- Email: [email protected]
Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller
BBE 200 1–10 biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Available online at www.sciencedirect.com
ScienceDirect journal homepage: www.elsevier.com/locate/bbe 1 2 3
Original Research Article
Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller
4 5 6
7 8 9 10
Q1
Rahul Kumar Chaurasiya a,*, Narendra D. Londhe b, Subhojit Ghosh b a
Department of Electronics and Telecommunication Engineering, National Institute of Technology, Raipur, Raipur, PIN-492010, India b Department of Electrical Engineering, National Institute of Technology, Raipur, India
article info
abstract
Article history:
P300 speller-based brain-computer interface (BCI) allows a person to communicate with a
Received 5 October 2016
computer using only brain signals. In order to achieve better reliability and user continence,
Received in revised form
it is desirable to have a system capable of providing accurate classification with as few EEG
9 March 2017
channels as possible. This article proposes an approach based on multi-objective binary
Accepted 20 April 2017
differential evolution (MOBDE) algorithm to optimize the system accuracy and number of
Available online xxx
EEG channels used for classification. The algorithm on convergence provides a set of paretooptimal solutions by solving the trade-off between the classification accuracy and the
Keywords:
number of channels for Devanagari script (DS)-based P300 speller system. The proposed
BCI
method is evaluated on EEG data acquired from 9 subjects using a 64 channel EEG acquisition
Devanagari
device. The statistical analysis carried out in the article, suggests that the proposed method
Multi-objective Optimization
not only increases the classification accuracy but also increases the over-all system reliabil-
Binary DE
ity in terms of improved user-convenience and information transfer rate (ITR) by reducing
P300-speller
the EEG channels. It was also revealed that the proposed system with only 16 channels was
SVM
able to achieve higher classification accuracy than a system which uses all 64 channel's data for feature extraction and classification. © 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier B.V. All rights reserved.
13 11 12 14 15 16
17 18 19
1.
Introduction
P300 speller-based brain-computer interface (BCI) allows a person to communicate with a computer using only brain signals [1]. The communication does not require any physical
movements, and is particularly useful for patients suffering from severe motor disabilities but having cognitive abilities [2]. The most widely used P300 speller works in an odd-ball paradigm-based experimental environment [1,3]. In the oddball experiment, the subjects are randomly presented with two types of events, one of which rarely occurs (the odd-ball). The
* Corresponding author at: Department of Electronics and Telecommunication Engineering, National Institute of Technology, Raipur, Raipur-C.G, PIN-492010, India. E-mail addresses: [email protected] (R.K. Chaurasiya), [email protected] (N.D. Londhe), [email protected] (S. Ghosh). http://dx.doi.org/10.1016/j.bbe.2017.04.006 0208-5216/© 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier B.V. All rights reserved. Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
20 21 22 23 24 25
BBE 200 1–10
2 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
rare events generates a P300 event related potential (ERP) in the recorded electroencephalogram (EEG). Farwell and Donchin first developed a P300 speller for English alphabetic script using a 6 6 matrix [1]. All rows and columns of the matrix were randomly intensified and the subject was asked to focus on the target character (that one, which he wants to communicate). The speller is commonly known as row/column (RC) paradigm-based speller. The multiple trails of the intensifications were repeated and the averaged signals were used to improve the signal-to-noise ratio (SNR). Several methods have been reported to further improve the system reliability and information transfer rate (ITR). The method includes improvement in display paradigm by changing matrix size, background color, font size, and inter stimulus interval [4–7]. Different classification methods such as SWLDA, Baysian linear discriminant analysis (BLDA), support vector machines (SVMs), and artificial neural networks (ANN), have been successfully applied for classification in P300 spellers [8–15]. Although the RC paradigm is still the most widely used paradigm for P300 spellers, experiments with single character (SC) paradigm [16], region-based (RB) paradigm [17], and check board (CB) paradigms [18] have also been tried in recent years. The performance of classifiers used for detection of P300 ERPs significantly depends on the choice of features and hence only the most discriminative features should be ideally used for classification. However, in the case of P300 spellers, the channel set providing the most relevant information varies from subject to subject [19,20]. For an EEG device having 64 EEG channels, there are total 264 possible subsets and practically it is impossible to select the best channel subset using exhaustive search. Hence, different channel selection methods such as channel selection using recursive channel elimination [21], jump-wise regression [22], Gibbs sampling [23], multi-ganglion [24], particle swarm optimization (PSO) [25] and genetic algorithm (GA), have been proposed for improving the classifier performance by selecting the best channel subset. Although, most of research work on the use of visual P300 spellers to aid communication has concentrated on languages that are written with English alphabetic script, a P300 speller systems capable of communicating text in Devanagari script (DS) has been developed in [15], which aimed at improving the system reliability by maximize the classification accuracy. A binary differential evolution (DE)-based optimization method was used for selection of channel subset with a single objective of achieving maximum classification accuracy. In the proposed study, we have extended the work of [15] for improving the user convenience and ITR, in addition to increasing the classification accuracy of DS-based P300 speller system. An improved ITR and user-convenience in addition to accuracy is expected to further improve the reliability of the system. The ITR and user-convenience are directly related to the number of channels used for data acquisition. In this regard, the present work aims at the multiple objectives of improved classification accuracy and reduced number of channels. Two different approaches are generally used for solving multi-objective optimization problem, the first approach combine all the objective functions into a single composite objective function by assigning different
weightages to different objectives, while the second approach determine an entire pareto-optimal solution set. Considering the limitations related to the selection of proper weightage function, which is not known a priori, the second approach is preferable as it provides a set of pareto-optimal solutions and user can decide about which solution he wants to use, based on the priorities of the objectives. In this article, a multiobjective binary DE (MOBDE) algorithm is proposed for finding the pareto-optimal solution set for solving the trade-off between number of channels and classification accuracy. Due to the requirement of lesser number of algorithm specific parameters and simplicity of the algorithm, MOBDE has been preferred over other optimization approaches. The rest of the paper is organized as follows: The description of signal acquisition procedure and the dataset is given in Section 2. The character detection mechanism is also described in Section 2. The framing of the multi-objective problem of accuracy maximization & channel minimization as an optimization problem and its solution using MOBDE algorithm is described in Section 3. The results are presented in Section 4. With discussions in Section 5, the study is concluded in Section 6.
2.
Materials and methods
2.1.
The dataset
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
Total 9 healthy volunteer subjects of age range 21–29 (mean 26.5) were used to collect EEG responses. All subjects were able to communicate using DS. The EEG responses were recorded with a BrainAmp DC hardware which was equipped with a 64channel actiCAP. The EEG responses were collected at 500 Hz sampling frequency. At the time of recording, a digital bandpass filter of 1 to 250 Hz was also applied. General purpose BCI2000 software was used for stimulus presentation and data collection using DS-based display paradigm [26]. The DS paradigm used for stimulation is shown in Fig. 1(a) and the channel configuration used for data collection is shown in Fig. 1(b). Total 100 characters were presented in 20 runs as target characters for each subject. In a particular run, for each character, the rows and the column of the display matrix were randomly and successively intensified for 120 ms, the followed by 80 ms non-intensification period. In order to enhance the SNR of the acquired EEG signals, the sequences of intensifications of 16 rows and columns were repeated 15 times for each character. After 15 trials of one character, the recording for the next character was started with a gap of 10 s between the two characters. The 10 second gap was incorporated to make the subjects relax and to ensure that they can comfortably find the next target character in the display matrix. Total 240 responses (16 rows/columns 15 trials) were recorded per character. Two out of 16 responses were supposed to contain P300 ERPs.
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
2.2.
136
Preprocessing and feature extraction
The methodology for preprocessing and feature extraction in the present work has been adopted from reported works on P300 speller [2,27,28]. The EEG samples posterior to 600 ms
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
137 138 139
BBE 200 1–10
3
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Fig. 1 – (a) The 8 T 8 matrix containing 13 vowels, 37 consonants, and 10 digits of DS. Total 4 special characters were also used in stimulus presentation. (b) A 64 channel configuration for acquiring EEG responses. Channels AFz and FCz were also used as ground and reference, respectively.
140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
from the starting of each flashing were extracted from all 64 channels. As the interest was to capture the P300 ERPs, the selected time window of 600 ms was sufficient to capture the relevant information. The extracted samples were passed through a band-pass filter of cut-off frequencies between 1 and 10 Hz. The filter samples were than decimated with a frequency of 10 Hz. At this stage, each EEG response consists of 6 samples per channel. For these samples, normalization was carried out independently for each channel. Afterwards, feature vectors were formed by concatenation of the signal samples of all 64 channels. Thus for a single character, there are total 240 feature vectors, each of size 384 (6 samples 64 channels). In these 240 feature vectors, 30 are from class +1 (1 row and 1 column per trial 15 trials) and are expected to contain P300 ERPs. The rest of the feature vectors are in class 1. Sample variations of the average of the class +1 and 1 EEG signals for all nine subjects for channel CPz are depicted in Fig. 2. A peak in class +1 signal near 300 ms shows that the P300 ERPs were properly captured in the oddball experiment.
