:172-9. DOI: 10.1016/j.nbd.2014.11.025</RefTotal>
<RefLink>https://doi.org/10.1016/j.nbd.2014.11.025</RefLink>
</Reference>
<Reference refNo="6">
<RefAuthor>Wetzel D</RefAuthor>
<RefAuthor>Winterlich J</RefAuthor>
<RefAuthor>Schaal T</RefAuthor>
<RefTitle>Rehabilitation after Stroke: Practicable Functionalities and Benefiting Patients of Brain- computer Interfaces in Combination with Functional Electrical Stimulation - A Qualitative Interview Study</RefTitle>
<RefYear>2022</RefYear>
<RefJournal>Int J Sci Stud</RefJournal>
<RefPage>54-8</RefPage>
<RefTotal>Wetzel D, Winterlich J, Schaal T. Rehabilitation after Stroke: Practicable Functionalities and Benefiting Patients of Brain- computer Interfaces in Combination with Functional Electrical Stimulation - A Qualitative Interview Study. Int J Sci Stud. 2022;10(4):54-8. Available from: https://www.ijss-sn.com/uploads/2/0/1/5/20153321/12_ijss_jul_22_oa11_-_2022.pdf</RefTotal>
<RefLink>https://www.ijss-sn.com/uploads/2/0/1/5/20153321/12_ijss_jul_22_oa11_-_2022.pdf</RefLink>
</Reference>
<Reference refNo="7">
<RefAuthor>Xygonakis I</RefAuthor>
<RefAuthor>Athanasiou A</RefAuthor>
<RefAuthor>Pandria N</RefAuthor>
<RefAuthor>Kugiumtzis D</RefAuthor>
<RefAuthor>Bamidis PD</RefAuthor>
<RefTitle>Decoding Motor Imagery through Common Spatial Pattern Filters at the EEG Source Space</RefTitle>
<RefYear>2018</RefYear>
<RefJournal>Comput Intell Neurosci</RefJournal>
<RefPage>7957408</RefPage>
<RefTotal>Xygonakis I, Athanasiou A, Pandria N, Kugiumtzis D, Bamidis PD. Decoding Motor Imagery through Common Spatial Pattern Filters at the EEG Source Space. Comput Intell Neurosci. 2018 Aug 1;2018:7957408. DOI: 10.1155/2018/7957408</RefTotal>
<RefLink>https://doi.org/10.1155/2018/7957408</RefLink>
</Reference>
<Reference refNo="8">
<RefAuthor>Zahid SZ</RefAuthor>
<RefAuthor>Aqil M</RefAuthor>
<RefAuthor>Tufail M</RefAuthor>
<RefAuthor>Nazir MS</RefAuthor>
<RefTitle>Online Classification of Multiple Motor Imagery Tasks Using Filter Bank Based Maximum-a-Posteriori Common Spatial Pattern Filters</RefTitle>
<RefYear>2020</RefYear>
<RefJournal>IRBM</RefJournal>
<RefPage>141-50</RefPage>
<RefTotal>Zahid SZ, Aqil M, Tufail M, Nazir MS. Online Classification of Multiple Motor Imagery Tasks Using Filter Bank Based Maximum-a-Posteriori Common Spatial Pattern Filters. IRBM. 2020;41(3):141-50. DOI: 10.1016/J.IRBM.2019.11.002</RefTotal>
<RefLink>https://doi.org/10.1016/J.IRBM.2019.11.002</RefLink>
</Reference>
<Reference refNo="9">
<RefAuthor>Blanco-Diaz CF</RefAuthor>
<RefAuthor>Antelis JM</RefAuthor>
<RefAuthor>Ruiz-Olaya AF</RefAuthor>
<RefTitle>Comparative analysis of spectral and temporal combinations in CSP-based methods for decoding hand motor imagery tasks</RefTitle>
<RefYear>2022</RefYear>
<RefJournal>J Neurosci Methods</RefJournal>
<RefPage>109495</RefPage>
<RefTotal>Blanco-Diaz CF, Antelis JM, Ruiz-Olaya AF. Comparative analysis of spectral and temporal combinations in CSP-based methods for decoding hand motor imagery tasks. J Neurosci Methods. 2022 Apr 1;371:109495.
