A discriminative subject-specific spatio-spectral filter selection approach for EEG based motor-imagery task classification

Expert Systems With Applications 64 (2016) 375–384

Contents lists available at ScienceDirect

Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa

A discriminative subject-specific spatio-spectral filter selection approach for EEG based motor-imagery task classification A.K. Das, S. Suresh∗, N. Sundararajan School of Computer Science and Engineering, Nanyang Technological University, 50 Nanyang Avenue, 639798, Singapore

a r t i c l e

i n f o

Article history: Received 1 November 2015 Revised 27 April 2016 Accepted 1 August 2016 Available online 3 August 2016 Keywords: Spectral filters Band elimination Spatiospectral filters Non-stationarity Interval type-2 neuro-Fuzzy inference system

a b s t r a c t Motor-imagery tasks generate event related synchronization and de-synchronization in certain subjectspecific frequency ranges of the subject’s ElectroEncephaloGraphy (EEG) signals. The selection of frequency ranges for each subject is important for obtaining better classification accuracy of motor-imagery based Brain Computer Interface (BCI). Further, the spatial filters extracted corresponding to the selected spectral ranges also influence the classification accuracy. In this paper, a subject-specific spatio-spectral filter selection approach using a cognitive fuzzy inference system for classification of the motor-imagery tasks in a two step approach is presented. The cognitive fuzzy inference system (CFIS) employs an evolving interval type-2 system to classify the non-stationary features. The classifier employs a meta-cognitive sequential algorithm to determine both the structure and parameters of the CFIS. In the first step, the CFIS classifier is used to find the desired spectral filters by eliminating those frequency bands that do not affect the classification performance. In the second step, CFIS is used to eliminate those spatial filters which do not affect the performance. The performance of CFIS based spatio-spectral scheme has been evaluated using two publicly available BCI competition data sets and compared with other existing algorithms like FBCSP, DCSP and BSSFO. The results indicate that the proposed approach outperforms the CSP method by approximately 15–18% and other algorithms like FBCSP, DCSP by 8–10%. Compared to a recently proposed algorithm BSSFO, it achieves an improvement of 2%, but is simpler in comparison to BSSFO. The main impact of the work is its ability to handle non-stationarity using interval type-2 sets and provide good classification performance. In general, the proposed CFIS algorithm can be applied in the field of expert and intelligent systems where it is necessary to deal with non-stationary signals. © 2016 Elsevier Ltd. All rights reserved.

1. Introduction A BCI aims to provide a direct communication pathway between the brain and any external device (Curran & Stokes, 2003; Vallabhaneni, Wang, & He, 2005). In recent years, most of the BCI research has been in the area of mental imagination of motor tasks (called motor-imagery based BCI), where the imagination of movements is translated into user commands without performing any muscle activity. Motor-imagery based BCI helps the users with motor disabilities to communicate and control external devices so that they can achieve greater freedom (Birbaumer, 2006; Pfurtscheller & Neuper, 2001; Wolpaw, Birbaumer, McFarland, Pfurtscheller, & Vaughan, 2002). Motor-imagery based BCI has also been used in the field of neuro-rehabilitation in stroke pa-

∗ Corresponding author at: No. 02b-67, Block N4, School of Computer Science and Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore 639798. Fax:+65 6792 6559. E-mail addresses: [email protected] (A.K. Das), [email protected] (S. Suresh), [email protected] (N. Sundararajan).

http://dx.doi.org/10.1016/j.eswa.2016.08.007 0957-4174/© 2016 Elsevier Ltd. All rights reserved.

tients (Ang & Guan, 2015), device control (Doud, Lucas, Pisansky, & He, 2011; LaFleur et al., 2013), computer games and entertainment (Ahn, Lee, Choi, & Jun, 2014; Nijholt, Bos, & Reuderink, 2009) and locomotion (Mak & Wolpaw, 2009). During mental imagination of different motor tasks, event related de-synchronization/synchronization patterns are generated in the EEG. Here, Event-Related De-synchronization (ERD) pattern (short-lasting attenuation or blocking of rhythms) occurs in the (μ (8–12) Hz and β (14–30) Hz) frequency bands. The opposite phenomenon of ERD is called Event-Related Synchronization (ERS) pattern, where an increase in spectral peak occurs in μ and β bands. These patterns are useful phenomena in motorimagery tasks for accurate detection of the task. For more details on ERD, and ERS, refer to INOUE, Mori, Sugioka, Pfurtscheller, and Kumamaru (2004); Zhang, Wang, Gao, Hong, and Gao (2007). These ERD/ERS patterns occur in different frequency ranges for different subjects (Ang, Chin, Zhang, & Guan, 2008). Based on these subject-specific patterns, different spatial locations of the brain have different activations which makes feature extraction from a specific location challenging. Hence, there is a need to find the

376

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

subject-specific frequency ranges (using band-pass filters referred to as spectral filters) followed by spatial filters for accurate motorimagery task classification. Together, this process is referred to as spatio-spectral filter selection. The accurate detection of spatiospectral filters for a given subject is challenging due to the presence of non-stationarity in EEG signals (Wolpaw et al., 20 0 0). The research works in the literature for solving the subjectspecific spatio-spectral filter selection can be grouped into three major categories based on the nature of EEG signal processing. In the first category, time-delayed signals are embedded into raw EEG signals and the Common Spatial Pattern (CSP) algorithm (Blankertz et al., 2006; Ramoser, Muller-Gerking, & Pfurtscheller, 20 0 0) is used to optimize both the spatial and spectral finite impulse response filters (Dornhege et al., 2006; Lemm, Blankertz, Curio, & Muller, 2005). In the Common Spatio-Spectral Pattern (CSSP) method (Lemm et al., 2005), the CSP algorithm has been extended by appending a time-delayed EEG signal as the signal from additional channels resulting in additional finite impulse response filters at each electrode. However, the CSSP algorithm is not flexible due to the use of a single time-delay filter (Dornhege et al., 2006). Later, Dornhege et al. (2006) extended the CSSP algorithm by allowing a larger time-delay. However, the selection of time-delay parameter is computationally expensive (Tomioka et al., 2006). In the second category, the spectral filters are optimized in the frequency domain (Tomioka et al., 2006; Wu, Gao, Hong, & Gao, 2008). Here, an iterative procedure is employed in which the spatial filters are optimized using the CSP algorithm followed by the optimization of the spectral weighting coefficients in the frequency domain. However, the optimization in the frequency domain is difficult as EDR/ERS patterns occur in specific frequency bands which makes the tuning difficult. In the last category, the frequency range is discretized into multiple smaller bands using a filter bank. Here, researchers have focused on finding the subject-specific spatiospectral filters by constructing a filter bank and applying the CSP algorithm on each of the frequency bands. In Novi, Guan, Dat, and Xue (2007); Thomas, Guan, Lau, Vinod, and Ang (2009), a broad frequency range was divided into sub-bands and a reduced set of filters were selected. In Novi et al. (2007), a score for each subband was calculated using the Linear Discriminant Analysis (LDA) and the top bands with higher scores were selected. In the Discriminative Filter Bank CSP (DCSP) paper (Thomas et al., 2009), the Fisher ratio of signals at each band output was used to select the optimum bands. These algorithms only find the optimal spectral filters that look at the frequencies and they do not consider the spatial aspects. In order to address the spatial aspects, several algorithms were proposed with an aim to optimize both the spectral and spatial filters (Ang et al., 2008; Meng, Yao, Sheng, Zhang, & Zhu, 2015; Suk & Lee, 2013; Zhang, Chin, Ang, Guan, & Wang, 2011). These algorithms employed a mutual information based approach to find the spatio-spectral filters. Filter Bank CSP (FBCSP) (Ang et al., 2008) employed a mutual information based filter selection approach to select a subset of filters generated by a filter bank, followed by the band selection. In the Optimal SpatioSpectral Filter Network (OSSFN) work (Zhang et al., 2011), spatial filter optimization was performed at frequency band levels followed by the selection of that band with the maximal mutual information, while in Meng et al. (2015), the spatio-spectral filters were optimized simultaneously. All these algorithms used a fixed filter bank. Recently, Bayesian Spatio-Spectral Filter Optimization (BSSFO) (Suk & Lee, 2013) was proposed which assigned different weights to different frequency bands. It employed a wrapper around method based on Bayesian framework for optimizing the spatio-spectral filters. It combined the outputs from multiple classifiers using the weights of different frequency bands. The use of multiple classifiers makes this approach complex in nature. It should be pointed out here that all the above methods use the CSP

