Novel spatial filter for SSVEP-based BCI: A generated reference filter approach

Accepted Manuscript Novel spatial filter for SSVEP-based BCI: A generated reference filter approach Abdullah Talha Sözer, Can Bülent Fidan PII:

S0010-4825(18)30049-0

DOI:

10.1016/j.compbiomed.2018.02.019

Reference:

CBM 2905

To appear in:

Computers in Biology and Medicine

Received Date: 29 November 2017 Revised Date:

10 February 2018

Accepted Date: 24 February 2018

Please cite this article as: A.T. Sözer, Can.Bü. Fidan, Novel spatial filter for SSVEP-based BCI: A generated reference filter approach, Computers in Biology and Medicine (2018), doi: 10.1016/ j.compbiomed.2018.02.019. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT ARTICLE TITLE: NOVEL SPATIAL FILTER FOR SSVEP-BASED BCI: A GENERATED REFERENCE FILTER APPROACH

Abdullah Talha Sözera, Can Bülent Fidanb a

b

RI PT

Karabuk University / Electrical and Electronics Engineering Department, Karabuk, 78050, Turkey, [email protected], +903704332021

AC C

EP

TE D

M AN U

Corresponding author: Abdullah Talha Sözer

SC

Karabuk University / Mechatronics Engineering Department, Karabuk, 78050, Turkey, [email protected]

1

ACCEPTED MANUSCRIPT NOVEL SPATIAL FILTER FOR SSVEP-BASED BCI: A GENERATED REFERENCE FILTER APPROACH Abdullah Talha Sözera, Can Bülent Fidanb Electrical and Electronics Engineering Department, Karabuk University, Karabük,

RI PT

a

78050, Turkey b

Mechatronics Engineering Department, Karabuk University, Karabuk, 78050, Turkey

SC

ABSTRACT

Steady state visual evoked potential (SSVEP)-based brain computer interface (BCI)

M AN U

systems can be realised using only one electrode, however, due to the inter-user and inter-trial differences, the handling of multiple electrode is preferred. This raises the problem of evaluating information from multiple electrode signals. To solve this problem, we developed a novel spatial filtering method (Generated Reference Filter) for

TE D

SSVEP-based BCIs. In our method an artificial reference signal is generated by a combination of reference electrode signals. Multiple regression analysis (MRA) was used to determine the optimal weight coefficients for signal combination. The filtered

EP

signal was obtained by subtraction. The method was tested on a SSVEP dataset and compared with minimum energy combination and common reference methods, namely

AC C

the surface Laplacian technique and common average referencing. The newly developed method provided more effective filtering and therefore higher SSVEP detection accuracy was obtained. It was also more robust against subject-to-subject and trial-totrial variability as the artificial reference signal was recalculated for each detection round. No special preparation is required, and the method is easy to implement. These experimental results indicate that the proposed method can be used confidently with SSVEP-based BCI systems.

2

ACCEPTED MANUSCRIPT Keywords: Steady state visual evoked potential (SSVEP); brain computer interface (BCI); spatial filter; multiple regression analysis (MRA) 1

INTRODUCTION

RI PT

Human computer interaction is a promising technique, especially for people with disabilities. The main driver of this technology is the brain computer interface (BCI), which allows the user to communicate with the outside world solely through brain

SC

signals. It allows individuals with severe disabilities such as amyotrophic lateral sclerosis (ALS), multiple sclerosis, brainstem stroke, cerebral palsy to communicate

M AN U

with their environment [1–3].

BCIs based on electroencephalographic (EEG) signals use various electrophysiological mechanisms and events such as the P300 potential and sensorimotor activities [4] and steady state visual evoked potentials (SSVEPs) that are generated in the occipital area

TE D

when an individual looks at a light flashing at a frequency of more than 4 Hz. SSVEPbased BCIs have several advantages [5–7].

SSVEPs have clinical applications in the examination of the eyes and the optic nervous

EP

system. SSVEPs are recognised as a good signal source for BCI systems because the

AC C

neural response is robust, and it does not need to be trained as it occurs naturally when the brain processes visual information. The basic design of SSVEP-based BCI is shown in Figure 1. In the signal acquisition step EEG signals are acquired from the occipital region; generally the SSVEP amplitude is highest at the Oz position [8]. Although it is possible to design a simple and useful BCI based on a single electrode such systems have disadvantages as the location of the maximum SSVEP amplitude varies from person to person [9] and within-person

3

ACCEPTED MANUSCRIPT according to the frequency of the visual stimuli (this is referred to as ‘travelling property of SSVEP’) [10]. This means that use of single-electrode measurement systems can result in incorrect analysis of the EEG signal, which has a negative effect on the

RI PT

system’s performance. Multiple-electrode EEG measurement is preferred as it eliminates this problem.