DOI: 10.1016/j.jneumeth.2022.109495</RefTotal>
<RefLink>https://doi.org/10.1016/j.jneumeth.2022.109495</RefLink>
</Reference>
<Reference refNo="10">
<RefAuthor>Wang Y</RefAuthor>
<RefAuthor>Gao S</RefAuthor>
<RefAuthor>Gao X</RefAuthor>
<RefTitle>Common Spatial Pattern Method for Channel Selelction in Motor Imagery Based Brain-computer Interface</RefTitle>
<RefYear>2005</RefYear>
<RefJournal>Conf Proc IEEE Eng Med Biol Soc</RefJournal>
<RefPage>5392-5</RefPage>
<RefTotal>Wang Y, Gao S, Gao X. Common Spatial Pattern Method for Channel Selelction in Motor Imagery Based Brain-computer Interface. Conf Proc IEEE Eng Med Biol Soc. 2005;2005:5392-5. DOI: 10.1109/IEMBS.2005.1615701</RefTotal>
<RefLink>https://doi.org/10.1109/IEMBS.2005.1615701</RefLink>
</Reference>
<Reference refNo="11">
<RefAuthor>Nouri A</RefAuthor>
<RefAuthor>Ghanbari Z</RefAuthor>
<RefAuthor>Aslani MR</RefAuthor>
<RefAuthor>Moradi MH</RefAuthor>
<RefTitle>A new approach to feature extraction in MI-based BCI systems</RefTitle>
<RefYear>2022</RefYear>
<RefBookTitle>Artificial Intelligence-Based Brain-Computer Interface</RefBookTitle>
<RefPage>75-98</RefPage>
<RefTotal>Nouri A, Ghanbari Z, Aslani MR, Moradi MH. A new approach to feature extraction in MI-based BCI systems. In: Bajaj V, Sinha GR, editors. Artificial Intelligence-Based Brain-Computer Interface. London: Academic Press; 2022. p. 75-98. DOI: 10.1016/B978-0-323-91197-9.00002-3</RefTotal>
<RefLink>https://doi.org/10.1016/B978-0-323-91197-9.00002-3</RefLink>
</Reference>
<Reference refNo="12">
<RefAuthor>Jayaram V</RefAuthor>
<RefAuthor>Barachant A</RefAuthor>
<RefTitle>MOABB: trustworthy algorithm benchmarking for BCIs</RefTitle>
<RefYear>2018</RefYear>
<RefJournal>J Neural Eng</RefJournal>
<RefPage>066011</RefPage>
<RefTotal>Jayaram V, Barachant A. MOABB: trustworthy algorithm benchmarking for BCIs. J Neural Eng. 2018 Dec;15(6):066011. DOI: 10.1088/1741-2552/aadea0</RefTotal>
<RefLink>https://doi.org/10.1088/1741-2552/aadea0</RefLink>
</Reference>
<Reference refNo="13">
<RefAuthor>Wetzel D</RefAuthor>
<RefAuthor>Spahn N</RefAuthor>
<RefAuthor>Heilemann M</RefAuthor>
<RefAuthor>Löffler MM</RefAuthor>
<RefAuthor>Seidel M</RefAuthor>
<RefAuthor>Kolbig S</RefAuthor>
<RefAuthor>Winkler D</RefAuthor>
<RefTitle>Evaluation of electroencephalography analysis methods</RefTitle>
<RefYear>2019</RefYear>
<RefJournal>Annu Int Conf IEEE Eng Med Biol Soc</RefJournal>
<RefPage>767-73</RefPage>
<RefTotal>Wetzel D, Spahn N, Heilemann M, Löffler MM, Seidel M, Kolbig S, Winkler D. Evaluation of electroencephalography analysis methods. Annu Int Conf IEEE Eng Med Biol Soc. 2019 Jul;2019:767-73. DOI: 10.1109/EMBC.2019.8857230</RefTotal>
<RefLink>https://doi.org/10.1109/EMBC.2019.8857230</RefLink>
</Reference>
<Reference refNo="14">
<RefAuthor>Yi W</RefAuthor>
<RefAuthor>Qiu S</RefAuthor>
<RefAuthor>Wang K</RefAuthor>
<RefAuthor>Qi H</RefAuthor>
<RefAuthor>Zhang L</RefAuthor>
<RefAuthor>Zhou P</RefAuthor>
<RefAuthor>He F</RefAuthor>
<RefAuthor>Ming D</RefAuthor>
<RefTitle>Evaluation of EEG oscillatory patterns and cognitive process during simple and compound limb motor imagery</RefTitle>
<RefYear>2014</RefYear>
<RefJournal>PLoS One</RefJournal>
<RefPage>e114853</RefPage>
<RefTotal>Yi W, Qiu S, Wang K, Qi H, Zhang L, Zhou P, He F, Ming D. Evaluation of EEG oscillatory patterns and cognitive process during simple and compound limb motor imagery. PLoS One. 2014 Dec 9;9(12):e114853.