159
2.3.
f ðxÞ ¼
(1)
171 172 173 174 175 176 177
Srjc
J 1X ð jÞ ¼ f xrjc J j¼1
(2)
where f(xr|c) is the score assigned to features xr|c of a give row/column (r/c), with J trials (J can be chosen to be between 1 and 15). The character which is common in predicted row and column is the target character.
3. Optimizing for accuracy and number of channels
Classification
The classification task of predicting the row and column containing the P300 ERP is a two-class classification problem. Because of its better generalization capability over other machine learning algorithms [29], an SVM classifier was employed for this task. From the training data points Xi, i = 1, 2 . . . ..N (with respective class labels yi =1 or + 1, i = 1, 2 . . . ..N), SVM learns maximizes-margin hyper-plane while also trying to minimize the total errors in classification. The learned separating hyper-plane then can be used to assign a score to a new test data point X is represented as:
170
i¼1
where li, i = 1, 2 . . . ..N are the Lagrange's multipliers. Since in the presented work 15 trials have been recorded for each character, the rows/column are predicted based on the scores obtained by different rows/columns, as per the formulation of Eq. (2).
3.1. 160 161 162 163 164 165 166 167 168 169
N X yi li ðX:Xi Þ þ b
Binary differential evolution
DE is a population-based evolutionary optimization method. As compared to other evolutionary approaches, it is simple yet effective algorithm and requires only two control parameters [30]. The population of DE consists of continues valued floating-point encoded vectors. The population at iteration 0 is generated as a group of NP random vectors. Mutation, crossover and selection operation are generally performed in DE for updating the positions for coming iterations. Suppose we have NP, D-dimensional target vector xti ; i ¼ 1; 2; . . .NP at
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
179 178 180 181 182 183 184 185 186
187 188 189 190 191 192 193 194 195
BBE 200 1–10
4
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
S2
Normalized Amp.
S1
1
1
0
0
0
-1
0
100
200
300
-1
0
100
200
300
1
1
0
0
0
0
100
200
300
-1
0
100
S7
200
300
200
300
0
200
300
100 200 Sample No.
300
100
1
Class +1 Class -1 0
0
100 200 Sample No.
-1
S9
1
0
100
S8
1
-1
0
S6
1
-1
Normalized Amp.
-1
S5
S4 Normalized Amp.
S3
1
300
-1
0
0
100 200 Sample No.
300
-1
0
Fig. 2 – Averaged Sample variations of recorded EEG signals (average over all the signals for first character) of class +1 and class S1 for channel CPz. A peak in class +1 signal near 300 ms shows that the P300 ERPs were properly captured in the oddball experiment. Class S1 signal are shown by dotted red.
196 197 198
iteration number t, then, in mutation stage, a mutation vector for next iteration is generated for each bit of every target utþ1 ij vector as
199
utþ1 ¼ xtr1;j þ F xtr2;j xtr3;j ij
(3)
200 201 202 203
where index of dimensionality j varies from 1 to D. F is a positive constant and xr1,j, xr2,j and xr3,j are three bits of randomly chosen individuals with indexes r1 6¼ r2 6¼ r3 6¼ i.
204 205 206
In crossover stage, a trial individual vi is generated by crossing over the target vector xi with the corresponding mutant vector ui as follows
207
( vtþ1 ij
¼
utþ1 ij ; xtij ;
if ðrand jCRÞorð j ¼ randðiÞÞ otherwise
better. Otherwise, the target individual xti is carried forward for next stage. Mathematically, the target individual for next iteration is selected as
(4)
208 209 210 211
where CR is the crossover probability between (0,1); rand j are stochastic random number uniformly distributed within ½0; 1Þ; rand(i) are random integers within 1, 2.., D.
212 213
In selection stage, the trial individual vtþ1 replaces the i , if its fitness value is target individual xti for generation of xtþ1 i
xtþ1 i
¼
tþ1 > f xti vtþ1 i ; if f vi xti ; otherwise
214 215 216 217
(5)
A number of improved variants of DE have been proposed in recent years [31]. It has also been reported that DE can perform better than other optimization for real world problems [32]. However, the standard DE operates in continuous space and is not suitable for solving binary optimization problems. Hence, its binary versions have also been proposed to solve such problems [33–35]. In BDE algorithms, each bit of target vector is represented by either a 0 or 1. The methodology used to update the population in BDE is similar to DE; involving crossover after mutation, and finally selection operations. In order to ensure that the target vector consists of only 0 s and 1 s, a probability estimation operator is used to generate mutant vector in the mutation stage as follows
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
219 218 220 221 222 223 224 225 226 227 228 229 230 231 232 233
BBE 200 1–10
5
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
234
8 1 tþ1 > > 3 ¼2 > P xij > 2bðMO0:5Þ > > < 6 1 þ 2F 7 41 þ e 5 > > > > > > : MO ¼ xt þ F xt xt r1;j
235 236 237 238 239 240
r2;j
Initialize 100 target vectors (6)
Decode target vector to obtain channel subset
r3;j
where b is a positive valued bandwidth factor. The mutant operator MO in BDE is analogous to the mutation operation used in standard DE as in Eq. (3). A binary-coded mutant vector uijtþ1 for target vector xtij is generated as ( uijtþ1 ¼
1;
if randð ÞP xtþ1 ij
0;
otherwise
Obtain the feature subset corresponding to selected channels
Train SVM with training data
Repeat 5 times for 5-fold CV
(7)
241 242 243 244 245 246 247 248 249 250
The crossover operation and selection process in binary DE are same as in Eqs. (4) and (5) respectively. The BDE has one more algorithm parameters (i.e. bandwidth factor b) in addition to two algorithm specific parameters F and CR used in DE. In this article, the method of Wang et al. has been adopted for cannel selection using MOBDE [34]. The values for algorithm specific parameters were chosen to be same as suggested in [34].
251
3.2.
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
In this article, the reduction in number of channels and improvement in the classification accuracy has been framed as a binary optimization problem in. A binary coded target vector in the population represents a set of channel. As the dataset was recorded from 64 channels, a 64-dimensional binary vector xi ¼ xi;1 ; xi;2 ; ::; xi;64 constituted the target vector. Total 100 target vectors were randomly generated as initial population. In target vector xi, at iteration number t, the features for classification are extracted from channel j (1 ≤ j ≤ 64), if xtij holds a value '1'. The update mechanism of the target vectors for the next iteration is same as described in Section 3.1. Each target vector in the population is a candidate solution and represents a possible combination of set of channels. The optimization process tries to search for a solution that maximizes the classification accuracy and at the same time minimizes the number of channels. In multi-objective optimization frame work, the two objective functions, viz. maximizing the classification accuracy and minimizing the number of channels has been mathematically formulated as Objective 1: maxA(xi) Objective 2: minC(xi)
273
s:t: xi x
274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
In objective 1, A(xi) represents the accuracy of character detection, obtained from the SVM-based classification approach while choosing the channels for target vector xi. In objective 2, C(xi) is the count of number of 1 s present in target vector (xi). In Eq. (8), the constraint ‘‘xi x’’ indicates that target vectors xis contain a channel-set which is subset of x containing all 64 channels (i.e. all the entries in x are 1). The MOBDE algorithm was applied for each subject separately. It was observed that for each subject, there were no changes in the accuracy and number of channels after 40 iterations. Fig. 3 depicts the flow of the proposed methodology, involving SVM-based character detection and MOBDE-based channel selection method.
Obtain the classification accuracy with test data
Evaluate the fitness of the two objective functions
Update the target vectors using MOBDE algorithm
Find the pareto-optimal solutions
Multi-objective optimization framework
(8)
Iteration no.>40
No
Yes End Fig. 3 – Flow of the proposed methodology.
4.
Experimental results
289
In BCI, channel selection is usually an off-line process where the data acquired from all channels is used to select the optimal channel subset. Afterwards, the optimal channel subset is used for real-time applications. The contribution of the proposed MOBDE-based channel selection method for minimizing the number of channels and maximizing the classification accuracy has been evaluated in this section, which has been further divided into the following subsections: The classification accuracies obtained using SVM classifiers are presented in first subsection. The results of proposed MOBDE-based method for optimal channel selection have been presented in second subsection. The statistical analysis carried out in the third subsection reflects the effectiveness of the proposed channel selection method.
290 291 292 293 294 295 296 297 298 299 300 301 302 303
4.1.