When multiple electrodes are used for EEG measurement one must decide how to

SC

evaluate the acquired signals. There is variation in spontaneous EEG signals observed at different scalp locations and in the phase of the SSVEP signal. In this situation the

M AN U

spatial filtering carried out as part of the pre-processing step becomes important. Spatial filtering involves combining EEG signals acquired at different scalp locations to facilitate the detection of SSVEP.

Various spatial filtering methods are used in the SSVEP-based BCI studies. Common

TE D

background activities are eliminated in reference methods (Bipolar Combination [11], Common Average Reference (CAR) [12,13], and Surface Laplacian (SL) [14]). Thus, signal-to-noise ratio (SNR) of the SSVEP signal increases. In the minimum energy

EP

combination (MEC) [14] method the channels are combined so as to minimise the energy of EEG signals that are unrelated to the SSVEP. In maximum contrast

AC C

combination (MCC) [14], the energy of the SSVEP signals is increased and the energy of signals unrelated to the SSVEP is decreased. In the Common Spatial Patterns (CSP) technique [15], a non-stimulus condition is used to increase the SNR of the SSVEP. The analytic common spatial patterns (ACSPs) method [16], involves combining spatial filtering and feature extraction techniques; the signals are identified analytically using a Hilbert transform. Thus, the frequency and phase information of the SSVEP signal are clearly expressed. In the double-partial least-squares (DPLS) [17] method the effects of

4

ACCEPTED MANUSCRIPT noise are reduced by using multiple linear regression to detect the latent variables in the EEG signal. Reliable component analysis (RCA) [18] reduces the size of the SSVEP and increases the SNR using the fundamental assumption of evoked responses. The idle

RI PT

state detection problem has been studied using maximum evoked response spatial filtering [19]. Complex sparse spatial filters are used to detect both the frequency and phase of the target visual stimulus in multi-channel EEG and solve the problem of phase

SC

discrepancies between channels [20]. Adaptive weighted average referencing (AdWAR) and adaptive local average referencing (AdCAR) reduce common background noise

M AN U

whilst preserving useful signals. In these methods the adjacent electrodes and their weights are calculated based on information in the electrode of interest [21]. Collaborative representation projection-based spatial filtering increases the accuracy with which error-related potentials (ErrPs) are detected and forms the basis of error-

TE D

aware SSVEP-based BCIs that detect and ignore false interactions [22]. Existing reference methods rely on defining reference channel with fixed weight, but this is not sufficient to eliminate noise from the active channel. When fixed weight is

EP

used, information in electrodes is not taken into account, so using a fixed weight may not eliminate background noise and may actually add noise. In this study we developed

AC C

a novel spatial filtering method, called generated reference (GR) filtering. This method involves generating a reference signal that can be used to eliminate background activity more effectively anew for each detection round by combining reference channels in the optimal way. The weight coefficients for combination are identified using multiple regression analysis (MRA), taking into account the active channel signal. Filtering involves subtracting the reference signal from the active channel signal in order to obtain an SSVEP signal with a high SNR. This signal can be then subjected to various

5

ACCEPTED MANUSCRIPT feature extraction analyses. SSVEP detection accuracy was used to compare the effectiveness of GR, CAR, SL and MEC filters and it was shown that the new method allowed more accurate SSVEP detection than the others.

RI PT

This paper is organised as follows. Section 2 describes the implementation of GR filtering, briefly compares various spatial filters, the technique used to analyse the EEG data and the dataset used. Section 3 presents the results and section 4 highlights the

2

SC

advantages of the GR method and makes suggestions for future studies. MATERIAL AND METHOD

M AN U

This section describes the new method and compares it with other methods of spatial filtering and target identification. The EEG dataset is discussed briefly. 2.1

Generated Reference Filter Approach

Reference methods can be applied successfully to SSVEP-based BCIs. Channels that

TE D

have background activities similar to those of the active channel and lower SSVEP amplitude are suitable for reference. Thus, while SSVEP signal is retained, common background activities are cancelled and detection of the SSVEP signal is facilitated

EP

[9,11,23].

In conclusion, it is claimed that SSVEP signal detection is facilitated by the generation

AC C

of a reference signal which has background components that are very similar to those present in the active channel. The GR method presented in this paper generates a reference signal by linear combination of several reference channels. Formulations of the GR method are given in Equations 1 and 2. is the generated signal,

denotes reference channels and

coefficients.

6

is the active channel, the optimal weight

ACCEPTED MANUSCRIPT −

=

−

(1)

=

−

=

−

…

: :

(2)

RI PT

=

Figure 2 shows a flow diagram of the GR method. In step 1 the components corresponding to the visual stimulus frequencies and its harmonics are extracted from

SC

the channel signals. In the method, the idea is based on similarity of background noise.

=S− "=

=

−

!

−

=

(3)

!

−

[14].

M AN U

Hence any potential SSVEP activity is removed using orthogonal projection

(4)

and " contain only the background components. in Equations 3 and 4, contain sine

= % % … %&

TE D

and cosine signals at the visual stimulus frequencies and its harmonics. (5)

… '() 2+,& ) × ℎ /0' 2+,& ) × ℎ

(6)

EP

%& = '() 2+,& ) × 1 /0' 2+,& ) × 1

In Equation 5, k is equal to the number of visual stimuli. In Equation 6, h is the number

AC C

of harmonics.