algorithm as the base to generate the features for the classification of motor-imagery tasks. However, the features generated using the CSP algorithm are inherently non-stationary in nature (Lotte & Guan, 2011; Wolpaw et al., 20 0 0). The earlier works either employ LDA or Support Vector Machine (SVM) as the classifier which are unable to handle non-stationary EEG signals (Nguyen, Khosravi, Creighton, & Nahavandi, 2015) effectively and using a CSP based approach with these classifiers might lead to a lower classification performance. This is the main motivation behind this paper that develops a new approach that is capable of finding the optimal subject-specific spatio-spectral filters in the presence of nonstationarity in EEG signals and also provide good classification performance. This paper presents such an approach using a Cognitive Fuzzy Inference System (CFIS) to find the subject-specific spatio-spectral filters for motor-imagery based BCI. CFIS employs an interval type2 fuzzy system capable of handling noisy and non-stationary EEG signals effectively. Some of early works on using type-2 systems in BCI is presented in Alhaddad, Mohammed, Kamel, and Hagras (2015); Herman, Prasad, and McGinnity (20 05); 20 08); Nguyen et al. (2015). In this paper, the interval type-2 neuro-fuzzy inference system presented in Subramanian, Das, Suresh, and Ramasamy (2014) is used to find the optimal spatio-spectral filters. The spatio-spectral filter selection is cast into the problem of band selection in a filter bank followed by spatial filter selection in the selected bands. It has a two step approach wherein the first step spectral filter selection is performed followed by the second step where the spatial filter selection in done. This approach is shown in Fig. 1. In the first step, CFIS is used to eliminate unnecessary bands from a filter bank to arrive at a set of optimal spectral filters, referred to as Spectral-CFIS (S-CFIS) method. Here, frequency bands are iteratively eliminated such that the overall classification performance is better. In this step, the raw EEG signals are bandpassed using six equally partitioned bands of 4 Hz each as given in Ang et al. (2008); Thomas et al. (2009), from 8 to 32 Hz covering the μ and β rhythms followed by the CSP algorithm based feature extraction for each of the bands. The non-stationary nature of features generated by the CSP algorithm is handled as uncertainty using the CFIS classifier. In the second step, CFIS takes those CSP filters from the selected bands in the first step and eliminates unnecessary spatial filters to obtain an optimal set of spatial filters. Similar approach has been employed to remove the spatial filters from the selected spectral bands. At the end of the second step, the proposed method is able to identify the desired spatio-spectral filters for a given subject. This procedure is referred to as SpatioSpectral-CFIS (SS-CFIS) method. A preliminary version of this work was presented in Das, Subramanian, and Suresh (2015), where only the subject-specific bands were selected using the McIT2FIS algorithm (Subramanian et al., 2014) for only one task and spatiospectral aspect was not considered. Performances of S-CFIS and SS-CFIS are evaluated using two data sets viz. the data set-IIa from the BCI competition IV (Naeem, Brunner, Leeb, Graimann, & Pfurtscheller, 2006) (six classification tasks) and the data set-IIIa (Schlögl, Lee, Bischof, & Pfurtscheller, 2005) from the BCI competition III (Blankertz et al., 2006). The evaluation has been carried out as described below: First, the performance of S-CFIS algorithm is evaluated in selecting the spectral filters for the classification tasks followed by the evaluation of SS-CFIS algorithm in selection of spatio-spectral filters. The results show the existence of common frequency bands ((8–16) Hz and (20–32) Hz) for all subjects across all the classification tasks. It is also observed that the order of spatial and spectral filtering has little effect on the performance, however, ‘spectral followed by spatial filtering’ is computationally simple. The performance of S-CFIS and SS-CFIS is compared with other existing algorithms like CSSP, FBCSP, DCSP, OSSFN (OSSFNwFBCSP-OSSFN utiliz-

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

377

Fig. 1. Block diagram of two step spatio-spectral filtering using CFIS classifier.

ing FBCSP, OSSFNwFBCSP-OSSFN utilizing DCSP) and BSSFO. It gives an improvement of 15–18% on an average over the CSP algorithm and 2% over the next best performing algorithm. Rest of the paper is organized as follows. In Section 2, the spatio-spectral filter selection using the CFIS algorithm is presented. Section 3 presents the performance evaluation of the proposed approach and Section 4 presents the conclusions.

In this section, the two step approach to find the subjectspecific spatio-spectral filters using the CFIS classifier is described. The idea is to use the CFIS algorithm to find the desired spatiospectral filters in a two step approach. In the first step, the spectral filters are found by eliminating those bands from a filter bank that do not affect the CFIS classification accuracy. In the second step, CFIS is used to find the optimal spatial filters by eliminating the irrelevant spatial filters obtained from those frequency bands that were selected in the first step. In this section, first, the generation of features using a filter bank is described. Next, the CFIS classifier for motor-imagery task classification is highlighted followed by the selection procedure for both spectral and spatial filters. 2.1.1. Feature generation from EEG using a filter bank The filter bank comprises of six frequency bands of 4 Hz each (Ang et al., 2008) in the range of 8–32 Hz (8–12 Hz, 12–16 Hz and so on) covering the μ and β rhythms. It has been shown in Ang, Chin, Wang, Guan, and Zhang (2012); Thomas et al. (2009), that the frequency bands of width 4 Hz yield a stable frequency response and provide the best results in comparison to other bands of varying widths in the range of 2–6 Hz. The EEG signals generated from the motor-imagery tasks are processed using each of these frequency bands. The band passed signals are then spatially filtered using the CSP algorithm to generate the log-variance features for each frequency band (Fukunaga, 1990; Ramoser et al., 20 0 0). The log-variance feature computation using the CSP algorithm for two-class problem is described below. Let us assume that for a particular frequency band, the band passed EEG data of the p p-th trial and the l-th class label is denoted as El ∈ M×S , where M denotes the total number of channels and S is the number of p recorded samples in each channel. The covariance matrix Cl of the EEG data of the p-th trial is given by T T

trace(Elp Elp )

; l = 1, 2.