Results for character classification using SVM
304
Table 1 depicts the accuracy of character detection obtained for 5, 10 and 15 trials using SVM classifier (for all 64 channels). A 5-fold cross validation methodology was adopted for training and testing the classifier. Average accuracy of 65.3%, 80.9% and 86.9% is obtained for 5, 10 and 15 trials, respectively.
305 306 307 308 309 310
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
BBE 200 1–10
6
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Table 1 – Percentage classification accuracy obtained using SVM classifier for the 5, 10 and 15 trials of the dataset for all 64 channels. Subject number
minimize the number of channel. The set of pareto-optimal solutions encountered by the target vectors have been shown as a pareto-front diagram in Fig. 4. The pareto-front diagram access the trade-off relationship between the number of channels and the classification accuracy. The pareto-front solutions shown in Fig. 4 represent a boundary at which an improvement of classification accuracy necessarily requires an increment in number of channels or, conversely, an attempt for reduction in the number of channels results in loss of accuracy. For 5, 10 and 15 trials, the classification accuracy obtained and the corresponding numbers of channel selected by the pareto-optimal solutions points with maximum accuracy are presented in Table 2. It can be observed from Fig. 4 that a total of 35 paretooptimal solution points were selected for 15 trials across all nine subjects. In these points, each of the 64 channels might have got selected between 0 to 35 times. Similarly, a total of 37 and 38 pareto-optimal solution points were selected for 5 and 10 trials, respectively. In order to analyze the locations of the channels that can provide more discriminatory information and better accuracy, the topographical map of channelfrequency among all the pareto-optimal solutions were plotted. Fig. 5 shows the topographical maps (considered over
Accuracy with different number of trials
1 2 3 4 5 6 7 8 9 Average
5
10
15
68 59 70 56 62 80 56 60 76 65.2
84 74 82 79 78 87 72 80 86 80.2
85 82 91 88 84 91 88 89 90 87.6
311
4.2.
312 313 314 315
The MOBDE was used with SVM-based character detection method for optimum channel selection using 5, 10 and 15 trials. As mentioned in Section 3.2, the optimization objectives were to maximize the accuracy of character detection and
Results for optimization using MOBDE
Subject 1
Subject 2
Channel numbers
64 visited places Pareto front
48
64 visited places Pareto front
48 32
32
16
16
16
75
80
85
90
95
0 100 70
75
85
90
95
0 70
visited places Pareto front
visited places Pareto front
48
32
16
16
16
80
85
90
95
100
0 70
75
Subject 7
80
85
90
95
visited places Pareto front
visited places Pareto front
48
32
16
16
16
80 85 90 Acc uracy (%)
75
95
100
0 70
100
80
85
90
95
100
75
80 85 90 Acc uracy (%)
95
100
visited places Pareto front
48
32
75
95
64
32
0 70
90
Subject 9
64
48
0 100 70
Subject 8
64
85
visited places Pareto front
48
32
75
80
Subject 6
32
0 70
75
64
64
48
100
Subject 5
Subject 4 64 Channel numbers
80
visited places Pareto front
48
32
0 70
Channel numbers
Subject 3
64
95
100 70
75
80 85 90 Acc uracy (%)
Fig. 4 – Illustration of the pareto-optimal solutions found by the MOBDE algorithmin for different subjects. The horizontal and vertical axes denote the classification accuracy and the number of selected channels, respectively. The 'visited places' points show the fitness of all the positions visited by the target particles. The 'pareto front' points represent the positions that belong to the pareto-optimal solution set. Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
BBE 200 1–10
7
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Table 2 – Percentage classification accuracy (A) and the corresponding number of channels selected (Nc) for 5, 10 and 15 trials obtained by pareto-optimal solution with maximum classification accuracy. Subject No.
Accuracy (A) and the corresponding number of channels selected (Nc) for 5, 10 and 15 trials obtained by pareto-optimal solution with maximum classification accuracy 5
1 2 3 4 5 6 7 8 9 Average
10
15
A
Nc
A
Nc
A
Nc
75 70 83 70 68 88 73 70 87 76.0
28 28 27 32 37 22 29 22 30 28.3
89 83 87 89 85 93 87 87 91 87.9
28 33 26 36 33 22 28 20 30 28.4
96 90 96 93 89 94 92 90 95 92.8
29 24 19 33 34 21 25 21 29 26.1
Fig. 5 – The topographical maps (common for all 9 subjects) showing the channel selection frequency in a total of 37, 38 and 35 pareto-optimal solution points for 5, 10 and 15 trials, respectively.
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
all 9 subjects) of the frequency of channel selection for 5, 10 and 15 trials. From Fig. 5, it can be observed that the parietal (P), occipital (O) and central (C) regions on scalp are most frequently selected. However, the optimal channel selection is a subject dependent problem i.e. there is inter subject variability in the optimal set [19,20,36]. Hence, in the present work, the topographical maps for individual subjects have also been analyzed. In order to ensure a sufficient number of data points are available for analysis, the pareto-optimal points obtained with 5, 10 and 15 trails were considered for each subject. As from Fig. 4, the total number of pareto-optimal points may vary in different solutions. After the experimentations for the present case, in total 110 pareto-optimal points (37 for 5 trials, 38 for 10 trials, and 35 for 15 trials) 12, 9, 13, 15, 8, 13, 14, 12 and 14 points were present in the pareto-optimal solution set of subject 1 to 9, respectively. The subject-wise topographical maps are shown in Fig. 6.
357
4.3.
358 359 360 361 362 363 364
In order to validate the statistical significance of the proposed method at different trials, the Friedman test was employed [37]. The test was applied for two different methods (i.e. SVM with all channels and SVM with MOBDE-based channel selection) for 5, 10 and 15 trials across all 9 subjects. The maximum classification accuracy values were used for statistical comparison. Under the test, the methods were
Statistical analysis
ranked based on the results for different subjects. The p-value was obtained as 4.48 108, which was much lesser than 0.05. Hence, the null hypothesis (there is no significant difference among the different methods) was rejected. Afterwards, a critical difference (CD) of 2.5135 was computed using a post hoc Nemenyi test [38]. The results of the post hoc test are depicted in Fig. 7. The average ranks of two methods with different trials are shown in increasing order on horizontal axis. Two methods connected by a colored line (below the horizontal axis) shows that the there is no significant difference between them.
365 366 367 368 369 370 371 372 373 374 375
5.
376
Discussions
From the statistical analysis presented in the previous section, it can be concluded that the classification accuracy increases with number of trials. The improvement in the system performance with respect to number of trials can be explained by Eq. (2), where the averaging over number of trials (j = 1 to J) is used to decide the target character. The averaging over the number of trials supresses the noise component and increases the SNR. However, increasing the number trials decreases ITR and user convenience. Hence, for better ITR and more userconvenience, it is required to decrease the number of channels in addition to increase the classification accuracy. In this regard and to further enhance the system efficiency, a MOBDE-based optimal channel selection method was
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
377 378 379 380 381 382 383 384 385 386 387 388 389
BBE 200 1–10
8
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Fig. 6 – The topographical maps showing the channel selection frequency in pareto-optimal solution points for each subject, separately.
390 391 392 393 394 395 396 397 398
proposed. The aim of the method was used to optimize the trade-off between the number of channels and classification accuracy. The effectiveness of the proposed technique can be verified by Fig. 7 which provides the statistical comparison of different approaches. From Fig. 7, it is also evident the average rank of the proposed MOBDE-based approach with 15 trials was 1, i.e. the approach performs best for all subjects. From Fig. 7, it can be seen that the average rank of MOBDE with 10 trials was better than the approach which used 15 trials of all
Fig. 7 – Visualization of post hoc Nemenyi test for showing the effectiveness of the proposed channel selection approach.
64 channels. This suggests that the proposed approach not only increase the reliability (by increasing the accuracy) and user convenience (by reducing the channels), but also increases the ITR by reducing the number of trials. It can also be observed from Fig. 6 that for given number of trials, there is no significant difference between the two approaches. Further, the proposed method provides a set of paretooptimal solutions for optimizing the trade-off between the number of channels and classification accuracy. So, based on the priority of objectives, a user can choose any solution from the pareto-front for optimal performance. For example, If minimizing number of channels is the main objective, then considering the average over the pareto-optimal solutions corresponding to the minimum number of channels (across all subjects), an average accuracy of 88% can be achieved with only 16 channels (Fig. 4). As the maximum accuracy of 87.6% was achieved without channel selection, it can be concluded that the proposed system with only 16 channels was able to produce better classification accuracy then a system which uses all 64 channel's data for feature extraction and classification.