Step 2 is the calculation of optimal weight coefficients, . The weight coefficients sought are the values at which the correlation between = 234526 |/033 ,

" |

(7)

7

and " is maximum.

ACCEPTED MANUSCRIPT MRA ( , " ) was used to calculate the optimal weight coefficients. MRA focuses on the relationship between one variable and a set of variables by finding a linear equation. Regression coefficients in the linear equation are the weight coefficients sought. signal is obtained from the weighted reference signals. In

step 4 a background component-free

RI PT

In the third step of the

signal is obtained by subtraction. The

signal

thus obtained can be analysed using existing SSVEP detection methods. SSVEP based BCI Dataset

SC

2.2

The analyses in this paper were performed on EGG SSVEP Dataset II from the Centre

M AN U

for Research and Technology Hellas (CERTH). Eleven healthy volunteers aged between 25 and 39 years (8 men; 3 women) participated in the experiments [24]. EEG signals were recorded at a sampling rate of 250 Hz from 256 channels. A 22” LCD monitor with a 60 Hz refresh rate was used to present visual stimuli. The visual stimuli

TE D

are five boxes flickering simultaneously at frequencies of 6.66, 7.5, 8.57, 10 and 12 Hz (Figure 3) [24].

The flickering stimuli were presented for 5 seconds. Details of the experimental

EP

procedure are given elsewhere [24]. We divided the EEG signals into 2.5-second epochs

2.3

AC C

for the purpose of SSVEP detection. Data Analysis

To confirm the efficacy of the GR filter we carried out SNR analysis of the filtered signal and SSVEP detection. As shown in Figure 4, the signals were filtered using the SL, CAR, MEC or GR filtering and then SSVEP detection was performed using power estimation (PE) or canonical correlation analysis (CCA).

8

ACCEPTED MANUSCRIPT 2.3.1

EEG Signal Filtering

Active and reference channel signals were filtered with a 6-24 bandpass filter, followed by a CAR, SL, MEC or GR spatial filter. The second spatial difference of the surface

9

is the surrounding electrode signal and ) is the number of surrounding

electrodes. −

:;

1 − < )

=

9=

9

(8)

M AN U

=

SC

8, where

RI PT

EEG was computed using the SL technique. The SL method is applied using Equation

The weight coefficients are fixed in SL, but they are variable in GR. The effectiveness of the optimal weight coefficients was examined by comparing GR filtering with SL filtering.

TE D

The CAR filter is widely used in SSVEP-based BCI studies [12,25–27]. In the CAR method the average of all the electrode signals in use is subtracted from the channel of interest. The CAR method is implemented through Equation 9, where is the channel

−

>?

=

1 − < )

AC C

=

EP

of interest and ) is the number of channels in use [28].

9=

9

(9)

The MEC method is a spatial filter which works by combining the multiple electrode signals so as to minimise the power of the nuisance components. The MEC method is a linear transformation procedure in which the appropriate weight coefficients are determined from the eigenvector of EEG signals [5,14]. Equation 10 gives the filtered signal

based on appropriate weights

and multiple electrode signals @.

9

ACCEPTED MANUSCRIPT =@

(10)

Possible SSVEP components are extracted using Equation 11:

where

!

@=@−

@

(11)

RI PT

" = @− @

is the possible SSVEP components given in Equation 6. Then weights for

minimising the power of signal @ are computed. " CD = min C @ " @ " C min D@

(12)

C B

SC

C B

M AN U

" @ " matrix is The eigenvector that corresponds to the minimum eigenvalue of the @ computed by optimising Equation 12. Thus, a signal component is obtained using

with a minimised nuisance

. Additional signal filtering can be done using the next

smallest eigenvalue. Detection of targets in the filtered signal is achieved through PE. A

2.3.2

SNR Analysis

TE D

detailed explanation of MEC can be found elsewhere [5,14].

The SNR was calculated as the ratio of the Fourier power at visual stimulus frequency ,

EP

and the average Fourier power at neighbouring frequencies as shown in Equation 13

E

)×F ,

AC C

[11,29]. , =

∑&=/

F , + I∆, + F , − I∆,



(13)

In Equation 13 , is visual stimulus frequency, F is Fourier power, ) is the number of neighbouring frequencies and ∆, is the difference between neighbouring frequencies. In this study ) = 6 and ∆, = 0.25 Hz.

10

ACCEPTED MANUSCRIPT 2.3.3

Channel Selection

Using EEG data acquired from 256 channels we tested our newly developed method with various active and reference channels. The Oz, O1 and O2 channels had the highest

RI PT

SSVEP amplitude, as shown in Figure 5, and were chosen as the active channel. We want the reference signal to be similar to that of the active channel except with respect to the SSVEP component. If the reference channels are located far from the active

SC

channel they will not carry similar background noise, but if they are next to the active channel then both the background noise and SSVEP components will be similar.