(2)

The eigen value decomposition of Cc is given by: (3)

where, Gc denotes the matrix of normalized eigenvectors with corresponding eigenvalues χ . Then whitening transformation P matrix is given by:

2.1. Spatio-spectral filter selection using CFIS

Elp Elp

Cc = C1 + C2 Cc = Gc χ Gc T

2. Materials and methods

Clp =

where, T is the transpose operator and trace(.) represents the sum of the diagonals. The covariance matrices of the respective classes, C1 and C2 are calculated by computing the mean over all the trials. The composite covariance matrix Cc is then given by:

(1)

P=



χ −1 Gc T

(4)

and decorrelates the composite covariance matrix Cc as given below:

PCc PT = P(C1 + C2 )P = I T



C1 = PC1 PT ,

(5)



C2 = PC2 PT

(6)

The common eigenvectors U for matrices, as: 

C1 = Uβ1 UT ,



C2 = Uβ2 UT ,

 C1

β1 + β2 = I

and

 C2

are calculated (7)

where, I is the identity matrix. Here, the eigenvector with the   largest eigenvalue for C1 has the smallest eigenvalue for C2 and vice-versa. For further details on the CSP algorithm, refer to Fukunaga (1990); Ramoser et al. (20 0 0). The projection matrix V = (UT P ) are called spatial fil ters. From V, the three most dominant spatial filters V = [v1 v2 v3 v2m−2 v2m−1 v2m ] for task 1 and task 2 are chosen, respectively. The projected signal from the selected spatial filters can be used to map a trial E is given by: 

Z=V ×E

(8)

In this paper, the log-variances of the signals are computed as features and are given by:

bj = log(var (Z ) ); j = 1, 2, . . . , 6

(9)

where, bj = [b j1 , . . . , b j6 ]T ∈ 6 is a 6-dimensional feature vector in the j-th band. These spatial features are the log-variances of the projected EEG signal from the selected spatial filters. The spatial filters provide a linear combination of raw EEG signals measured at different locations of the brain. These filters are computed such that they project the EEG signals into an orthogonal space. Eqs. (8) and (9) are repeated for every trial to compute the logvariance features.

378

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

Fig. 2. Architecture and learning scheme of CFIS.

2.1.2. Cognitive fuzzy inference system motor-imagery task classifier In this section, the description of the CFIS architecture and its learning algorithm (Subramanian et al., 2014) is presented. Let the training data generated by the j-th band be {(b1j , c1 ), . . . , (btj , ct ), . . . , (bNj , cN )}, where, btj = [btj1 , . . . , btj6 ]T ∈ 6 is the 6-dimensional input vector of the tth sample in the j-th band and ct ∈ [1, 2] is the corresponding class label. The binary class label yt is given as:



t

y =

1, −1,

I f ct == 1 I f ct == 2

(10)

The objective of the CFIS classifier is to approximate the underlying functional relationship (btj → yt ) between the input (btj ) and output

(yt ).



(ut − μ )2 φik = exp − i 2 ik 2 σk

j = 1, 2, . . . , 6; i = 1, 2, . . . , 6.

(11)

The band dependence j is dropped hereafter for convenience. Layer 2- Fuzzification layer: Each node in this layer handles the non-stationary nature of the input features as an uncertainty

 ≡φ



 μik , σk , uti ,

(12)

where, μik ∈ [1 μik , 2 μik ] and σ k ∈ [1 σ k , 2 σ k ] are the left and right limits of the center and width of the kth rule. The footprint of the uncertainty of this membership function is represented as a bounded interval in terms of an upper membership function φ up and a lower membership function φ lo , given by:

φikup,t

2.1.2.1. Cognitive fuzzy inference system architecture. The architecture of CFIS shown in Fig. 2 is a five layered Interval type-2 neurofuzzy inference system used to predict the class label for the current sample. Let us assume that the network has grown K rules after processing t − 1 data samples. The detailed description of each layer when presenting the tth sample in jth band is as follows: Layer 1- Input layer: The input layer consists of six nodes, one each for the six frequency bands. The output of the ith node of the jth band of the input layer is given by:

utji = btji ;

bound by employing an Interval type-2 fuzzy variable. The membership of the ith feature with the kth rule is given by:

φiklo,t

⎧   ⎨φ 1 μik , 1 σk , uti = 1 ⎩φ 2 μ , 2 σ , ut  ik k i

1

uti <1 μik μik ≤ uti ≤2 μik uti >2 μik

⎧ 1 2   ⎪ ⎨ φ 2 μik , 2 σk , uti uti ≤ ( μik + μik ) 2 = 1 2 ⎪ ⎩φ 1 μik , 1 σk , uT  ut > ( μik + μik ) i i

(13)

(14)

2

The output of each node in the fuzzification layer is represented up,t lo,t by the interval ik = [φik , φik ]. Layer 3- Firing layer: The nodes in this layer represent the upup,t per Fk and lower Fklo,t membership of a rule given by:



Fkt = Fklo,t , Fkup,t ; k = 1, . . . , K.

(15)

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

and the output matrix d ∈ RK × 1 is given by:

where 6 

Fklo,t =

φiklo,t

and Fkup,t =

i=1

6 

φikup,t ; k = 1, . . . , K.

(16)

i=1

Fkt = α Fklo,t + (1 − α )Fkup,t ; k = 1, . . . , K.

(17)

where α is the weighted measure of uncertainty. In this study, α is chosen as 0.5. Layer 5-Output layer: This layer calculates the predicted output of network given by:

K

k=1

t

yˆ =

K

wk Fkt

(18)

t p=1 Fp

where, wk is the output weight connecting the k-th rule with the output node. The predicted class label corresponding to yˆt is given by cˆt , and is defined as:



1, If yˆt >= 0 2, otherwise

(19)

Next, the meta-cognitive learning algorithm used in the above fuzzy inference system is presented. The objective of this learning algorithm is to find the rules and the associated parameters to approximate the functional relationship between the inputs and the outputs. 2.1.2.2. Cognitive fuzzy inference system learning algorithm. The CFIS learning algorithm employs the Projection Based Learning algorithm (PBL) (Babu & Suresh, 2013) to calculate the weights of the network. The PBL algorithm finds the optimal weights of the network by minimizing the total energy of the error function. The total energy function, J(w) is given by:

J (w ) =

1 2

t 

Ji

(20)

i=1

where,



Jt =

yt −

K

k=1

0,

wk F¯kt

2

,

if yt yˆt < 1 otherwise

(21)

here

Ft F¯kt = K k

t p=1 Fp

dp =

t 

F¯pi yi ; p = 1, . . . , K

(26)

i=1

Layer 4- Type reduction layer: Each node in this layer converts the Interval type-1 fuzzy set to a fuzzy number for a particular rule using the Nie-Tan type-reduction approach (Nie & Tan, 2008). This approach is a closed form approximation of the well-known Karnik–Mendel algorithm (Mendel & Liu, 2012). The output of each node Fkt is given by:

cˆt =

379

; k = 1, . . . , K

(22)