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
BBE 200 1–10 biocybernetics and biomedical engineering xxx (2017) xxx–xxx
420 421 422 423 424 425 426 427 428 429 430 431 432 433
From the topographical maps shown in Fig. 5, it can be observed that the location of channel frequency is almost similar for 5, 10 and 15 trials. Additionally, taking the average over all the subjects, channel Pz(25), Oz(30), O2(31), CPz(53), P2 (58), PO7(60), PO4(63), and PO8(64) were selected in top 10 channel lists for 5, 10 and 15 trails. In other words, for majority of cases, the aforementioned channels comprised the optimal channel subset with significant information. The results obtained by the proposed algorithm are also in correspondence with the results of [24,25] for most informative channels. Hence, if it is not possible (or feasible) to apply channel selection methods for individual users, the channels from the occipital and parietal regions should be selected for more discriminative features in a P300-based BCI application.
434
6.
435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 465 464
A novel method is proposed to optimize the performance of DS-based P300 speller systems. An SVM-based classification method was used to detect the target characters on the dataset collected from 9 healthy subjects. Unlike most of the existing works on P300 spellers, which have only concentrated on improving the classification accuracy, motivated by the significance of reducing the number of channels for better user-convenience and improved ITR, the present work aims at optimizing the trade-off between the twin objectives of maximizing the classification accuracy and minimizing the number of channels. The application of MOBDE algorithm is proposed to solve the aforementioned multi-objective problem. The proposed algorithm provided a set of pareto-optimal solutions with various configuration of number of channels and corresponding accuracy values. With the pareto-front obtained after convergence of MOBDE, based on the requirements, the user gets a choice to select any optimal solution. The statistical analysis presented in the paper reflects that the proposed method not only increase the classification accuracy, but also increases the ITR and user-convenience by reducing the number of channels. Further, the proposed system with only 16 channels achieved higher classification accuracy than that achieved by a traditional method with all 64 channels. The concept of DS-based P300 speller presented in this paper was based on RC paradigm; exploration of differed types of paradigm such as CB paradigm is proposed as a future task on P300 spellers. It is also planned to concentrate on the application of other types of feature extraction techniques such as features from combined time-frequency domain and higher order statistics.
Conclusion
466 467
Acknowledgements
468 469 470 471 472 473 474 475
The authors would like to thank to the people associated with MILE Lab and Primates Research Lab of IISc, Bangalore, India, for providing us necessary support and facilities for recording the dataset. Authors acknowledge Department of Science and Technology, Government of India for financial support vide Reference No. SR/CSRI/38/2015 (G) under Cognitive Science Research Initiative (CSRI) to carry out this work.
9
references
476
[1] Farwell LA, Donchin E. Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalogr Clin Neurophysiol 1988 Dec;70:510–23. [2] Akcakaya M, et al. Noninvasive brain-computer interfaces for augmentative and alternative communication. IEEE Rev Biomed Eng 2014;7:31–49. [3] Wolpaw JR, et al. Brain-computer interfaces for communication and control. Clin Neurophysiol 2002;113:767–91. [4] Allison BZ, Pineda JA. ERPs evoked by different matrix sizes: implications for a brain computer interface (BCI) system. Neural Systems and Rehabilitation Engineering IEEE Transactions on 2003;11:110–3. [5] Sellers EW, et al. A P300 event-related potential brain– computer interface (BCI): the effects of matrix size and inter stimulus interval on performance. Biol Psychol 2006;73:242–52. [6] Allison BZ, Pineda JA. Effects of SOA and flash pattern manipulations on ERPs, performance, and preference: implications for a BCI system. Int J Psychophysiol 2006;59:127–40. [7] Salvaris M, Sepulveda F. Visual modifications on the P300 speller BCI paradigm. J Neural Eng 2009;6:046011. [8] Kaper M, et al. BCI Competition 2003–Data set IIb: support vector machines for the P300 speller paradigm. IEEE Trans Biomed Eng 2004;51:1073–6. [9] Hoffmann U, et al. An efficient P300-based brain-computer interface for disabled subjects. J Neurosci Methods 2008;167:115–25. [10] Krusienski DJ, et al. A comparison of classification techniques for the P300 Speller. J Neural Eng 2006;3:299–305. [11] Manyakov NV, et al. Comparison of classification methods for P300 brain-computer interface on disabled subjects. Comput Intell Neurosci 2011;2011:519868. [12] Cecotti H. Toward shift invariant detection of event-related potentials in non-invasive brain-computer interface. Pattern Recognition Letters 2015;66:127–34. [13] Chaurasiya RK, et al. An efficient P300 speller system for Brain-Computer Interface. Signal Processing, Computing and Control (ISPCC), 2015 International Conference on. 2015. pp. 57–62. [14] Bhatnagar V, et al. A modified approach to ensemble of SVM for P300 based brain computer interface. 2016 International Conference on Advances in Human Machine Interaction (HMI). 2016. pp. 1–5. [15] Chaurasiya RK, et al. Binary DE-Based Channel Selection and Weighted Ensemble of SVM Classification for Novel Brain–Computer Interface Using Devanagari Script-Based P300 Speller Paradigm. Int J Human–Comput Interaction 2016;1–17. [16] Guan C, et al. High performance P300 speller for braincomputer interface. Biomedical Circuits and Systems, 2004 IEEE International Workshop on; 2004. pp. S3/5/INVS3/13-16. [17] Fazel-Rezai R, Abhari K. A region-based P300 speller for brain-computer interface. Canadian J Electr Comput Eng 2009;34:81–5. [18] Townsend G, et al. A novel P300-based brain–computer interface stimulus presentation paradigm: moving beyond rows and columns. Clin Neurophysiol 2010;121:1109–20. [19] Blankertz B, et al. The Berlin Brain–Computer Interface: accurate performance from first-session in BCI-naive subjects. IEEE Trans Biomed Eng 2008 Oct;55:2452–62.
477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
Q2
BBE 200 1–10
10 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
[20] Xu M, et al. Channel selection based on phase measurement in P300-based brain-computer interface. PLoS ONE 2013;8:e60608. [21] Schröder M, et al. Robust EEG channel selection across subjects for brain-computer interfaces. EURASIP J Appl Signal Process 2005;2005:3103–12. [22] Colwell KA, et al. Channel selection methods for the P300 Speller. J Neurosci Methods 2014 Jul;232:6–15. [23] Speier W, et al. A method for optimizing EEG electrode number and configuration for signal acquisition in P300 speller systems. Clin Neurophysiol 2015;126:1171–7. [24] Gao W, et al. Multi-ganglion ANN based feature learning with application to P300-BCI signal classification. Biomed Signal Process Control 2015;18:127–37. [25] Jin J, et al. P300 Chinese input system based on Bayesian LDA. Biomed Tech (Berl) 2010 Feb;55:5–18. [26] Schalk G, et al. BCI2000: a general-purpose brain-computer interface (BCI) system. IEEE Trans Biomed Eng 2004;51: 1034–43. [27] Rakotomamonjy A, Guigue V. BCI competition III: dataset IIensemble of SVMs for BCI P300 speller. IEEE Trans Biomed Eng 2008 Mar;55:1147–54. [28] Kee C-Y, et al. Multi-objective genetic algorithm as channel selection method for P300 and motor imagery data set. Neurocomputing 2015;161:120–31. [29] Theodoridis S, Koutroumbas K. Pattern recognition. Fourth Edition. Academic Press; 2008.
[30] Storn R, Price K. Differential evolution-a simple and efficient adaptive scheme for global optimization over continuous spaces vol. 3. ICSI Berkeley; 1995. [31] Das S, Suganthan PN. Differential evolution: a survey of the state-of-the-art. IEEE Trans Evolut Comput 2011;15: 4–31. [32] Vesterstrom J, Thomsen R. A comparative study of differential evolution, particle swarm optimization, and evolutionary algorithms on numerical benchmark problems. Evolutionary Computation, 2004. CEC2004. Congress on. 2004. pp. 1980–7. [33] Pampará G, et al. Binary differential evolution. Evolutionary Computation, 2006. CEC 2006. IEEE Congress on. 2006. pp. 1873–9. [34] Wang L, et al. A novel modified binary differential evolution algorithm and its applications. Neurocomputing 2012;98:55–75. [35] Chen Y, et al. A binary differential evolution algorithm learning from explored solutions. Neurocomputing 2015;149:1038–47. [36] Rivet B, et al. Impact of spatial filters during sensor selection in a visual P300 brain-computer interface. Brain Topogr 2012 Jan;25:55–63. [37] Demšar J. Statistical comparisons of classifiers over multiple data sets. J Mach Learn Res 2006;7:1–30. [38] Nemenyi P. Distribution-free multiple comparisons. Biometrics 1962;263.