[11,30,31]. 2.3.4

Target Identification

M AN U

Therefore, channels close to the active channel were preferred as reference channel

TE D

Targets were identified using PE and CCA, which are common SSVEP detection methods. The power at stimulus frequency and its harmonics with Equation 14: TV TU

S= &=



(14)

contains the sine and cosine components at stimulus frequency or its

AC C

where

R& S D

EP

1 , = <
, can be estimated

harmonics, EQ is the number of harmonics,

is filtered signal and EP is the number of

signals (more than one filtered signal can be obtained using MEC). The frequency of target visual stimulus F is determined by Equation 15 [32] where W is the number of visual stimuli. F = 234526

, , ) = 1, 2, . . . , W

(15)

11

ACCEPTED MANUSCRIPT The CCA method is a statistical method for detecting a target stimulus using the correlation between two multi-dimensional datasets. In SSVEP-based BCI systems the first dataset is the EEG data and the second dataset consists artificial sines and cosines

sin 2+, Z cos 2+, Z T X = , Z = ] , , … , V ^ : RV RV RV sin 2+EQ , Z cos 2+EQ , Z

RI PT

at stimulus frequency and its harmonics as shown in Equation 16 [33,34].

SC

(16)

, in Equation 16 represents the visual stimulus frequencies. The frequency of the target

M AN U

visual stimulus F satisfies Equation 17. In Equation 17 _ , is the largest canonical correlation coefficient and W is the number of visual stimuli. F = 234526 _ , , ) = 1, 2, . . . , W 3

RESULTS

(17)

3.1

TE D

In this section we present the comparative results using figures and tables. SL vs. GR filtering

EP

It is claimed that filtering is better when the reference signal is more similar to the active channel signal. To prove this, similarity value, SNR analysis results and SSVEP

AC C

detection accuracy were compared with SL. Figure 6 shows the correlations between reference signals (

,

:; )

and the active signal ( ) in 50 samples.

In this case the active channel was Oz and the reference channels were those labelled channel set 1 in Table 1. As the figure shows, the similarity between the active and reference signals was much greater when the GR approach was used. Figure 7 shows SNR values normalised to [0,1] for 50 samples. To show that GR filtering increases SNR we compared the SNR of GR filtered signals, SL filtered signals

12

ACCEPTED MANUSCRIPT and spatially unfiltered signals. In this case the active channel was Oz and the reference channels were those labelled channel set 1 in Table 1. The average SNR for Oz, SL, and GR across the 50 samples was 0.23, 0.29 and 0.37 respectively: GR filtering increased

RI PT

SNR substantially. The effects of SL and GR filtering on SSVEP detection were compared by selecting O1, O2, and Oz as the active channel and varying the reference channels. The channels used

SC

and the resulting SSVEP detection accuracy are shown in Table 1. In the electrode maps in Table 1 active electrodes are shown in blue and reference electrodes in red. Table 1

M AN U

shows that detection accuracies were very low in some subjects. This suggested that the data collected from some subjects may have insufficient quality. Therefore, the average for all subjects was presented as Average1 and the average for all subjects, except 3, 4, 5, and 8, was presented as Average2. The results show that calculating optimal weight

3.2

TE D

coefficients using the GR method increased the accuracy of SSVEP detection. GR vs. CAR and MEC

The O1, O2, Oz, PO3, PO4, PO7, PO8, and POz channels, which are commonly

EP

selected for SSVEP-based BCI studies, were used for CAR and MEC filtering. Channel

AC C

set 1 was used for GR filtering. When the GR method was used, optimal reference signals were calculated for each active channel. Figure 8 shows the SSVEP detection accuracy for each subject and Average1 and Average2 values. As Figure 8 shows, detection was most accurate with the GR spatial filter. Furthermore, the standard deviation of the detection accuracy when subjects S3, S4, S5, and S8 were excluded was 6.3, 8.4 and 10.5 respectively, for GR, CAR and MEC. The GR filter is less sensitive to differences between subjects.

13

ACCEPTED MANUSCRIPT 4

DISCUSSION

The aim of the reference methods is to obtain a signal from which the common background noise has been eliminated. SSVEP-based BCI systems that use reference

RI PT

methods may use fixed reference channel, subject-specific reference channel or dynamic reference channel [9,11,14,30]. However, reference channel has fixed weight in these methods. But fixed weights are not entirely appropriate for reference methods

SC

as they do not take into account the information carried by channels. In other words, these spatial filters are not data-specific. Using fixed weights may result in removal of

introduce additional noise.

M AN U

useful information or insufficient filtering of background noise and may actually

The GR approach proposed here uses dynamic weights. In each decision round a reference signal that will eliminate the background noise is obtained by weighting the

TE D

reference channels with the optimal coefficients. These coefficients are calculated considering the background noise in active channel. Background noise is removed by subtracting the reference signal from the active channel signal and thus the SSVEP

EP

component becomes more prominent.