The aim of the PBL algorithm is to compute the optimal weights (w∗ ) such that total energy of the network is minimized.

w∗ = arg min J (w )

(23)

w

When a new sample (bt , ct ) is presented to the network, the CFIS learning algorithm monitors the knowledge content of the sample using the prediction error Et (Subramanian et al., 2014) and the spherical potential ψ t (Subramanian & Suresh, 2012) with respect to the knowledge already present in the network. It then decides whether the sample can be used for a parameter update or an addition of a rule or it can be discarded without learning or it can be considered as a reserve sample for learning at later point of time. For further details, refer to Subramanian et al. (2014). The working mechanism of the parameter update/rule growing, deletion of sample and sample reserve are briefly described below for completeness: Rule growing criterion: A new rule is added to the network if the prediction error is high and the spherical potential is below the ‘novelty’ threshold. The rule addition condition is given by:

ct = cˆt & E t > Ea &

ψ t < Es

(27)

where, Ea and Es are the ‘add’ and ‘novelty’ thresholds. Here, Ea is self-regulated as in Subramanian et al. (2014). In this work, Ea and Es are determined using a grid search with five-fold cross validation method. The suggested ranges for these thresholds are [1.01, 1.20] and [0.01, 0.60] respectively. The center μlK+1 for the newly added (K + 1 )th rule is initialized on the basis of the current sample as:



μlK+1 = bt × 0.9, bt × 1.1

(28)

where, the superscript l denotes the class of rule. The width of the newly added rule is assigned by considering the distance from the nearest rule in inter/intra classes (different/same class as the sample) rules as given in Subramanian et al. (2014). When the current sample (bt ) is closer to the rule belonging to the same class, the width is assigned as:

σK+1 = κ × dS

(29)

where, dS is the distance to the nearest rule in the same class and κ determines the overlap between the newly added rule and the nearest rule in the same class. It is determined using a grid-search with five-fold cross validation method in the range [0.5, 0.9]. When the current sample (bt ) is nearer to the rule belonging to a different class than the sample, the width is assigned as:

σK+1 = η × dI

(30)

where, dI is the distance to the nearest rule in inter class and η controls the overlap with the nearest rule in the inter class. Similar to κ , it is also determined using a grid-search with five-fold cross validation method in range [0.1, 0.4]. When a new rule is added based on the tth sample, the projection matrix A and the output matrix d are updated as given in Subramanian et al. (2014). The output weights are re-estimated as:

w∗ = A−1 d

(31)

The optimal output weights (w∗ ) are obtained by equating the first order partial derivatives of J(w) with respect to the output weights to zero and are given by:

Parameter update criterion: The output weights of the network are updated if the predicted class label is correct and the absolute hinge error for the current sample is greater than the ‘update’ threshold. The update criterion is given by:

w∗ = A−1 d

ct == cˆt & E t > El

where, the projection matrix

akp =

t  i=1

(24) A ∈ RK × K

is given by:

F¯ki F¯pi , k = 1, . . . , K; p = 1, . . . , K

(25)

(32)

Here, El is the self-adaptive parameter ‘update’ threshold and is adapted as in Subramanian et al. (2014). It is chosen using grid search with five-fold cross validation method in the range [0.04, 0.2].

380

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

CSP v11..v16

8-12 Hz

Logvariance features

B1 Raw EEG Signal Data Set

E

X

B

Spectral Filtered Signal

28-32 Hz

Spatio-Spectral Filtered Signal

v61..v66

B6

CFIS

Classification Accuracy

Classifier

Logvariance features

Band Pass Filters

Spatial Filter Selection Decision Block

Spectral filters

Band Selection Decision Block

Spatio-Spectral filters

Fig. 3. Schematic diagram of spatio-spectral filtering using CFIS classifier.

If the tth sample is used to update the parameters of the network, the matrices A ∈ K × K and d ∈ K × 1 are updated as given in Subramanian et al. (2014) and the output weights are updated as:

 T  t 

w = w + A−1 F¯ t

e

(33)

Sample deletion criterion: A sample is deleted if the predicted class label is the same as the actual class label and the prediction error for the current sample is lower than the ‘delete’ threshold. The sample delete condition is given by:

ct == cˆt & E t < Ed

(34)

where, Ed is the sample ‘delete’ threshold. Similar to Ea , it is also found using a grid-search with five-fold cross validation method and can be chosen in the range [0.008, 0.01]. Sample reserve criterion: If the current sample does not satisfy any of the rule growing or parameter update or sample delete criterion, it is pushed to the rear-end of the training data stream. Selfregulating nature of ‘add’ and ‘learn’ thresholds facilitate the usage of these samples at a later stage. 2.1.3. Spectral filters selection using cognitive fuzzy inference system In this section, the approach to find the subject-specific spectral filters using the CFIS algorithm is explained in detail. The CFIS algorithm is employed to remove all those bands from the filter bank, whose elimination will not have any effect on the performance. It is referred to as Spectral-CFIS (S-CFIS). This mechanism is illustrated in Fig. 3. The features obtained from the filter bank are passed as input to the CFIS algorithm. In each iteration, one of the bands is dropped and the performance is re-computed. If the performance does not drop, the particular band is termed as redundant and excluded from the final set of bands. This procedure is carried out for all the bands and a final smaller set of bands, Bnew is obtained. The above procedure is summarized in the pseudocode Algorithm 1 . 2.1.4. Spatial filters selection using cognitive fuzzy inference system In this section, the procedure for finding the subject-specific spatial filters is explained in detail. Let the bands selected in the previous step be given as Bi , Bj and Bk . The corresponding spatial filter set of these bands is given as the filter set,

Algorithm 1: Pseudo-code for spectral filters selection using CFIS. Initialization: Band set B = {(B1 . . . B j . . . B6 )} BestTrainAccuracy←− 0 TrainAccuracy ←− 0 Calculate training error using features from all the bands BestTrainAccuracy ←− CFIS(B) for j−→1 to 6 do Bnew ←− B \ B j TrainAccuracy ←− CFIS(Bnew ) if TrainAccuracy ≥ BestTrainAccuracy then BestTrainAccuracy ←− TrainAccuracy. else concatenate(Bnew , B j ) end Output:Bnew

f={(vi1 ..vi2m )..(vj1 ..vj2m )..(vk1 ..vk2m )}. Here, the appropriate spatial filters are selected inside the obtained bands after the band elimination procedure. The iterative procedure for finding the spatial filters is shown in pseudo-code Algorithm 2 and illustrated in Fig. 3. At each iteration, a spatial filter is dropped from the spatial filter set, f obtained by the selected bands in the step one. If the performance of CFIS does not drop, this spatial filter is considered redundant and excluded from the spatial filter set. This operation is carried out for all the spatial filters in the spatial filter set and a final set of spatial filters are obtained. The mechanism to find spectral filters followed by spatial filters using CFIS is referred to as spatio-spectral CFIS (SS-CFIS). In the next section, the performance of S-CFIS and SS-CFIS is evaluated using benchmark data sets. 3. Experimental results The performances of S-CFIS and SS-CFIS have been evaluated using two publicly available BCI competition data sets, viz., data set-IIa from the BCI competition IV (Naeem et al., 2006) and