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
Available online at www.sciencedirect.com
ScienceDirect journal homepage: www.elsevier.com/locate/bbe 1 2 3
Original Research Article
Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller
4 5 6
7 8 9 10
Q1
Rahul Kumar Chaurasiya a,*, Narendra D. Londhe b, Subhojit Ghosh b a
Department of Electronics and Telecommunication Engineering, National Institute of Technology, Raipur, Raipur, PIN-492010, India b Department of Electrical Engineering, National Institute of Technology, Raipur, India
article info
abstract
Article history:
P300 speller-based brain-computer interface (BCI) allows a person to communicate with a
Received 5 October 2016
computer using only brain signals. In order to achieve better reliability and user continence,
Received in revised form
it is desirable to have a system capable of providing accurate classification with as few EEG
9 March 2017
channels as possible. This article proposes an approach based on multi-objective binary
Accepted 20 April 2017
differential evolution (MOBDE) algorithm to optimize the system accuracy and number of
Available online xxx
EEG channels used for classification. The algorithm on convergence provides a set of paretooptimal solutions by solving the trade-off between the classification accuracy and the
Keywords:
number of channels for Devanagari script (DS)-based P300 speller system. The proposed
BCI
method is evaluated on EEG data acquired from 9 subjects using a 64 channel EEG acquisition
Devanagari
device. The statistical analysis carried out in the article, suggests that the proposed method
Multi-objective Optimization
not only increases the classification accuracy but also increases the over-all system reliabil-
Binary DE
ity in terms of improved user-convenience and information transfer rate (ITR) by reducing
P300-speller
the EEG channels. It was also revealed that the proposed system with only 16 channels was
SVM
able to achieve higher classification accuracy than a system which uses all 64 channel's data for feature extraction and classification. © 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier B.V. All rights reserved.
13 11 12 14 15 16
17 18 19
1.
Introduction
P300 speller-based brain-computer interface (BCI) allows a person to communicate with a computer using only brain signals [1]. The communication does not require any physical
movements, and is particularly useful for patients suffering from severe motor disabilities but having cognitive abilities [2]. The most widely used P300 speller works in an odd-ball paradigm-based experimental environment [1,3]. In the oddball experiment, the subjects are randomly presented with two types of events, one of which rarely occurs (the odd-ball). The
* Corresponding author at: Department of Electronics and Telecommunication Engineering, National Institute of Technology, Raipur, Raipur-C.G, PIN-492010, India. E-mail addresses: [email protected] (R.K. Chaurasiya), [email protected] (N.D. Londhe), [email protected] (S. Ghosh). http://dx.doi.org/10.1016/j.bbe.2017.04.006 0208-5216/© 2017 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier B.V. All rights reserved. Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
20 21 22 23 24 25
BBE 200 1–10
2 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
rare events generates a P300 event related potential (ERP) in the recorded electroencephalogram (EEG). Farwell and Donchin first developed a P300 speller for English alphabetic script using a 6 6 matrix [1]. All rows and columns of the matrix were randomly intensified and the subject was asked to focus on the target character (that one, which he wants to communicate). The speller is commonly known as row/column (RC) paradigm-based speller. The multiple trails of the intensifications were repeated and the averaged signals were used to improve the signal-to-noise ratio (SNR). Several methods have been reported to further improve the system reliability and information transfer rate (ITR). The method includes improvement in display paradigm by changing matrix size, background color, font size, and inter stimulus interval [4–7]. Different classification methods such as SWLDA, Baysian linear discriminant analysis (BLDA), support vector machines (SVMs), and artificial neural networks (ANN), have been successfully applied for classification in P300 spellers [8–15]. Although the RC paradigm is still the most widely used paradigm for P300 spellers, experiments with single character (SC) paradigm [16], region-based (RB) paradigm [17], and check board (CB) paradigms [18] have also been tried in recent years. The performance of classifiers used for detection of P300 ERPs significantly depends on the choice of features and hence only the most discriminative features should be ideally used for classification. However, in the case of P300 spellers, the channel set providing the most relevant information varies from subject to subject [19,20]. For an EEG device having 64 EEG channels, there are total 264 possible subsets and practically it is impossible to select the best channel subset using exhaustive search. Hence, different channel selection methods such as channel selection using recursive channel elimination [21], jump-wise regression [22], Gibbs sampling [23], multi-ganglion [24], particle swarm optimization (PSO) [25] and genetic algorithm (GA), have been proposed for improving the classifier performance by selecting the best channel subset. Although, most of research work on the use of visual P300 spellers to aid communication has concentrated on languages that are written with English alphabetic script, a P300 speller systems capable of communicating text in Devanagari script (DS) has been developed in [15], which aimed at improving the system reliability by maximize the classification accuracy. A binary differential evolution (DE)-based optimization method was used for selection of channel subset with a single objective of achieving maximum classification accuracy. In the proposed study, we have extended the work of [15] for improving the user convenience and ITR, in addition to increasing the classification accuracy of DS-based P300 speller system. An improved ITR and user-convenience in addition to accuracy is expected to further improve the reliability of the system. The ITR and user-convenience are directly related to the number of channels used for data acquisition. In this regard, the present work aims at the multiple objectives of improved classification accuracy and reduced number of channels. Two different approaches are generally used for solving multi-objective optimization problem, the first approach combine all the objective functions into a single composite objective function by assigning different
weightages to different objectives, while the second approach determine an entire pareto-optimal solution set. Considering the limitations related to the selection of proper weightage function, which is not known a priori, the second approach is preferable as it provides a set of pareto-optimal solutions and user can decide about which solution he wants to use, based on the priorities of the objectives. In this article, a multiobjective binary DE (MOBDE) algorithm is proposed for finding the pareto-optimal solution set for solving the trade-off between number of channels and classification accuracy. Due to the requirement of lesser number of algorithm specific parameters and simplicity of the algorithm, MOBDE has been preferred over other optimization approaches. The rest of the paper is organized as follows: The description of signal acquisition procedure and the dataset is given in Section 2. The character detection mechanism is also described in Section 2. The framing of the multi-objective problem of accuracy maximization & channel minimization as an optimization problem and its solution using MOBDE algorithm is described in Section 3. The results are presented in Section 4. With discussions in Section 5, the study is concluded in Section 6.
2.
Materials and methods
2.1.
The dataset
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109
Total 9 healthy volunteer subjects of age range 21–29 (mean 26.5) were used to collect EEG responses. All subjects were able to communicate using DS. The EEG responses were recorded with a BrainAmp DC hardware which was equipped with a 64channel actiCAP. The EEG responses were collected at 500 Hz sampling frequency. At the time of recording, a digital bandpass filter of 1 to 250 Hz was also applied. General purpose BCI2000 software was used for stimulus presentation and data collection using DS-based display paradigm [26]. The DS paradigm used for stimulation is shown in Fig. 1(a) and the channel configuration used for data collection is shown in Fig. 1(b). Total 100 characters were presented in 20 runs as target characters for each subject. In a particular run, for each character, the rows and the column of the display matrix were randomly and successively intensified for 120 ms, the followed by 80 ms non-intensification period. In order to enhance the SNR of the acquired EEG signals, the sequences of intensifications of 16 rows and columns were repeated 15 times for each character. After 15 trials of one character, the recording for the next character was started with a gap of 10 s between the two characters. The 10 second gap was incorporated to make the subjects relax and to ensure that they can comfortably find the next target character in the display matrix. Total 240 responses (16 rows/columns 15 trials) were recorded per character. Two out of 16 responses were supposed to contain P300 ERPs.
110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
2.2.
136
Preprocessing and feature extraction
The methodology for preprocessing and feature extraction in the present work has been adopted from reported works on P300 speller [2,27,28]. The EEG samples posterior to 600 ms
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
137 138 139
BBE 200 1–10
3
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Fig. 1 – (a) The 8 T 8 matrix containing 13 vowels, 37 consonants, and 10 digits of DS. Total 4 special characters were also used in stimulus presentation. (b) A 64 channel configuration for acquiring EEG responses. Channels AFz and FCz were also used as ground and reference, respectively.
140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
from the starting of each flashing were extracted from all 64 channels. As the interest was to capture the P300 ERPs, the selected time window of 600 ms was sufficient to capture the relevant information. The extracted samples were passed through a band-pass filter of cut-off frequencies between 1 and 10 Hz. The filter samples were than decimated with a frequency of 10 Hz. At this stage, each EEG response consists of 6 samples per channel. For these samples, normalization was carried out independently for each channel. Afterwards, feature vectors were formed by concatenation of the signal samples of all 64 channels. Thus for a single character, there are total 240 feature vectors, each of size 384 (6 samples 64 channels). In these 240 feature vectors, 30 are from class +1 (1 row and 1 column per trial 15 trials) and are expected to contain P300 ERPs. The rest of the feature vectors are in class 1. Sample variations of the average of the class +1 and 1 EEG signals for all nine subjects for channel CPz are depicted in Fig. 2. A peak in class +1 signal near 300 ms shows that the P300 ERPs were properly captured in the oddball experiment.