The SNR analysis and comparison of GR with other reference methods and with MEC,

AC C

which has high noise immunity [35], have shown that the GR method provides better filtering. The GR method is independent of noise level provided that the active channel and reference signal have similar background noise, because filtered signal is obtained by subtraction. One of the main features of BCI systems that needs to be improved for real life applications is their consistency in performance. This inconsistency in performance can manifest itself when used on different subjects, or sometimes even during different trials

14

ACCEPTED MANUSCRIPT on the same subject. The GR method may make BCI systems more consistent as the reference signal is recalculated in each decision round, based on the active channel signal.

RI PT

As shown in Table 1, the detection accuracy obtained from different subjects varied according to the reference channels used, indicating that the method will perform better when subject-specific reference channels are identified in preliminary experiments.

SC

In the GR method the reference signal is generated according to background activities similarity. Using different or additional criteria it may be possible to obtain a reference

5

M AN U

signal which offers even better SSVEP detection. CONCLUSIONS

We have developed a novel spatial filtering method. The method generates a reference signal that reduces the power of the nuisance components in a channel of interest. When

TE D

the reference signal is subtracted from the active channel signal, the SNR of the SSVEP increases and SSVEP detection accuracy is improved. The method offers more accurate detection than SL, CAR or MEC filtering. Furthermore, because the reference signal is

EP

recalculated in each detection round the method is more consistent against inter-subject

AC C

differences. The method is easy to implement as no preparation or calibration stage is required. It is therefore appropriate for SSVEP-based BCI systems. CONFLICTS OF INTEREST None declared.

ACKNOWLEDGMENT This work was supported by the Research Fund of Karabuk University (Project Number: KBÜ-BAP-14/2-DR-011).

15

ACCEPTED MANUSCRIPT REFERENCES [1]

J. Van Kokswijk, M. Van Hulle, Self adaptive BCI as service-oriented information system for patients with communication disabilities, New Trends Inf. Sci. Serv. Sci.

[2]

RI PT

(NISS), 2010 4th Int. Conf. (2010) 264–269. J. Santhosh, M. Bhatia, S. Sahu, S. Anand, Quantitative EEG analysis for assessment to “plan” a task in amyotrophic lateral sclerosis patients: a study of executive functions (planning) in ALS patients, Brain Res Cogn Brain Res. 22 (2004) 59–66.

J.R. Wolpaw, N. Birbaumer, D.J. McFarland, G. Pfurtscheller, T.M. Vaughan, Brain–

M AN U

[3]

SC

doi:10.1016/j.cogbrainres.2004.07.009.

computer interfaces for communication and control, Clin. Neurophysiol. 113 (2002) 767–791. doi:10.1016/S1388-2457(02)00057-3. [4]

A. Bashashati, M. Fatourechi, R.K. Ward, G.E. Birch, A survey of signal processing algorithms in brain–computer interfaces based on electrical brain signals, J. Neural Eng.

[5]

TE D

4 (2007) R32–R57. doi:10.1088/1741-2560/4/2/R03. I. Volosyak, SSVEP-based Bremen–BCI interface—boosting information transfer rates,

[6]

EP

J. Neural Eng. 8 (2011) 36020. doi:10.1088/1741-2560/8/3/036020. J. Tong, D. Zhu, Multi-phase cycle coding for SSVEP based brain-computer interfaces,

[7]

AC C

Biomed. Eng. Online. 14 (2015) 5. doi:10.1186/1475-925X-14-5. A. Combaz, M.M. Van Hulle, Simultaneous detection of P300 and steady-state visually

evoked potentials for hybrid brain-computer interface, PLoS One. 10 (2015) e0121481.

doi:10.1371/journal.pone.0121481.

[8]

Z. Oralhan, M. Tokmakçi, The Effect of Duty Cycle and Brightness Variation of Visual Stimuli on SSVEP in Brain Computer Interface Systems, IETE J. Res. 62 (2016) 795– 803. doi:10.1080/03772063.2016.1176543.

16

ACCEPTED MANUSCRIPT [9]

Z. Wu, S. Su, A dynamic selection method for reference electrode in SSVEP-based BCI, PLoS One. 9 (2014) e104248. doi:10.1371/journal.pone.0104248. G.R. Burkitt, R.B. Silberstein, P.J. Cadusch, A.W. Wood, Steady-state visual evoked potentials

and

travelling

waves,

Clin.

Neurophysiol.

doi:10.1016/S1388-2457(99)00194-7. [11]

(2000)

246–258.

Y. Wang, R. Wang, X. Gao, B. Hong, S. Gao, A practical VEP-based brain-computer interface,

IEEE

Trans.

Neural

Syst.

Rehabil.

Eng.

14

(2006)

234–239.

SC

doi:10.1109/TNSRE.2006.875576.