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

Algorithm 2: Pseudo-code for spatial filters selection using CFIS. Initialization: Filter set f ={(vi1 ..vi2m )..(v j1 ..v j2m )..(vk1 ..vk2m )} n ←− |f | BestTrainAccuracy←− 0 TrainAccuracy ←− 0 Calculate training error using all the spatial filters BestTrainAccuracy ←− CFIS(f ) for j−→1 to n do f f inal ←− f \ f j TrainAccuracy ←− CFIS(f f inal ) if TrainAccuracy ≥ BestTrainAccuracy then BestTrainAccuracy ←− TrainAccuracy. else concatenate(f f inal , f j ) end Output:f f inal

data set-IVa from the BCI competition III (Blankertz et al., 2006). First, the performance of selecting the desired spectral and spatiospectral filters is evaluated using nine subjects from the data setIIa from the BCI competition IV, performing the following six binary classification tasks, viz. left vs right hand, left hand vs foot, left hand vs tongue, right hand vs foot, right hand vs tongue and foot vs tongue. This study is also supported by performance evaluation using additional five subjects from the data set-IVa of BCI competition III. The performance of S-CFIS and SS-CFIS is compared with six state of art algorithms (Suk & Lee, 2013). 3.1. Performance evaluation of S-CFIS and SS-CFIS using the BCI competition IV data set-IIa Data set IIa (Naeem et al., 2006) consists of data of nine subjects. The EEG signals are recorded using twenty two non-invasive electrodes and sampled at two hundred fifty Hz. The experiment is conducted over two sessions on two separate days for each subject. Each session consists of two hundred eighty eight (seventy two for each class) trials which are distributed over six runs separated by short breaks. In the experiment, subjects performed movement of left hand (class one), right hand (class two), foot (class three) and tongue (class four) motor-imagery tasks. The training and testing data consisted of seventy two trials for each class. The signals between 0.5 s and 2.5 s after the onset of cue has been considered for both calibration and evaluation. Table 2 shows the selected bands and classification accuracies of S-CFIS and SS-CFIS for each subject for all the six classification tasks (left vs right hand classification, left vs foot classification, left vs tongue classification, right vs foot classification, right vs tongue classification and foot vs tongue classification). It may be observed from the Table 2 that both S-CFIS and SS-CFIS achieve the highest accuracy for left hand vs tongue motor-imagery task implying that these motor-imagery tasks are most distinguishable, closely followed by the left vs foot and the right vs foot tasks. It may also be noted that the spatio-spectral filter selection leads to improvement in the average classification performance for all the classification tasks. Based on the experimental results, existence of common frequency bands of ((8–16) Hz and (20–32) Hz) for all subjects across the classification tasks were observed. If these common frequency bands across all subjects are used, the performance is comparable with individual spatial-spectral filter selection. In order to further study the effect of the order of spatial and spectral filtering, three experiments were conducted: Case A) Simultaneous spatial and spectral filtering. Case B) Spatial filtering

381

followed by spectral filtering. Case C) Spectral filtering followed by spatial filtering. Table 1 shows the selected spatial filters and spectral filters and the classification accuracy of the subject A5 for left vs right task for all the three cases. From Table 1, it is seen that “Spectral followed by spatial filtering” achieves the best performance among all the three cases. In Case A, none of the spectral bands are eliminated and hence there is no effect of spectral filtering. Also the time taken is more, as the algorithm eliminates the spatial filters one by one with all the frequency bands intact. For Case B, it is seen that changing the order of spectral and spatial filtering has little effect on the performance. However, at the start the total number of spatial filters is very high. Hence, performing spatial filtering before spectral filtering is computationally complex. Next, the performances of S-CFIS and SS-CFIS are compared with other algorithms like Common Spatio-Spectral Pattern (CSSP) method (Lemm et al., 2005), the Filter Bank CSP (FBCSP) method (Ang et al., 2008), the Discriminative Filter Bank CSP (DCSP) method (Thomas et al., 2009), the Optimal Spatio-Spectral Filter Network (OSSFN) method (Zhang et al., 2011) (OSSFNwFBCSPOSSFN utilizing FBCSP, OSSFNwFBCSP-OSSFN utilizing DCSP) and the Bayesian Spatio-Spectral Filter Optimization (BSSFO) method (Suk & Lee, 2013). Table 3 shows the classification performance (subject mean and standard deviation (nine subjects)) for all the algorithms for all the six classification tasks. It may be noted here that all the compared algorithms use the SVM as classifier. For all the algorithms except S-CFIS and SS-CFIS, the results have been adapted from (Suk & Lee, 2013). From the table, it can be observed that S-CFIS and SS-CFIS performs better than all the algorithms across all the tasks. In comparison to the baseline CSP algorithm, S-CFIS alone achieves an increment of 18–20% on an average. It achieves the highest gain of 19% compared to CSP in case of left vs foot classification task. This shows the advantages of spectral and spatial filter selection as S-CFIS and SS-CFIS when compared to the baseline performance. In comparison to the time domain method CSSP, the proposed method achieves an average increment of 6%. In comparison to the filter bank based methods like the BCI competition winning FBCSP algorithm, it achieves an average increment of 9% with highest gain for right vs foot classification task. This might be due to efficient handling of non-stationarity in EEG signals. In comparison to the next best performing algorithm BSSFO, SS-CFIS achieves an average increment of 2%. However, BSSFO uses an ensemble of multiple classifiers from multiple bands in an iterative manner which makes the algorithm complex in nature, whereas the proposed approach uses a simpler approach employing fixed size bands with a single classifier. Also, the standard deviation of SS-CFIS is lower than BSSFO for all the tasks which shows that the proposed approach is more consistent across subjects in comparison to BSSFO. In case of complex task like foot vs tongue the SSCFIS achieves the highest increment of 4% over BSSFO. This shows the effectiveness of the proposed method in case of complex tasks. 3.2. Performance evaluation of S-CFIS and SS-CFIS using the BCI competition III data set-IVa The data set-IVa (Blankertz et al., 2006) consists of five subjects (‘aa’, ‘al’, ‘av’, ‘aw’, ‘ay’) and is recorded using one hundred eighteen electrodes. The subjects performed left hand, right hand and right foot motor-imagery tasks. In this study, the data for the right hand (class one) and foot (class two) motor-imagery tasks are used for classification. The number of trials in both the training and testing phase varies for each subject. Subject ‘aa’ has one hundred sixty eight trials in training and one hundred twelve trials in the testing set, subject ‘al’ has two hundred twenty four trials in training and fifty six trials in the testing set, subject ‘av’ has eighty four trials in the training set and one hundred ninety six trials in the testing