159
2.3.
f ðxÞ ¼
(1)
171 172 173 174 175 176 177
Srjc
J 1X ð jÞ ¼ f xrjc J j¼1
(2)
where f(xr|c) is the score assigned to features xr|c of a give row/column (r/c), with J trials (J can be chosen to be between 1 and 15). The character which is common in predicted row and column is the target character.
3. Optimizing for accuracy and number of channels
Classification
The classification task of predicting the row and column containing the P300 ERP is a two-class classification problem. Because of its better generalization capability over other machine learning algorithms [29], an SVM classifier was employed for this task. From the training data points Xi, i = 1, 2 . . . ..N (with respective class labels yi =1 or + 1, i = 1, 2 . . . ..N), SVM learns maximizes-margin hyper-plane while also trying to minimize the total errors in classification. The learned separating hyper-plane then can be used to assign a score to a new test data point X is represented as:
170
i¼1
where li, i = 1, 2 . . . ..N are the Lagrange's multipliers. Since in the presented work 15 trials have been recorded for each character, the rows/column are predicted based on the scores obtained by different rows/columns, as per the formulation of Eq. (2).
3.1. 160 161 162 163 164 165 166 167 168 169
N X yi li ðX:Xi Þ þ b
Binary differential evolution
DE is a population-based evolutionary optimization method. As compared to other evolutionary approaches, it is simple yet effective algorithm and requires only two control parameters [30]. The population of DE consists of continues valued floating-point encoded vectors. The population at iteration 0 is generated as a group of NP random vectors. Mutation, crossover and selection operation are generally performed in DE for updating the positions for coming iterations. Suppose we have NP, D-dimensional target vector xti ; i ¼ 1; 2; . . .NP at
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
179 178 180 181 182 183 184 185 186
187 188 189 190 191 192 193 194 195
BBE 200 1–10
4
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
S2
Normalized Amp.
S1
1
1
0
0
0
-1
0
100
200
300
-1
0
100
200
300
1
1
0
0
0
0
100
200
300
-1
0
100
S7
200
300
200
300
0
200
300
100 200 Sample No.
300
100
1
Class +1 Class -1 0
0
100 200 Sample No.
-1
S9
1
0
100
S8
1
-1
0
S6
1
-1
Normalized Amp.
-1
S5
S4 Normalized Amp.
S3
1
300
-1
0
0
100 200 Sample No.
300
-1
0
Fig. 2 – Averaged Sample variations of recorded EEG signals (average over all the signals for first character) of class +1 and class S1 for channel CPz. A peak in class +1 signal near 300 ms shows that the P300 ERPs were properly captured in the oddball experiment. Class S1 signal are shown by dotted red.
196 197 198
iteration number t, then, in mutation stage, a mutation vector for next iteration is generated for each bit of every target utþ1 ij vector as
199
utþ1 ¼ xtr1;j þ F xtr2;j xtr3;j ij
(3)
200 201 202 203
where index of dimensionality j varies from 1 to D. F is a positive constant and xr1,j, xr2,j and xr3,j are three bits of randomly chosen individuals with indexes r1 6¼ r2 6¼ r3 6¼ i.
204 205 206
In crossover stage, a trial individual vi is generated by crossing over the target vector xi with the corresponding mutant vector ui as follows
207
( vtþ1 ij
¼
utþ1 ij ; xtij ;
if ðrand jCRÞorð j ¼ randðiÞÞ otherwise
better. Otherwise, the target individual xti is carried forward for next stage. Mathematically, the target individual for next iteration is selected as
(4)
208 209 210 211
where CR is the crossover probability between (0,1); rand j are stochastic random number uniformly distributed within ½0; 1Þ; rand(i) are random integers within 1, 2.., D.
212 213
In selection stage, the trial individual vtþ1 replaces the i , if its fitness value is target individual xti for generation of xtþ1 i
xtþ1 i
¼
tþ1 > f xti vtþ1 i ; if f vi xti ; otherwise
214 215 216 217
(5)
A number of improved variants of DE have been proposed in recent years [31]. It has also been reported that DE can perform better than other optimization for real world problems [32]. However, the standard DE operates in continuous space and is not suitable for solving binary optimization problems. Hence, its binary versions have also been proposed to solve such problems [33–35]. In BDE algorithms, each bit of target vector is represented by either a 0 or 1. The methodology used to update the population in BDE is similar to DE; involving crossover after mutation, and finally selection operations. In order to ensure that the target vector consists of only 0 s and 1 s, a probability estimation operator is used to generate mutant vector in the mutation stage as follows
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
219 218 220 221 222 223 224 225 226 227 228 229 230 231 232 233
BBE 200 1–10
5
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
234
8 1 tþ1 > > 3 ¼2 > P xij > 2bðMO0:5Þ > > < 6 1 þ 2F 7 41 þ e 5 > > > > > > : MO ¼ xt þ F xt xt r1;j
235 236 237 238 239 240
r2;j
Initialize 100 target vectors (6)
Decode target vector to obtain channel subset
r3;j
where b is a positive valued bandwidth factor. The mutant operator MO in BDE is analogous to the mutation operation used in standard DE as in Eq. (3). A binary-coded mutant vector uijtþ1 for target vector xtij is generated as ( uijtþ1 ¼
1;
if randð ÞP xtþ1 ij
0;
otherwise
Obtain the feature subset corresponding to selected channels
Train SVM with training data
Repeat 5 times for 5-fold CV
(7)
241 242 243 244 245 246 247 248 249 250
The crossover operation and selection process in binary DE are same as in Eqs. (4) and (5) respectively. The BDE has one more algorithm parameters (i.e. bandwidth factor b) in addition to two algorithm specific parameters F and CR used in DE. In this article, the method of Wang et al. has been adopted for cannel selection using MOBDE [34]. The values for algorithm specific parameters were chosen to be same as suggested in [34].
251
3.2.
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
In this article, the reduction in number of channels and improvement in the classification accuracy has been framed as a binary optimization problem in. A binary coded target vector in the population represents a set of channel. As the dataset was recorded from 64 channels, a 64-dimensional binary vector xi ¼ xi;1 ; xi;2 ; ::; xi;64 constituted the target vector. Total 100 target vectors were randomly generated as initial population. In target vector xi, at iteration number t, the features for classification are extracted from channel j (1 ≤ j ≤ 64), if xtij holds a value '1'. The update mechanism of the target vectors for the next iteration is same as described in Section 3.1. Each target vector in the population is a candidate solution and represents a possible combination of set of channels. The optimization process tries to search for a solution that maximizes the classification accuracy and at the same time minimizes the number of channels. In multi-objective optimization frame work, the two objective functions, viz. maximizing the classification accuracy and minimizing the number of channels has been mathematically formulated as Objective 1: maxA(xi) Objective 2: minC(xi)
273
s:t: xi x
274 275 276 277 278 279 280 281 282 283 284 285 286 287 288
In objective 1, A(xi) represents the accuracy of character detection, obtained from the SVM-based classification approach while choosing the channels for target vector xi. In objective 2, C(xi) is the count of number of 1 s present in target vector (xi). In Eq. (8), the constraint ‘‘xi x’’ indicates that target vectors xis contain a channel-set which is subset of x containing all 64 channels (i.e. all the entries in x are 1). The MOBDE algorithm was applied for each subject separately. It was observed that for each subject, there were no changes in the accuracy and number of channels after 40 iterations. Fig. 3 depicts the flow of the proposed methodology, involving SVM-based character detection and MOBDE-based channel selection method.
Obtain the classification accuracy with test data
Evaluate the fitness of the two objective functions
Update the target vectors using MOBDE algorithm
Find the pareto-optimal solutions
Multi-objective optimization framework
(8)
Iteration no.>40
No
Yes End Fig. 3 – Flow of the proposed methodology.
4.
Experimental results
289
In BCI, channel selection is usually an off-line process where the data acquired from all channels is used to select the optimal channel subset. Afterwards, the optimal channel subset is used for real-time applications. The contribution of the proposed MOBDE-based channel selection method for minimizing the number of channels and maximizing the classification accuracy has been evaluated in this section, which has been further divided into the following subsections: The classification accuracies obtained using SVM classifiers are presented in first subsection. The results of proposed MOBDE-based method for optimal channel selection have been presented in second subsection. The statistical analysis carried out in the third subsection reflects the effectiveness of the proposed channel selection method.
290 291 292 293 294 295 296 297 298 299 300 301 302 303
4.1.
Results for character classification using SVM
304
Table 1 depicts the accuracy of character detection obtained for 5, 10 and 15 trials using SVM classifier (for all 64 channels). A 5-fold cross validation methodology was adopted for training and testing the classifier. Average accuracy of 65.3%, 80.9% and 86.9% is obtained for 5, 10 and 15 trials, respectively.