S.M.T. Muller, T.F. Bastos-Filho, M. Sarcinelli-Filho, Using a SSVEP-BCI to command

M AN U

[12]

111

RI PT

[10]

a robotic wheelchair, in: 2011 IEEE Int. Symp. Ind. Electron., IEEE, 2011: pp. 957–962. doi:10.1109/ISIE.2011.5984288. [13]

M. Nakanishi, Y. Wang, Y.-T. Wang, Y. Mitsukara, T.-P. Jung, A High-Speed Brain Speller Using Steady-State Visual Evoked Potentials, Int. J. Neural Syst. 24 (2014)

[14]

TE D

1450019. doi:10.1142/S0129065714500191. O. Friman, I. Volosyak, A. Graser, Multiple Channel Detection of Steady-State Visual Evoked Potentials for Brain-Computer Interfaces, IEEE Trans. Biomed. Eng. 54 (2007)

S. Parini, L. Maggi, A.C. Turconi, G. Andreoni, A Robust and Self-Paced BCI System

AC C

[15]

EP

742–750. doi:10.1109/TBME.2006.889160.

Based on a Four Class SSVEP Paradigm: Algorithms and Protocols for a High-TransferRate Direct Brain Communication, Comput. Intell. Neurosci. 2009 (2009) 1–11. doi:10.1155/2009/864564.

[16]

O. Falzon, K. Camilleri, J. Muscat, Complex-Valued Spatial Filters for SSVEP-Based BCIs With Phase Coding, IEEE Trans. Biomed. Eng. 59 (2012) 2486–2495. doi:10.1109/TBME.2012.2205246.

[17]

S. Ge, R. Wang, Y. Leng, H. Wang, P. Lin, K. Iramina, A Double-Partial Least-Squares

17

ACCEPTED MANUSCRIPT Model for the Detection of Steady-State Visual Evoked Potentials, IEEE J. Biomed. Heal. Informatics. 21 (2017) 897–903. doi:10.1109/JBHI.2016.2546311. J.P. Dmochowski, A.S. Greaves, A.M. Norcia, Maximally reliable spatial filtering of steady

state

visual

evoked

potentials,

Neuroimage.

doi:10.1016/j.neuroimage.2014.12.078. [19]

109

(2015)

63–72.

RI PT

[18]

D. Zhang, B. Huang, W. Wu, S. Li, An Idle-State Detection Algorithm for SSVEPBased Brain–Computer Interfaces Using a Maximum Evoked Response Spatial Filter,

N. Morikawa, T. Tanaka, Sparse spatial filtering in frequency domain of multi-channel

M AN U

[20]

SC

Int. J. Neural Syst. 25 (2015) 1550030. doi:10.1142/S0129065715500306.

EEG for frequency and phase detection, in: 2016 Asia-Pacific Signal Inf. Process. Assoc. Annu. Summit Conf., IEEE, 2016: pp. 1–7. doi:10.1109/APSIPA.2016.7820779. [21]

D. Delisle-Rodriguez, A. Villa-Parra, T. Bastos-Filho, A. López-Delis, A. Frizera-Neto, S. Krishnan, E. Rocon, Adaptive Spatial Filter Based on Similarity Indices to Preserve

TE D

the Neural Information on EEG Signals during On-Line Processing, Sensors. 17 (2017) 2725. doi:10.3390/s17122725. [22]

F.P. Kalaganis, E. Chatzilari, S. Nikolopoulos, N.A. Laskaris, Y. Kompatsiaris, A

EP

Collaborative Representation Approach to Detecting Error-Related Potentials in SSVEP-

AC C

BCIs, in: Proc. Themat. Work. ACM Multimed. 2017 - Themat. Work. ’17, ACM Press, New York, New York, USA, 2017: pp. 262–270. doi:10.1145/3126686.3129334.

[23]

G. Garcia-Molina, D. Zhu, Optimal spatial filtering for the steady state visual evoked

potential: BCI application, in: 2011 5th Int. IEEE/EMBS Conf. Neural Eng., IEEE, 2011: pp. 156–160. doi:10.1109/NER.2011.5910512.

[24]

Physionet.org, MAMEM Steady State Visually Evoked Potential EEG Database, (2016). https://physionet.org/physiobank/database/mssvepdb/ (accessed August 24, 2016).

[25]

S.N. Carvalho, T.B.S. Costa, L.F.S. Uribe, D.C. Soriano, G.F.G. Yared, L.C. Coradine,

18

ACCEPTED MANUSCRIPT R. Attux, Comparative analysis of strategies for feature extraction and classification in SSVEP

BCIs,

Biomed.

Signal

Process.

Control.

21

(2015)

34–42.

doi:10.1016/j.bspc.2015.05.008. K.B. Ng, R. Cunnington, A.P. Bradley, Enhancing the classification accuracy of Steady-

RI PT

[26]

State Visual Evoked Potential-based Brain-Computer Interface using Component Synchrony Measure, in: 2012 Int. Jt. Conf. Neural Networks, IEEE, 2012: pp. 1–6. doi:10.1109/IJCNN.2012.6252686.