382

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384 Table 1 Classification performance in %, selected bands and spatial filters obtained for subject A5 for the left-right classification task for the BCI competition IV data set-IIa. Order of filter

Selected bands

Selected spatial filters

Testing accuracy

CASE A CASE B CASE C

1,2,3, 4,5,6 4,5,6 2,4,6

v11 v12 v13 v14 v15 v16 , v21 v22 v23 v24 , v31 v32 v33 v35 v36 , v41 v42 v43 v44 v45 v46 , v41 v42 v43 v44 v45 v46 , v61 v62 v64 v66 v41 v42 v43 v44 v45 v46 , v41 v42 v43 v44 v45 v46 , v61 v62 v64 v66 v21 v22 v23 v24 v25 , v43 v44 v45 v46 , v61 v62 v63 v64 v65 v66

72.59 79.26 80.28

Table 2 Classification performance (Subject Mean (SM) in % and standard deviation(SD)) of S-CFIS, SS-CFIS and Bands obtained for each subject for all the six classification tasks for the BCI competition IV data set-IIa. Subject

Selected

Testing accuracy

bands

S-CFIS

SS-CFIS

(a) Left vs right A1 1,2,4,6 A2 1,3,5,6 A3 1,3,6 A4 1,2,3,4,5,6 A5 2,5,6 A6 2,6 A7 3,4,5,6 A8 1,2,3,4,5,6 A9 1,2,3,4,5

96.45 66.90 96.35 78.45 79.58 74.29 80.00 98.51 93.85

97.78 69.01 97.81 78.45 80.28 74.29 80.00 98.51 94.62

SM SD

84.93 11.51

(d) Right vs foot A1 1,2 A2 2,4,5,6 A3 1,2,4,6 A4 2,3,4,5 A5 3,4,5,6 A6 1,2,3,4,5,6 A7 2,3,4,5 A8 1,2,3,4,6 A9 1,2,4,5,6 SM SD

Subject

Selected

Testing accuracy

bands

S-CFIS

SS-CFIS

(b) Left vs foot A1 1,2,3,4,5,6 A2 3,4,5,6 1,2,5 A3 A4 2 A5 2,5,6 A6 2,3,4,5,6 A7 2,3,4,5,6 A8 1,2,4,5,6 A9 1,3,4,5,6

99.29 82.86 94.81 93.28 79.58 80.37 99.30 94.81 97.76

100 83.57 97.04 93.38 80.28 80.37 100 94.81 97.76

85.63 11.50

SM SD

91.34 8.11

100 80 98.55 93.10 83.21 77.06 99.29 92.70 91.79

100 80 98.55 95.69 78.90 78.90 100 94.16 92.54

(e) Right vs tongue A1 1,3,4,5 A2 2,4,5,6 A3 1,2,6 A4 3,4,5 A5 2,5,6 A6 1,2,3,6 A7 3,4,5,6 A8 1,2,4,6 A9 2,3,5,6

90.63 8.58

90.97 9.13

SM SD

Subject

Selected

Testingaccuracy

bands

S-CFIS

SS-CFIS

(c) Left vs tongue A1 1,4,5 A2 1,2,3,5,6 A3 1,2,5,6 A4 1,3,4,5 A5 1,4,6 A6 1,2,5,6 A7 1,2,5,6 A8 2,4,5 A9 4

97.89 76.92 95.56 96.43 84.89 77.36 98.54 97.76 98.46

99.30 78.32 97.76 95.54 84.89 77.36 99.27 97.76 98.46

91.91 8.21

SM SD

91.53 9.18

92.07 9.21

100 73.43 99.35 89.09 84.33 77.78 98.52 93.38 91.54

100 75.52 99.28 89.09 90.30 80.56 100 93.38 91.54

(f) Foot vs tongue A1 1,2,3 A2 3,4,6 A3 2 A4 3,4,5,6 A5 1,2,3,4,5,6 A6 2,3,4,5,6 A7 1,2,4,5 A8 1,4 A9 1,4

77.14 83.27 92.65 83.83 69.50 73.83 87.59 92.79 94.78

78.57 84.40 94.12 84.82 70.21 79.44 89.78 93.43 96.27

89.71 9.55

91.07 8.58

SM SD

83.93 8.96

85.67 8.59

Table 3 Performance comparison (Subject Mean (SM) in % and standard deviation(SD)) of S-CFIS and SS-CFIS on the BCI competition IV data set-IIa. Algorithms

left vs right

CSP CSSP FBCSP DCSP OSSFNwFBCSP OSSFNwDCSP BSSFO S-CFIS SS-CFIS

73.46 79.78 76.31 77.55 76.31 75.62 83.80 84.93 85.63

± ± ± ± ± ± ± ± ±

16.93 13.69 17.90 18.33 18.59 18.84 13.09 11.51 11.50

left vs foot

left vs tong.

72.76 ± 17.93 86.19 ± 12.47 81.33 ± 13.18 80.71 ± 17.95 80.94 ± 14.18 84.11 ± 12.15 89.89 ± 10.71 91.34 ± 8.11 91.91 ± 8.21

74.07 84.80 80.86 79.71 79.01 78.32 90.43 91.53 92.07

set, subject ‘aw’ has fifty six trials in the training and two hundred twenty four trials in the testing set and subject ‘ay’ has twenty eight trials in the training set and two hundred fifty two trials in the testing set. Table 4 presents the number of train, test samples, selected bands and classification accuracies achieved for all the five subjects for both S-CFIS and SS-CFIS. It may be seen from this table that for the subject ‘aw’, SS-CFIS improves the performance of SCFIS by 6% highlighting the importance of spatio-spectral filtering. Performance of both the proposed algorithms have been compared with CSSP, FBCSP, DCSP, OSSFN (OSSFNwFBCSP, OSSFNwFBCSP) and BSSFO. Table 5 shows the performance for individual subjects and the mean accuracy and standard deviation for all the algorithms. The results of other algorithms except S-CFIS and SS-CFIS have

± ± ± ± ± ± ± ± ±

16.57 13.96 15.17 18.28 11.02 10.81 10.03 9.18 9.21

right vs foot

right vs tong.