DOI: 10.1371/journal.pone.0114853</RefTotal>
<RefLink>https://doi.org/10.1371/journal.pone.0114853</RefLink>
</Reference>
<Reference refNo="15">
<RefAuthor>Tangermann M</RefAuthor>
<RefAuthor>Müller KR</RefAuthor>
<RefAuthor>Aertsen A</RefAuthor>
<RefAuthor>Birbaumer N</RefAuthor>
<RefAuthor>Braun C</RefAuthor>
<RefAuthor>Brunner C</RefAuthor>
<RefAuthor>Leeb R</RefAuthor>
<RefAuthor>Mehring C</RefAuthor>
<RefAuthor>Miller KJ</RefAuthor>
<RefAuthor>Müller-Putz GR</RefAuthor>
<RefAuthor>Nolte G</RefAuthor>
<RefAuthor>Pfurtscheller G</RefAuthor>
<RefAuthor>Preissl H</RefAuthor>
<RefAuthor>Schalk G</RefAuthor>
<RefAuthor>Schlögl A</RefAuthor>
<RefAuthor>Vidaurre C</RefAuthor>
<RefAuthor>Waldert S</RefAuthor>
<RefAuthor>Blankertz B</RefAuthor>
<RefTitle>Review of the BCI Competition IV</RefTitle>
<RefYear>2012</RefYear>
<RefJournal>Front Neurosci</RefJournal>
<RefPage>55</RefPage>
<RefTotal>Tangermann M, Müller KR, Aertsen A, Birbaumer N, Braun C, Brunner C, Leeb R, Mehring C, Miller KJ, Müller-Putz GR, Nolte G, Pfurtscheller G, Preissl H, Schalk G, Schlögl A, Vidaurre C, Waldert S, Blankertz B. Review of the BCI Competition IV. Front Neurosci. 2012 Jul 13;6:55. DOI: 10.3389/fnins.2012.00055</RefTotal>
<RefLink>https://doi.org/10.3389/fnins.2012.00055</RefLink>
</Reference>
<Reference refNo="16">
<RefAuthor>Cho H</RefAuthor>
<RefAuthor>Ahn M</RefAuthor>
<RefAuthor>Ahn S</RefAuthor>
<RefAuthor>Kwon M</RefAuthor>
<RefAuthor>Jun SC</RefAuthor>
<RefTitle>EEG datasets for motor imagery brain-computer interface</RefTitle>
<RefYear>2017</RefYear>
<RefJournal>Gigascience</RefJournal>
<RefPage>1-8</RefPage>
<RefTotal>Cho H, Ahn M, Ahn S, Kwon M, Jun SC. EEG datasets for motor imagery brain-computer interface. Gigascience. 2017 Jul 1;6(7):1-8. DOI: 10.1093/gigascience/gix034</RefTotal>
<RefLink>https://doi.org/10.1093/gigascience/gix034</RefLink>
</Reference>
<Reference refNo="17">
<RefAuthor>Goldberger AL</RefAuthor>
<RefAuthor>Amaral LA</RefAuthor>
<RefAuthor>Glass L</RefAuthor>
<RefAuthor>Hausdorff JM</RefAuthor>
<RefAuthor>Ivanov PC</RefAuthor>
<RefAuthor>Mark RG</RefAuthor>
<RefAuthor>Mietus JE</RefAuthor>
<RefAuthor>Moody GB</RefAuthor>
<RefAuthor>Peng CK</RefAuthor>
<RefAuthor>Stanley HE</RefAuthor>
<RefTitle>PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals</RefTitle>
<RefYear>2000</RefYear>
<RefJournal>Circulation</RefJournal>
<RefPage>E215-20</RefPage>
<RefTotal>Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation. 2000 Jun 13;101(23):E215-20. DOI: 10.1161/01.cir.101.23.e215</RefTotal>