305 306 307 308 309 310
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
BBE 200 1–10
6
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Table 1 – Percentage classification accuracy obtained using SVM classifier for the 5, 10 and 15 trials of the dataset for all 64 channels. Subject number
minimize the number of channel. The set of pareto-optimal solutions encountered by the target vectors have been shown as a pareto-front diagram in Fig. 4. The pareto-front diagram access the trade-off relationship between the number of channels and the classification accuracy. The pareto-front solutions shown in Fig. 4 represent a boundary at which an improvement of classification accuracy necessarily requires an increment in number of channels or, conversely, an attempt for reduction in the number of channels results in loss of accuracy. For 5, 10 and 15 trials, the classification accuracy obtained and the corresponding numbers of channel selected by the pareto-optimal solutions points with maximum accuracy are presented in Table 2. It can be observed from Fig. 4 that a total of 35 paretooptimal solution points were selected for 15 trials across all nine subjects. In these points, each of the 64 channels might have got selected between 0 to 35 times. Similarly, a total of 37 and 38 pareto-optimal solution points were selected for 5 and 10 trials, respectively. In order to analyze the locations of the channels that can provide more discriminatory information and better accuracy, the topographical map of channelfrequency among all the pareto-optimal solutions were plotted. Fig. 5 shows the topographical maps (considered over
Accuracy with different number of trials
1 2 3 4 5 6 7 8 9 Average
5
10
15
68 59 70 56 62 80 56 60 76 65.2
84 74 82 79 78 87 72 80 86 80.2
85 82 91 88 84 91 88 89 90 87.6
311
4.2.
312 313 314 315
The MOBDE was used with SVM-based character detection method for optimum channel selection using 5, 10 and 15 trials. As mentioned in Section 3.2, the optimization objectives were to maximize the accuracy of character detection and
Results for optimization using MOBDE
Subject 1
Subject 2
Channel numbers
64 visited places Pareto front
48
64 visited places Pareto front
48 32
32
16
16
16
75
80
85
90
95
0 100 70
75
85
90
95
0 70
visited places Pareto front
visited places Pareto front
48
32
16
16
16
80
85
90
95
100
0 70
75
Subject 7
80
85
90
95
visited places Pareto front
visited places Pareto front
48
32
16
16
16
80 85 90 Acc uracy (%)
75
95
100
0 70
100
80
85
90
95
100
75
80 85 90 Acc uracy (%)
95
100
visited places Pareto front
48
32
75
95
64
32
0 70
90
Subject 9
64
48
0 100 70
Subject 8
64
85
visited places Pareto front
48
32
75
80
Subject 6
32
0 70
75
64
64
48
100
Subject 5
Subject 4 64 Channel numbers
80
visited places Pareto front
48
32
0 70
Channel numbers
Subject 3
64
95
100 70
75
80 85 90 Acc uracy (%)
Fig. 4 – Illustration of the pareto-optimal solutions found by the MOBDE algorithmin for different subjects. The horizontal and vertical axes denote the classification accuracy and the number of selected channels, respectively. The 'visited places' points show the fitness of all the positions visited by the target particles. The 'pareto front' points represent the positions that belong to the pareto-optimal solution set. Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
BBE 200 1–10
7
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Table 2 – Percentage classification accuracy (A) and the corresponding number of channels selected (Nc) for 5, 10 and 15 trials obtained by pareto-optimal solution with maximum classification accuracy. Subject No.
Accuracy (A) and the corresponding number of channels selected (Nc) for 5, 10 and 15 trials obtained by pareto-optimal solution with maximum classification accuracy 5
1 2 3 4 5 6 7 8 9 Average
10
15
A
Nc
A
Nc
A
Nc
75 70 83 70 68 88 73 70 87 76.0
28 28 27 32 37 22 29 22 30 28.3
89 83 87 89 85 93 87 87 91 87.9
28 33 26 36 33 22 28 20 30 28.4
96 90 96 93 89 94 92 90 95 92.8
29 24 19 33 34 21 25 21 29 26.1
Fig. 5 – The topographical maps (common for all 9 subjects) showing the channel selection frequency in a total of 37, 38 and 35 pareto-optimal solution points for 5, 10 and 15 trials, respectively.
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
all 9 subjects) of the frequency of channel selection for 5, 10 and 15 trials. From Fig. 5, it can be observed that the parietal (P), occipital (O) and central (C) regions on scalp are most frequently selected. However, the optimal channel selection is a subject dependent problem i.e. there is inter subject variability in the optimal set [19,20,36]. Hence, in the present work, the topographical maps for individual subjects have also been analyzed. In order to ensure a sufficient number of data points are available for analysis, the pareto-optimal points obtained with 5, 10 and 15 trails were considered for each subject. As from Fig. 4, the total number of pareto-optimal points may vary in different solutions. After the experimentations for the present case, in total 110 pareto-optimal points (37 for 5 trials, 38 for 10 trials, and 35 for 15 trials) 12, 9, 13, 15, 8, 13, 14, 12 and 14 points were present in the pareto-optimal solution set of subject 1 to 9, respectively. The subject-wise topographical maps are shown in Fig. 6.
357
4.3.
358 359 360 361 362 363 364
In order to validate the statistical significance of the proposed method at different trials, the Friedman test was employed [37]. The test was applied for two different methods (i.e. SVM with all channels and SVM with MOBDE-based channel selection) for 5, 10 and 15 trials across all 9 subjects. The maximum classification accuracy values were used for statistical comparison. Under the test, the methods were
Statistical analysis
ranked based on the results for different subjects. The p-value was obtained as 4.48 108, which was much lesser than 0.05. Hence, the null hypothesis (there is no significant difference among the different methods) was rejected. Afterwards, a critical difference (CD) of 2.5135 was computed using a post hoc Nemenyi test [38]. The results of the post hoc test are depicted in Fig. 7. The average ranks of two methods with different trials are shown in increasing order on horizontal axis. Two methods connected by a colored line (below the horizontal axis) shows that the there is no significant difference between them.
365 366 367 368 369 370 371 372 373 374 375
5.
376
Discussions
From the statistical analysis presented in the previous section, it can be concluded that the classification accuracy increases with number of trials. The improvement in the system performance with respect to number of trials can be explained by Eq. (2), where the averaging over number of trials (j = 1 to J) is used to decide the target character. The averaging over the number of trials supresses the noise component and increases the SNR. However, increasing the number trials decreases ITR and user convenience. Hence, for better ITR and more userconvenience, it is required to decrease the number of channels in addition to increase the classification accuracy. In this regard and to further enhance the system efficiency, a MOBDE-based optimal channel selection method was
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
377 378 379 380 381 382 383 384 385 386 387 388 389
BBE 200 1–10
8
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
Fig. 6 – The topographical maps showing the channel selection frequency in pareto-optimal solution points for each subject, separately.
390 391 392 393 394 395 396 397 398
proposed. The aim of the method was used to optimize the trade-off between the number of channels and classification accuracy. The effectiveness of the proposed technique can be verified by Fig. 7 which provides the statistical comparison of different approaches. From Fig. 7, it is also evident the average rank of the proposed MOBDE-based approach with 15 trials was 1, i.e. the approach performs best for all subjects. From Fig. 7, it can be seen that the average rank of MOBDE with 10 trials was better than the approach which used 15 trials of all
Fig. 7 – Visualization of post hoc Nemenyi test for showing the effectiveness of the proposed channel selection approach.
64 channels. This suggests that the proposed approach not only increase the reliability (by increasing the accuracy) and user convenience (by reducing the channels), but also increases the ITR by reducing the number of trials. It can also be observed from Fig. 6 that for given number of trials, there is no significant difference between the two approaches. Further, the proposed method provides a set of paretooptimal solutions for optimizing the trade-off between the number of channels and classification accuracy. So, based on the priority of objectives, a user can choose any solution from the pareto-front for optimal performance. For example, If minimizing number of channels is the main objective, then considering the average over the pareto-optimal solutions corresponding to the minimum number of channels (across all subjects), an average accuracy of 88% can be achieved with only 16 channels (Fig. 4). As the maximum accuracy of 87.6% was achieved without channel selection, it can be concluded that the proposed system with only 16 channels was able to produce better classification accuracy then a system which uses all 64 channel's data for feature extraction and classification.