73.07 ± 15.82 82.64 ± 14.09 79.86 ± 15.52 82.10 ± 15.41 75.54 ± 16.71 78.01 ± 15.38 87.89 ± 11.29 90.63 ± 8.58 90.07 ± 9.13

72.38 86.34 80.32 77.47 79.01 79.55 89.04 89.71 91.07

± ± ± ± ± ± ± ± ±

17.83 11.92 12.89 15.27 10.31 12.29 11.43 9.55 8.58

foot vs tong. 67.28 ± 11.40 77.16 ± 9.55 73.30 ± 11.17 73.30 ± 12.33 75.08 ± 12.39 76.08 ± 10.11 81.10 ± 9.73 83.93 ± 8.96 85.67 ± 8.59

Table 4 Classification performance (Subject Mean (SM) in % and standard deviation(SD)) of S-CFIS, SS-CFIS and Bands obtained for each subject for the BCI competition III data set-IVa. Subject

No. of samples

Selected

Testing accuracy

Train

bands

S-CFIS

SS-CFIS

1,2,3,5 1,3,4 1,2,4,5,6 2 1,3,4,5

82.14 100 60.20 76.79 60.32

82.14 100 63.27 83.04 60.32

75.89 16.65

77.75 16.24

Test

(a) Right hand vs foot aa 168 112 al 224 56 av 84 196 aw 56 224 ay 28 252 SM SD

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384 Table 5 Performance comparison of S-CFIS and SS-CFIS on BCI competition III data set-IVa. Algorithm

CSP CSSP FBCSP DCSP OSSFNwFBCSP OSSFNwDCSP BSSFO S-CFIS SS-CFIS

SM ± SD

Subjects aa

al

av

aw

ay

66.07 79.46 69.64 69.64 75.00 75.00 79.46 82.14 82.14

89.29 92.43 80.36 82.14 83.93 83.93 94.64 100 100

52.55 52.55 47.96 54.08 53.06 53.06 57.65 60.20 63.27

47.77 91.52 55.36 50.89 74.11 74.11 91.96 76.79 83.04

52.38 51.59 48.41 48.41 48.81 47.22 53.57 60.32 60.32

61.79 73.60 60.35 61.03 66.98 66.46 75.46 75.89 77.75

± ± ± ± ± ± ± ± ±

(16.98) (20.33) (14.21) (14.21) (15.22) (15.93) (19.06) (16.65) (16.24)

SM= Subject Mean, SD= Standard deviation

been adapted from Suk and Lee (2013). They have used SVM as the classifier. It can be observed from the table that in comparison to the baseline CSP method, S-CFIS and SS-CFIS achieve an increment of 14 and 16%, respectively. This highlights the importance of spectral and spatio-spectral filtering over the baseline. In case of subject ‘aw’, CSSP and BSSFO perform better than the proposed method. This might be due to use of time delay and ensemble of classifiers, respectively, however, on an average the proposed approach performs better than both of these algorithms. The proposed method performs significantly better than other similar methods like FBCSP (15–17%) and DCSP (14–16%) which use CSP as the baseline. This highlights the effectiveness of the proposed method. The frequency bands of width 4 Hz employed in this paper have been shown to yield a stable frequency response and provide the best results in comparison to other bands of varying widths in the range of 2–6 Hz (Ang et al., 2012; Thomas et al., 2009). In contrast, freely searching of the desired frequency bands will result in a higher computational overload. Therefore, the frequency bands of width 4 Hz have been selected in the range of 8–32 Hz. It is also possible to apply the proposed approach to data sets which comprise more than two classes either using a One-Versus-Rest or a Pair-Wise multi-classification approach. It can be concluded from the results that the spatial filtering is a minor improvement over spectral filtering. Spatial filtering is used if one wants to finetune the performance further after the spectral filtering is done. The main advantages of the proposed approach over the existing works is the ability to handle non-stationarity and provide good classification performance whereas the main limitation of the method is the usage of a filter bank of fixed size bands. Also, a cross-validation approach is employed for selection of CFIS algorithm parameters, which may not be suitable for online training. The next best performing algorithm, BSSFO uses an ensemble of multiple classifiers from multiple bands in an iterative manner. This makes the BSSFO algorithm complex in nature, whereas the proposed approach uses a simpler approach employing fixed size bands with a single classifier. The CFIS approach presented in the paper can be used for handling non-stationary signals in other domains. The proposed mechanism can be extended for stroke rehabilitation studies. Another important direction is to handle multiclass classification directly in an IT2FIS framework. 4. Conclusions A CFIS based subject-specific spatio-spectral filter selection approach has been presented for EEG based motor-imagery task classification in BCI. The proposed approach uses an interval type-2 fuzzy classifier, which can handle the non-stationarity in EEG signals. Further, the proposed approach selects the appropriate spectral and spatio-spatial filters. Performance of S-CFIS and SS-CFIS

383

has been evaluated using two publicly available BCI competition data sets, viz., data set-IVa (Dornhege, Blankertz, Curio, & Muller, 2004) from the BCI competition III and data set-IIa (Naeem et al., 2006) from the BCI competition IV. Based on the subject-wise studies, the experimental results indicate a significant performance improvement by both the spectral and spatio-spectral filter selection scheme. It also indicates presence of common frequency bands ((8–16) Hz and (20–32) Hz) for all subjects across all the classification tasks. It can be concluded that spectral filtering is more important than spatial filtering and order of filtering has little impact on the performance. Further, the subject-specific SS-CFIS outperforms the CSP method by approximately 15-18% and other algorithms like FBCSP, DCSP by 8–10%. Compared to a recently proposed algorithm BSSFO, it achieves an improvement of 2%, but is simpler in comparison to BSSFO, which uses multiple classifiers. Acknowledgment We would like to thank the reviewers for the valuable comments which have improved the quality of paper. The authors wish to extend their thanks to the Ministry of Education (MoE), Singapore, for providing financial support through tier I (No. M4011269) funding to conduct this study. References Ahn, M., Lee, M., Choi, J., & Jun, S. C. (2014). A review of brain-computer interface games and an opinion survey from researchers, developers and users. Sensors, 14, 14601–14633. Alhaddad, M. J., Mohammed, A., Kamel, M., & Hagras, H. (2015). A genetic interval type-2 fuzzy logic-based approach for generating interpretable linguistic models for the brain P300 phenomena recorded via brain–computer interfaces. Soft Computing, 19, 1019–1035. Ang, K. K., Chin, Z. Y., Wang, C., Guan, C., & Zhang, H. (2012). Filter bank common spatial pattern algorithm on BCI competition IV datasets 2a and 2b. Frontiers in Neuroscience, 6, 1–9. Ang, K. K., Chin, Z. Y., Zhang, H., & Guan, C. (2008). Filter bank common spatial pattern (FBCSP) in brain-computer interface. In International joint conference on neural networks, 2008. (world congress on computational intelligence) (pp. 2390–2397). IEEE. Ang, K. K., & Guan, C. (2015). Brain–computer interface for neurorehabilitation of upper limb after stroke. Proceedings of the IEEE, 103, 944–953. Babu, G. S., & Suresh, S. (2013). Sequential projection-based metacognitive learning in a radial basis function network for classification problems. IEEE Transactions on Neural Networks and Learning Systems, 24, 194–206. Birbaumer, N. (2006). Brain–computer-interface research: coming of age. Clinical Neurophysiology, 117, 479–483. Blankertz, B., Muller, K., Krusienski, D. J., Schalk, G., Wolpaw, J. R., Schlogl, A., et al. (2006). The BCI competition III: Validating alternative approaches to actual BCI problems. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 14, 153–159. Curran, E. A., & Stokes, M. J. (2003). Learning to control brain activity: a review of the production and control of EEG components for driving brain–computer interface (BCI) systems. Brain and Cognition, 51, 326–336. Das, A., Subramanian, K., & Suresh, S. (2015). A subject-specific frequency band selection for efficient BCI- an interval type-2 fuzzy inference system approach. In Ieee international conference on fuzzy systems. (pp. 1–8). IEEE. Dornhege, G., Blankertz, B., Curio, G., & Muller, K. (2004). Boosting bit rates in noninvasive EEG single-trial classifications by feature combination and multiclass paradigms. IEEE Transactions on Biomedical Engineering, 51, 993–1002. Dornhege, G., Blankertz, B., Krauledat, M., Losch, F., Curio, G., & Muller, K. (2006). Combined optimization of spatial and temporal filters for improving brain-computer interfacing. IEEE Transactions on Biomedical Engineering, 53, 2274–2281. Doud, A. J., Lucas, J. P., Pisansky, M. T., & He, B. (2011). Continuous three-dimensional control of a virtual helicopter using a motor imagery based brain-computer interface. PloS One, 6, e26322. Fukunaga, K. (1990). Introduction to statistical pattern recognition. Academic Press. Herman, P., Prasad, G., & McGinnity, T. M. (2005). Investigation of the type-2 fuzzy logic approach to classification in an EEG-based brain-computer interface. In 27th annual international conference on of the engineering in medicine and biology society (pp. 5354–5357). IEEE. Herman, P., Prasad, G., & McGinnity, T. M. (2008). Design and on-line evaluation of type-2 fuzzy logic system-based framework for handling uncertainties in BCI classification. In 30th annual international conference on of the ieee engineering in medicine and biology society (pp. 4242–4245). IEEE. INOUE, K., Mori, D., Sugioka, K., Pfurtscheller, G., & Kumamaru, K. (2004). Feature extraction of EEG signals during right and left motor imagery. In Sice annual conference (pp. 2183–2187).