<RefLink>https://doi.org/10.1161/01.cir.101.23.e215</RefLink>
</Reference>
<Reference refNo="18">
<RefAuthor>Gramfort A</RefAuthor>
<RefAuthor>Luessi M</RefAuthor>
<RefAuthor>Larson E</RefAuthor>
<RefAuthor>Engemann DA</RefAuthor>
<RefAuthor>Strohmeier D</RefAuthor>
<RefAuthor>Brodbeck C</RefAuthor>
<RefAuthor>Parkkonen L</RefAuthor>
<RefAuthor>Hämäläinen MS</RefAuthor>
<RefTitle>MNE software for processing MEG and EEG data</RefTitle>
<RefYear>2014</RefYear>
<RefJournal>Neuroimage</RefJournal>
<RefPage>446-60</RefPage>
<RefTotal>Gramfort A, Luessi M, Larson E, Engemann DA, Strohmeier D, Brodbeck C, Parkkonen L, Hämäläinen MS. MNE software for processing MEG and EEG data. Neuroimage. 2014 Feb 1;86:446-60. DOI: 10.1016/j.neuroimage.2013.10.027</RefTotal>
<RefLink>https://doi.org/10.1016/j.neuroimage.2013.10.027</RefLink>
</Reference>
<Reference refNo="19">
<RefAuthor>Bayes T</RefAuthor>
<RefAuthor>Price R</RefAuthor>
<RefTitle>LII. An essay towards solving a problem in the doctrine of chances. By the late Rev. Mr. Bayes, F. R. S. communicated by Mr. Price, in a letter to John Canton, A. M. F. R. S</RefTitle>
<RefYear>1763</RefYear>
<RefJournal>Philos Trans R Soc Lond</RefJournal>
<RefPage>370-418</RefPage>
<RefTotal>Bayes T, Price R. LII. An essay towards solving a problem in the doctrine of chances. By the late Rev. Mr. Bayes, F. R. S. communicated by Mr. Price, in a letter to John Canton, A. M. F. R. S. Philos Trans R Soc Lond. 1763;53:370-418.
DOI: 10.1098/rstl.1763.0053</RefTotal>
<RefLink>https://doi.org/10.1098/rstl.1763.0053</RefLink>
</Reference>
<Reference refNo="20">
<RefAuthor>Kruschke JK</RefAuthor>
<RefTitle>Bayesian estimation supersedes the t test</RefTitle>
<RefYear>2013</RefYear>
<RefJournal>J Exp Psychol Gen</RefJournal>
<RefPage>573-603</RefPage>
<RefTotal>Kruschke JK. Bayesian estimation supersedes the t test. J Exp Psychol Gen. 2013 May;142(2):573-603.
DOI: 10.1037/a0029146</RefTotal>
<RefLink>https://doi.org/10.1037/a0029146</RefLink>
</Reference>
<Reference refNo="21">
<RefAuthor>Lal TN</RefAuthor>
<RefAuthor>Schröder M</RefAuthor>
<RefAuthor>Hinterberger T</RefAuthor>
<RefAuthor>Weston J</RefAuthor>
<RefAuthor>Bogdan M</RefAuthor>
<RefAuthor>Birbaumer N</RefAuthor>
<RefAuthor>Schölkopf B</RefAuthor>
<RefTitle>Support vector channel selection in BCI</RefTitle>
<RefYear>2004</RefYear>
<RefJournal>IEEE Trans Biomed Eng</RefJournal>
<RefPage>1003-10</RefPage>
<RefTotal>Lal TN, Schröder M, Hinterberger T, Weston J, Bogdan M, Birbaumer N, Schölkopf B. Support vector channel selection in BCI. IEEE Trans Biomed Eng. 2004 Jun;51(6):1003-10.