R.M.G. Tello, S.M.T. Muller, T. Bastos-Filho, A. Ferreira, A comparison of techniques

SC

[27]

and technologies for SSVEP classification, in: 5th ISSNIP-IEEE Biosignals Biorobotics Biosignals

Robot.

Better

Safer

Living,

IEEE,

2014:

M AN U

Conf.

pp.

1–6.

doi:10.1109/BRC.2014.6880956. [28]

D.J. McFarland, L.M. McCane, S. V. David, J.R. Wolpaw, Spatial filter selection for EEG-based communication, Electroencephalogr. Clin. Neurophysiol. 103 (1997) 386–

[29]

TE D

394. doi:10.1016/S0013-4694(97)00022-2.

F.-B. Vialatte, M. Maurice, J. Dauwels, A. Cichocki, Steady-state visually evoked potentials: Focus on essential paradigms and future perspectives, Prog. Neurobiol. 90

Y. Wang, X. Gao, B. Hong, C. Jia, S. Gao, Brain-Computer Interfaces Based on Visual

AC C

[30]

EP

(2010) 418–438. doi:10.1016/j.pneurobio.2009.11.005.

Evoked

Potentials,

IEEE

Eng.

Med.

Biol.

Mag.

27

(2008)

64–71.

doi:10.1109/MEMB.2008.923958.

[31]

Yijun Wang, Zhiguang Zhang, Xiaorong Gao, Shangkai Gao, Lead selection for SSVEP-

based brain-computer interface, in: 26th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., IEEE, 2004: pp. 4507–4510. doi:10.1109/IEMBS.2004.1404252. [32]

X. Gao, D. Xu, M. Cheng, S. Gao, A BCI-based environmental controller for the motion-disabled, IEEE Trans. Neural Syst. Rehabil. Eng. 11 (2003) 137–140.

19

ACCEPTED MANUSCRIPT doi:10.1109/TNSRE.2003.814449. [33]

Z. Lin, C. Zhang, W. Wu, X. Gao, Frequency recognition based on canonical correlation analysis for SSVEP-Based BCIs, IEEE Trans. Biomed. Eng. 54 (2007) 1172–1176.

[34]

RI PT

doi:10.1109/TBME.2006.889197. G. Bin, X. Gao, Z. Yan, B. Hong, S. Gao, An online multi-channel SSVEP-based brain– computer interface using a canonical correlation analysis method, J. Neural Eng. 6

M. Abu-Alqumsan, A. Peer, Advancing the detection of steady-state visual evoked potentials

in

brain-computer

interfaces,

J.

Neural

Eng.

M AN U

[35]

SC

(2009) 46002. doi:10.1088/1741-2560/6/4/046002.

13

(2016)

36005.

doi:10.1088/1741-2560/13/3/036005. [36]

Y. Zhang, G. Zhou, J. Jin, X. Wang, A. Cichocki, SSVEP recognition using common feature analysis in brain–computer interface, J. Neurosci. Methods. 244 (2015) 8–15.

AC C

EP

TE D

doi:10.1016/j.jneumeth.2014.03.012.

Figure 1. Basic design of SSVEP-based BCI.

20

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

EP

TE D

Figure 2. Flow diagram of the GR method.

AC C

Figure 3. Stimulus presentation [24].

21

RI PT

ACCEPTED MANUSCRIPT

M AN U

SC

Figure 4. Comparison of spatial filters used in the paper: a) GR vs. SL; b) GR vs. MEC vs. CAR.

AC C

EP

TE D

Figure 5. Scalp map distribution of power at 10 Hz visual stimulus [36].

Figure 6. The relationship of the Oz signal to the

22

and

:;

signals.

SC

RI PT

ACCEPTED MANUSCRIPT

EP

TE D

M AN U

Figure 7. SNR analysis of 50 samples.

AC C

Figure 8. Detection accuracy (%) for the GR, CAR and MEC methods.

23

ACCEPTED MANUSCRIPT

Table 1. Detection accuracy (%) with the GR and SL methods for various channel signals.