384

A.K. Das et al. / Expert Systems With Applications 64 (2016) 375–384

LaFleur, K., Cassady, K., Doud, A., Shades, K., Rogin, E., & He, B. (2013). Quadcopter control in three-dimensional space using a noninvasive motor imagery-based brain–computer interface. Journal of Neural Engineering, 10, 046003. Lemm, S., Blankertz, B., Curio, G., & Muller, K. (2005). Spatio-spectral filters for improving the classification of single trial EEG. IEEE Transactions on Biomedical Engineering, 52, 1541–1548. Lotte, F., & Guan, C. (2011). Regularizing common spatial patterns to improve BCI designs: unified theory and new algorithms. IEEE Transactions on Biomedical Engineering, 58, 355–362. Mak, J. N., & Wolpaw, J. R. (2009). Clinical applications of brain-computer interfaces: current state and future prospects. IEEE Reviews in Biomedical Engineering, 2, 187. Mendel, J. M., & Liu, X. (2012). New closed-form solutions for Karnik-Mendel algorithm+ defuzzification of an interval type-2 fuzzy set. In Ieee international conference on on fuzzy system (pp. 1–8). IEEE. Meng, J., Yao, L., Sheng, X., Zhang, D., & Zhu, X. (2015). Simultaneously optimizing spatial spectral features based on mutual information for EEG classification. IEEE Transactions on Biomedical Engineering, 62, 227–240. Naeem, M., Brunner, C., Leeb, R., Graimann, B., & Pfurtscheller, G. (2006). Seperability of four-class motor imagery data using independent components analysis. Journal of Neural Engineering, 3, 208. Nguyen, T., Khosravi, A., Creighton, D., & Nahavandi, S. (2015). EEG signal classification for BCI applications by wavelets and interval type-2 fuzzy logic systems. Expert Systems with Applications, 42, 4370–4380. Nie, M., & Tan, W. W. (2008). Towards an efficient type-reduction method for interval type-2 fuzzy logic systems. In Ieee international conference on fuzzy systems.(world congress on computational intelligence). (pp. 1425–1432). IEEE. Nijholt, A., Bos, D. P.-O., & Reuderink, B. (2009). Turning shortcomings into challenges: Brain–computer interfaces for games. Entertainment Computing, 1, 85–94. Novi, Q., Guan, C., Dat, T., & Xue, P. (2007). Sub-band common spatial pattern (SBCSP) for brain-computer interface. In 3rd international conference on neural engineering (pp. 204–207). Pfurtscheller, G., & Neuper, C. (2001). Motor imagery and direct brain-computer communication. Proceedings of the IEEE, 89, 1123–1134. Ramoser, H., Muller-Gerking, J., & Pfurtscheller, G. (20 0 0). Optimal spatial filtering of single trial EEG during imagined hand movement. IEEE Transactions on Rehabilitation Engineering, 8, 441–446.

Schlögl, A., Lee, F., Bischof, H., & Pfurtscheller, G. (2005). Characterization of four– class motor imagery EEG data for the BCI-competition 2005. Journal of Neural Engineering, 2, L14. Subramanian, K., Das, A., Suresh, S., & Ramasamy, S. (2014). A meta-cognitive interval type-2 fuzzy inference system and its projection based learning algorithm. Evolving Systems, 5, 219–230. Subramanian, K., & Suresh, S. (2012). A meta-cognitive sequential learning algorithm for neuro-fuzzy inference system. Applied Soft Computing, 12, 3603–3614. Suk, H.-I., & Lee, S.-W. (2013). A novel bayesian framework for discriminative feature extraction in brain-computer interfaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35, 286–299. Thomas, K. P., Guan, C., Lau, C. T., Vinod, A. P., & Ang, K. K. (2009). A new discriminative common spatial pattern method for motor imagery brain–computer interfaces. IEEE Transactions on Biomedical Engineering, 56, 2730–2733. Tomioka, R., Dornhege, G., Nolte, G., Blankertz, B., Aihara, K., & Müller, K.-R. (2006). Spectrally weighted common spatial pattern algorithm for single trial EEG classification. Deptartment of Mathematical Engineering, University of Tokyo, Tokyo, Japan, Technical Report, 40. Vallabhaneni, A., Wang, T., & He, B. (2005). Brain-computer interface. Neural Engineering, 85–121. Wolpaw, J. R., Birbaumer, N., Heetderks, W. J., McFarland, D. J., Peckham, P. H., Schalk, G., et al. (20 0 0). Brain-computer interface technology: a review of the first international meeting. IEEE Transactions on Rehabilitation Engineering, 8, 164–173. Wolpaw, J. R., Birbaumer, N., McFarland, D. J., Pfurtscheller, G., & Vaughan, T. M. (2002). Brain–computer interfaces for communication and control. Clinical Neurophysiology, 113, 767–791. Wu, W., Gao, X., Hong, B., & Gao, S. (2008). Classifying single-trial EEG during motor imagery by iterative spatio-spectral patterns learning (ISSPL). IEEE Trans. Biomedical Engineering, 55, 1733–1743. Zhang, D., Wang, Y., Gao, X., Hong, B., & Gao, S. (2007). An algorithm for idle-state detection in motor-imagery-based brain-computer interface. Computational Intelligence and Neuroscience, 2007, 5. Zhang, H., Chin, Z. Y., Ang, K. K., Guan, C., & Wang, C. (2011). Optimum spatio-spectral filtering network for brain–computer interface. IEEE Transactions on Neural Networks, 22, 52–63.