DOI: 10.1109/TBME.2004.827827</RefTotal>
<RefLink>https://doi.org/10.1109/TBME.2004.827827</RefLink>
</Reference>
<Reference refNo="22">
<RefAuthor>Yi W</RefAuthor>
<RefTitle></RefTitle>
<RefYear>2014</RefYear>
<RefBookTitle>EEG data of simple and compound limb motor imagery. V1</RefBookTitle>
<RefPage></RefPage>
<RefTotal>Yi W. EEG data of simple and compound limb motor imagery. V1. Harvard Dataverse; 2014. DOI: 10.7910/DVN/27306</RefTotal>
<RefLink>https://doi.org/10.7910/DVN/27306</RefLink>
</Reference>
<Reference refNo="23">
<RefAuthor>Brunner C</RefAuthor>
<RefAuthor>Leeb R</RefAuthor>
<RefAuthor>Müller-Putz G</RefAuthor>
<RefTitle></RefTitle>
<RefYear></RefYear>
<RefBookTitle>BCI Competition 2008–Graz data set A</RefBookTitle>
<RefPage></RefPage>
<RefTotal>Brunner C, Leeb R, Müller-Putz G. BCI Competition 2008–Graz data set A. IEEE DataPort. DOI: 10.21227/katb-zv89</RefTotal>
<RefLink>https://doi.org/10.21227/katb-zv89</RefLink>
</Reference>
<Reference refNo="24">
<RefAuthor>Cho H</RefAuthor>
<RefAuthor>Ahn M</RefAuthor>
<RefAuthor>Ahn S</RefAuthor>
<RefAuthor>Kwon M</RefAuthor>
<RefAuthor>Jun SC</RefAuthor>
<RefTitle></RefTitle>
<RefYear>2017</RefYear>
<RefBookTitle>Supporting data for "EEG datasets for motor imagery brain computer interface"</RefBookTitle>
<RefPage></RefPage>
<RefTotal>Cho H, Ahn M, Ahn S, Kwon M, Jun SC. Supporting data for "EEG datasets for motor imagery brain computer interface". GigaScience Database; 2017. DOI: 10.5524/100295</RefTotal>
<RefLink>https://doi.org/10.5524/100295</RefLink>
</Reference>
<Reference refNo="25">
<RefAuthor>Schalk S</RefAuthor>
<RefAuthor>Sarnacki WA</RefAuthor>
<RefAuthor>Joshi A</RefAuthor>
<RefAuthor>McFarland DJ</RefAuthor>
<RefAuthor>Wolpaw JR</RefAuthor>
<RefTitle></RefTitle>
<RefYear>2009</RefYear>
<RefBookTitle>EEG Motor Movement/Imagery Dataset</RefBookTitle>
<RefPage></RefPage>
<RefTotal>Schalk S, Sarnacki WA, Joshi A, McFarland DJ, Wolpaw JR. EEG Motor Movement/Imagery Dataset. PhysioNet; 2009.