O1

Channel Set 5

GR 61,2 77,6 51,2 30,8 29,6 34,4 50,8 29,6 73,6 46,0 70,8 50,5 59,2

O1 GR 70,4 85,2 39,2 32,0 50,4 60,8 76,8 24,4 68,4 61,6 82,4 59,2 72,2

SL 48,0 55,2 22,8 24,4 23,6 56,0 48,8 26,0 51,6 36,4 51,6 40,4 49,7

Channel Set 2

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

SL 38,4 69,6 42,0 27,2 24,8 47,6 38,4 25,2 58,0 55,6 63,6 44,6 53,0

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

SL 37,2 65,2 38,4 26,0 20,0 41,6 36,0 24,0 59,6 42,8 56,0 40,6 48,3

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

SL 37,2 54,4 40,0 29,6 19,6 38,0 30,0 22,8 54,8 39,6 54,0 38,2 44,0

O2

GR 74,0 75,6 37,6 30,8 39,6 46,0 62,0 31,2 74,4 54,8 62,4 53,5 64,2

SL 54,4 63,2 22,4 32,4 30,0 67,2 46,8 22,0 37,2 36,8 57,6 42,7 51,9

Oz

AC C

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

SL 49,2 58,4 22,4 27,6 28,0 56,0 47,2 24,4 47,6 42,0 50,4 41,2 50,1

SL 59,6 73,6 21,2 35,2 36,8 73,2 56,0 22,4 50,8 44,8 69,2 49,3 61,0

GR 85,6 80,8 38,8 36,4 41,6 57,2 67,6 23,6 59,6 53,6 70,4 55,9 67,8

TE D

GR 57,6 78,0 49,2 32,4 30,4 33,6 42,0 26,0 72,8 49,2 63,6 48,6 56,7 O1

SL 38,8 60,0 32,4 24,0 17,2 47,2 34,4 26,8 51,6 37,6 54,4 38,6 46,3

Oz

EP

Channel Set 3

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

SL 39,6 60,8 34,0 25,2 21,6 44,8 36,0 23,2 49,2 40,8 54,4 39,1 46,5

O2 GR 73,6 75,2 40,0 30,0 38,0 59,6 70,4 32,8 77,6 62,0 72,8 57,5 70,2

RI PT

SL 54,0 65,2 23,2 25,2 31,6 64,0 61,2 22,4 60,0 54,8 65,6 47,9 60,7

SC

Channel Set 1

S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 Average1 Average2

Oz GR 55,2 80,8 50,0 29,2 27,6 33,6 42,4 28,0 77,6 58,4 70,0 50,3 59,7

M AN U

O1 SL 42,8 69,6 38,8 24,8 24,0 43,2 40,4 23,6 61,2 53,2 59,2 43,7 52,8

GR 76,8 74,4 34,0 26,0 41,6 52,0 69,2 33,2 75,2 54,8 69,6 55,2 67,4

Channel Set 4

SL 51,2 68,0 28,8 26,4 33,6 63,6 62,8 24,8 53,6 57,2 65,6 48,7 60,3

GR 47,6 82,0 52,0 30,4 32,0 32,0 42,4 28,8 84,0 46,0 73,6 50,1 58,2

SL 45,2 62,4 23,6 24,8 28,4 63,6 56,0 24,8 55,2 47,2 60,8 44,7 55,8

GR 47,6 70,8 58,0 27,2 26,0 51,2 37,6 24,0 67,2 38,0 64,4 46,5 53,8

SL 49,2 65,6 28,8 26,0 27,2 61,6 55,2 25,2 56,8 44,8 60,0 45,5 56,2

O1

O2 SL 54,4 59,6 23,2 30,4 30,0 64,4 46,0 21,6 36,4 28,4 56,4 41,0 49,4

Oz GR 49,2 77,6 51,2 29,2 29,6 38,4 42,0 28,0 79,6 54,0 72,4 50,1 59,0

24

Channel Set 6

SL 59,2 75,2 28,0 42,0 41,6 75,6 57,6 21,6 42,8 47,6 72,4 51,2 61,5

GR 73,6 76,0 41,2 29,6 41,2 56,4 74,0 32,0 76,0 55,2 75,2 57,3 69,5

SL 50,4 68,8 22,8 36,8 36,4 71,6 50,0 23,2 42,0 40,8 63,2 46,0 55,3

GR 72,4 78,4 48,4 27,2 46,4 43,6 68,0 33,2 70,0 47,6 76,4 55,6 65,2

SL 54,0 59,6 21,6 42,4 26,8 72,8 50,4 21,2 46,0 35,2 62,8 44,8 54,4

Oz

O1 GR 84,0 81,6 36,4 33,6 50,8 57,2 72,4 23,6 56,0 48,4 76,4 56,4 68,0

O2 GR 72,0 74,4 38,0 23,2 37,6 58,0 71,6 30,0 76,8 60,4 78,4 56,4 70,2

GR 69,2 83,6 36,4 33,6 45,6 61,6 72,8 22,0 65,6 57,2 84,8 57,5 70,7 O2

Oz

GR 70,8 86,0 44,4 29,2 52,4 62,4 76,8 24,0 63,6 54,0 83,6 58,8 71,0 O2 GR 74,0 76,8 38,8 35,2 43,2 58,8 52,4 24,4 52,8 47,6 74,0 52,5 62,3

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

25

ACCEPTED MANUSCRIPT HIGHLIGHTS •

Artificial reference signal was generated taking into account the active channel signal

to reduce background noise. •

The proposed spatial filter method was compared with other spatial filter methods and

it provided better filtering.

EP

TE D

M AN U

SC

RI PT

The method is easy to implement, and no special preparation is required.

AC C

•