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419
BBE 200 1–10 biocybernetics and biomedical engineering xxx (2017) xxx–xxx
420 421 422 423 424 425 426 427 428 429 430 431 432 433
From the topographical maps shown in Fig. 5, it can be observed that the location of channel frequency is almost similar for 5, 10 and 15 trials. Additionally, taking the average over all the subjects, channel Pz(25), Oz(30), O2(31), CPz(53), P2 (58), PO7(60), PO4(63), and PO8(64) were selected in top 10 channel lists for 5, 10 and 15 trails. In other words, for majority of cases, the aforementioned channels comprised the optimal channel subset with significant information. The results obtained by the proposed algorithm are also in correspondence with the results of [24,25] for most informative channels. Hence, if it is not possible (or feasible) to apply channel selection methods for individual users, the channels from the occipital and parietal regions should be selected for more discriminative features in a P300-based BCI application.
434
6.
435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 465 464
A novel method is proposed to optimize the performance of DS-based P300 speller systems. An SVM-based classification method was used to detect the target characters on the dataset collected from 9 healthy subjects. Unlike most of the existing works on P300 spellers, which have only concentrated on improving the classification accuracy, motivated by the significance of reducing the number of channels for better user-convenience and improved ITR, the present work aims at optimizing the trade-off between the twin objectives of maximizing the classification accuracy and minimizing the number of channels. The application of MOBDE algorithm is proposed to solve the aforementioned multi-objective problem. The proposed algorithm provided a set of pareto-optimal solutions with various configuration of number of channels and corresponding accuracy values. With the pareto-front obtained after convergence of MOBDE, based on the requirements, the user gets a choice to select any optimal solution. The statistical analysis presented in the paper reflects that the proposed method not only increase the classification accuracy, but also increases the ITR and user-convenience by reducing the number of channels. Further, the proposed system with only 16 channels achieved higher classification accuracy than that achieved by a traditional method with all 64 channels. The concept of DS-based P300 speller presented in this paper was based on RC paradigm; exploration of differed types of paradigm such as CB paradigm is proposed as a future task on P300 spellers. It is also planned to concentrate on the application of other types of feature extraction techniques such as features from combined time-frequency domain and higher order statistics.
Conclusion
466 467
Acknowledgements
468 469 470 471 472 473 474 475
The authors would like to thank to the people associated with MILE Lab and Primates Research Lab of IISc, Bangalore, India, for providing us necessary support and facilities for recording the dataset. Authors acknowledge Department of Science and Technology, Government of India for financial support vide Reference No. SR/CSRI/38/2015 (G) under Cognitive Science Research Initiative (CSRI) to carry out this work.
9
references
476
[1] Farwell LA, Donchin E. Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalogr Clin Neurophysiol 1988 Dec;70:510–23. [2] Akcakaya M, et al. Noninvasive brain-computer interfaces for augmentative and alternative communication. IEEE Rev Biomed Eng 2014;7:31–49. [3] Wolpaw JR, et al. Brain-computer interfaces for communication and control. Clin Neurophysiol 2002;113:767–91. [4] Allison BZ, Pineda JA. ERPs evoked by different matrix sizes: implications for a brain computer interface (BCI) system. Neural Systems and Rehabilitation Engineering IEEE Transactions on 2003;11:110–3. [5] Sellers EW, et al. A P300 event-related potential brain– computer interface (BCI): the effects of matrix size and inter stimulus interval on performance. Biol Psychol 2006;73:242–52. [6] Allison BZ, Pineda JA. Effects of SOA and flash pattern manipulations on ERPs, performance, and preference: implications for a BCI system. Int J Psychophysiol 2006;59:127–40. [7] Salvaris M, Sepulveda F. Visual modifications on the P300 speller BCI paradigm. J Neural Eng 2009;6:046011. [8] Kaper M, et al. BCI Competition 2003–Data set IIb: support vector machines for the P300 speller paradigm. IEEE Trans Biomed Eng 2004;51:1073–6. [9] Hoffmann U, et al. An efficient P300-based brain-computer interface for disabled subjects. J Neurosci Methods 2008;167:115–25. [10] Krusienski DJ, et al. A comparison of classification techniques for the P300 Speller. J Neural Eng 2006;3:299–305. [11] Manyakov NV, et al. Comparison of classification methods for P300 brain-computer interface on disabled subjects. Comput Intell Neurosci 2011;2011:519868. [12] Cecotti H. Toward shift invariant detection of event-related potentials in non-invasive brain-computer interface. Pattern Recognition Letters 2015;66:127–34. [13] Chaurasiya RK, et al. An efficient P300 speller system for Brain-Computer Interface. Signal Processing, Computing and Control (ISPCC), 2015 International Conference on. 2015. pp. 57–62. [14] Bhatnagar V, et al. A modified approach to ensemble of SVM for P300 based brain computer interface. 2016 International Conference on Advances in Human Machine Interaction (HMI). 2016. pp. 1–5. [15] Chaurasiya RK, et al. Binary DE-Based Channel Selection and Weighted Ensemble of SVM Classification for Novel Brain–Computer Interface Using Devanagari Script-Based P300 Speller Paradigm. Int J Human–Comput Interaction 2016;1–17. [16] Guan C, et al. High performance P300 speller for braincomputer interface. Biomedical Circuits and Systems, 2004 IEEE International Workshop on; 2004. pp. S3/5/INVS3/13-16. [17] Fazel-Rezai R, Abhari K. A region-based P300 speller for brain-computer interface. Canadian J Electr Comput Eng 2009;34:81–5. [18] Townsend G, et al. A novel P300-based brain–computer interface stimulus presentation paradigm: moving beyond rows and columns. Clin Neurophysiol 2010;121:1109–20. [19] Blankertz B, et al. The Berlin Brain–Computer Interface: accurate performance from first-session in BCI-naive subjects. IEEE Trans Biomed Eng 2008 Oct;55:2452–62.
477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006
Q2
BBE 200 1–10
10 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568
biocybernetics and biomedical engineering xxx (2017) xxx–xxx
[20] Xu M, et al. Channel selection based on phase measurement in P300-based brain-computer interface. PLoS ONE 2013;8:e60608. [21] Schröder M, et al. Robust EEG channel selection across subjects for brain-computer interfaces. EURASIP J Appl Signal Process 2005;2005:3103–12. [22] Colwell KA, et al. Channel selection methods for the P300 Speller. J Neurosci Methods 2014 Jul;232:6–15. [23] Speier W, et al. A method for optimizing EEG electrode number and configuration for signal acquisition in P300 speller systems. Clin Neurophysiol 2015;126:1171–7. [24] Gao W, et al. Multi-ganglion ANN based feature learning with application to P300-BCI signal classification. Biomed Signal Process Control 2015;18:127–37. [25] Jin J, et al. P300 Chinese input system based on Bayesian LDA. Biomed Tech (Berl) 2010 Feb;55:5–18. [26] Schalk G, et al. BCI2000: a general-purpose brain-computer interface (BCI) system. IEEE Trans Biomed Eng 2004;51: 1034–43. [27] Rakotomamonjy A, Guigue V. BCI competition III: dataset IIensemble of SVMs for BCI P300 speller. IEEE Trans Biomed Eng 2008 Mar;55:1147–54. [28] Kee C-Y, et al. Multi-objective genetic algorithm as channel selection method for P300 and motor imagery data set. Neurocomputing 2015;161:120–31. [29] Theodoridis S, Koutroumbas K. Pattern recognition. Fourth Edition. Academic Press; 2008.
[30] Storn R, Price K. Differential evolution-a simple and efficient adaptive scheme for global optimization over continuous spaces vol. 3. ICSI Berkeley; 1995. [31] Das S, Suganthan PN. Differential evolution: a survey of the state-of-the-art. IEEE Trans Evolut Comput 2011;15: 4–31. [32] Vesterstrom J, Thomsen R. A comparative study of differential evolution, particle swarm optimization, and evolutionary algorithms on numerical benchmark problems. Evolutionary Computation, 2004. CEC2004. Congress on. 2004. pp. 1980–7. [33] Pampará G, et al. Binary differential evolution. Evolutionary Computation, 2006. CEC 2006. IEEE Congress on. 2006. pp. 1873–9. [34] Wang L, et al. A novel modified binary differential evolution algorithm and its applications. Neurocomputing 2012;98:55–75. [35] Chen Y, et al. A binary differential evolution algorithm learning from explored solutions. Neurocomputing 2015;149:1038–47. [36] Rivet B, et al. Impact of spatial filters during sensor selection in a visual P300 brain-computer interface. Brain Topogr 2012 Jan;25:55–63. [37] Demšar J. Statistical comparisons of classifiers over multiple data sets. J Mach Learn Res 2006;7:1–30. [38] Nemenyi P. Distribution-free multiple comparisons. Biometrics 1962;263.
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
Please cite this article in press as: Chaurasiya RK, et al. Multi-objective binary DE algorithm for optimizing the performance of Devanagari script-based P300 speller. Biocybern Biomed Eng (2017), http://dx.doi.org/10.1016/j.bbe.2017.04.006