DOI: 10.13026/C28G6P</RefTotal>
<RefLink>https://doi.org/10.13026/C28G6P</RefLink>
</Reference>
<Reference refNo="26">
<RefAuthor>Oxley BC</RefAuthor>
<RefTitle></RefTitle>
<RefYear>2017</RefYear>
<RefBookTitle>International 10-20 system for EEG-MCN.svg</RefBookTitle>
<RefPage></RefPage>
<RefTotal>Oxley BC. International 10-20 system for EEG-MCN.svg. Wikimedia Commons; 2017 Jul 11. Available from: https://commons.wikimedia.org/wiki/File:International_10-20_system_for_EEG-MCN.svg</RefTotal>
<RefLink>https://commons.wikimedia.org/wiki/File:International_10-20_system_for_EEG-MCN.svg</RefLink>
</Reference>
</References>
<Media>
<Tables>
<Table format="png">
<MediaNo>1</MediaNo>
<MediaID>1</MediaID>
<Caption><Pgraph><Mark1>Table 1: The combinations that where selected with approach 1 (and their p-values) and how often each combination is considered the best across the different variants (Best on PSA1 and Best on Comb) as well as the performance against the full baseline using each combination for all subjects (Accuracy improvement against BFull). Bottom lines show how frequently combinations from our algorithm were selected on Comb.</Mark1></Pgraph></Caption>
</Table>
<Table format="png">
<MediaNo>2</MediaNo>
<MediaID>2</MediaID>
<Caption><Pgraph><Mark1>Table 2: The performance of the different variants against BFull (Best bold)</Mark1></Pgraph></Caption>
</Table>
<NoOfTables>2</NoOfTables>
</Tables>
<Figures>
<Figure format="png" height="500" width="540">
<MediaNo>1</MediaNo>
<MediaID>1</MediaID>
<Caption><Pgraph><Mark1>Figure 1: The electrode positions used for pre-evaluation</Mark1><LineBreak></LineBreak><Mark1>Figure by Wetzel et al. [13], licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), and adapted from Oxley [26], licensed under CC0 1.0 Universal (https://creativecommons.org/publicdomain/zero/1.0/)</Mark1></Pgraph></Caption>
</Figure>
<Figure format="png" height="134" width="849">
<MediaNo>2</MediaNo>
<MediaID>2</MediaID>
<Caption><Pgraph><Mark1>Figure 2: Single pipeline for calculating the accuracy</Mark1></Pgraph></Caption>
</Figure>
<Figure format="png" height="267" width="1198">
<MediaNo>3</MediaNo>
<MediaID>3</MediaID>
<Caption><Pgraph><Mark1>Figure 3: The 8 components of the baseline of one subject of Yi2014, with the selected electrodes for each component with green border. The extracted electrode combination is Fz, FC1, FCz, CP3, CP1, CPz.</Mark1></Pgraph></Caption>
</Figure>
<Figure format="png" height="336" width="750">
<MediaNo>4</MediaNo>
<MediaID>4</MediaID>
<Caption><Pgraph><Mark1>Figure 4: The complete pipeline for approach 2 to calculate the best electrode combination</Mark1></Pgraph></Caption>
</Figure>
<Figure format="png" height="470" width="734">
<MediaNo>5</MediaNo>
<MediaID>5</MediaID>
<Caption><Pgraph><Mark1>Figure 5: The beta density functions for all variants over all subjects. The black lines (A1) show the beta distributions from the 26 selected combinations. The curve filled in blue is the baseline using all electrodes, which is also the base for comparison.</Mark1></Pgraph></Caption>
</Figure>
<Figure format="png" height="620" width="827">
<MediaNo>6</MediaNo>
<MediaID>6</MediaID>
<Caption><Pgraph><Mark1>Figure 6: The percentage change using a combination compared to full baseline. Upper image is subject 1 and lower image is subject 3 of the BNCI2014 dataset. One pair of blue and orange bars correspond to one electrode combination.</Mark1></Pgraph></Caption>
</Figure>
<NoOfPictures>6</NoOfPictures>
</Figures>
<InlineFigures>
<Figure format="png" height="35" width="133">
<MediaNo>1</MediaNo>
<MediaID>1</MediaID>
<AltText>Equation 1</AltText>
</Figure>
<Figure format="png" height="35" width="133">
<MediaNo>2</MediaNo>
<MediaID>2</MediaID>
<AltText>Equation 2</AltText>
</Figure>
<Figure format="png" height="13" width="7">
<MediaNo>3</MediaNo>
<MediaID>3</MediaID>
<AltText>Equation 3</AltText>
</Figure>
<NoOfPictures>3</NoOfPictures>
</InlineFigures>
<Attachments>
<NoOfAttachments>0</NoOfAttachments>
</Attachments>
</Media>
</OrigData>
</GmsArticle>]]></plaintext></textgroup></pgraph></abstract></origdata></gmsarticle>