Detection of P300 based on Artficial Bee Colony

Süleyman Abdullah Aytekin

and Tuba Kıyan

Department of Biomedical Engineering, Namik Kemal University, Tekirdag, Turkey

Department of Electronics and Communication Engineering, Yildiz Technical University, Istanbul, Turkey

Keywords: P300, Artificial Bee Colony, Brain-Computer Interface.

Abstract: A Brain-Computer Interface (BCI) is a system that allows users to communicate with their environment

through cerebral activity. P300 signal, which is used widely in BCI applications, is produced as a response

to a stimulus and can be measured in the parietal lobe of the brain. In this paper, an approach which is a

swarm intelligence technique, called Artificial Bee Colony (ABC) together with Multilayer Perceptron

(MLP) is used for the detection of P300 signals to achieve high accuracy. The system is based on the P300

evoked potential and is tested on four healthy subjects. It has two main blocks, feature extraction and

classification. In the feature extraction block, Power Spectrum Density (PSD) is used whereas ABC was

employed to train Multi Layer Perceptron (MLP) in the classification part. This method is compared to other

methods such as Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM). The best result

that is achieved in this work is 99.8%.

1 INTRODUCTION

A BCI is a system that allows a person to control or

communicate with a computer by using brain

signals. In various BCI systems,

electroencephalography (EEG) is still the most

common method because of being non-invasive. By

analyzing the EEG activities recorded from the

scalp, a computer can recognize the signal and

translate it to specific commands for output devices

such as a computer application or neuroprothesis in

order to accomplish user’s desired activity (Wolpaw

et al., 2002).

Infrequent or particularly significant visual

stimuli, when interspersed with frequent or routine

stimuli, evoke a positive peak at about 300 ms in the

EEG over parietal cortex. Donchin and et all have

used this ‘p300’ or ‘oddball’ response in a BCI

(Farwell and Donchin, 1988). Oddball paradigm

states that: “Rare expected stimuli produce a positive

deflection in the EEG after about 300 ms”. This

P300 component is present in nearly every human

(Rakotomamonjy and Guigue, 2008). P300 signal

can be seen in Figure 1. (Wolpaw et al., 2002).

They proposed a speller system based on P300,

which subjects were able to spell words by

sequentially choosing letters from the alphabet

(Serby et al., 2005). A 6x6 matrix containing the

letters of the alphabet and other symbols was

displayed on a computer screen. Rows and columns

of the matrix were flashed in random order. In order

to select a symbol, subjects count how often it was

flashed. Flashes of the row or column containing the

desired symbol evoked P300-like EEG signals,

while flashes of other rows and columns

corresponded to neutral EEG signals.

The target symbol could be inferred with a

simple algorithm that searched for the row and

column which evoked the largest P300 amplitude.

Since the work of Farwell and Donchin, many

researchers in the area of P300 based BCI systems

has concentrated on developing new application

scenarios (Polikof et al., 1995, Bayliss, 2003) and on

developing new algorithms for the detection of the

Figure 1: P300 signal (Wolpaw et al., 2002).

Aytekin, S. and Kiyan, T.

Detection of P300 based on Artﬁcial Bee Colony.

DOI: 10.5220/0005696001830189

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 4: BIOSIGNALS, pages 183-189

ISBN: 978-989-758-170-0

183

P300 signal from noisy data (Hoffmann et al., 2005,

Kaper et al.,, 2004, Rakotomamonjy et al., 2005;

Thulasidas et al., 2006, Xu et al., 2004).

In this paper, a pattern recognition system is used

for the detection of P300 signals. A five-choice P300

paradigm is tested. Five different images were

flashed in random order with a stimulus interval of

400 ms.

The MLP – ABC model is proposed for the

detection of P300 signals and compared to MLP,

Linear Discriminant Analysis (LDA) and Support

Vector Machine (SVM). Classification accuracies

were obtained for four healthy subjects. MLP trained

with the standard BP algorithm normally utilizes

computationally intensive training algorithms.

2 MATERIALS AND METHODS

A BCI takes EEG signal as input and delivers device

commands as output like any typical communication

system. Between input and output, there are

components that translate input into output.

Proposed BCI system with its elements and their

principal interactions is shown in Fig. 2 and each

element is described in this section in detail. Signal

processing and classification algorithms were

implemented by using MATLAB.

2.1 Data Acquisition

Four subjects are utilized in this study and they are

healthy college students whose ages are between 20

and 26. Three of them are male and the fourth

subject is female. Subjects were facing a screen on

which five images were displayed. The images are a

window shutter, a telephone, a lamp, a door, and a

television (see Fig. 3). The images were selected

according to an application scenario in which users

can control electrical appliances via a BCI system.

The images were flashed in random sequences, one

image at a time. Each flash of an image lasted for

100 ms and no images were shown during the

following 300 ms. One stimulus time is 400 ms.

Figure 2: Design and operation of the BCI system.

Figure 3: The display used for evoking the p300.

Total record time for each subject is five minutes. So

total flashes are 150 times and each image flashed

30 times in total record.

The EEG was recorded at 250 Hz sampling rate

from 64 electrodes which are placed at the standard

positions of the 10-20 international system. EGI

Geodesic EEG System 300 was used for

amplification and analog to digital conversion of the

EEG signals.

2.2 Preprocessing

In this step, in order to eliminate high and low

frequency noise, the signal is passed through a high

pass filter which has a cut off frequency of 0.1 Hz

and a low pass filter which has a cut off frequency of

40 Hz.

Another critical issue is to choose appropriate

channels. In order to avoid large number of channels

used for data recording and to avoid complicated

calculations, eight channels were used including

CFz, CP1, CP2, Fz, Pz, POz, P7 and P8 in this work

(Figure 4).

2.3 Feature Extraction

Before the classification, suitable features are

extracted from the raw signal. The goal of the

feature extraction is to remove noise and other

unnecessary information from the input signals. In

this paper, Power Spectral Density (PSD) method is

utilized which is generally used for characterizing

random processes. It can be calculated with the help

of Fourier Transform (Howard, 2002).

Discrete Fourier Transforms (DFTs) of filtered EEG

are calculated using Equation 1.

X(k) = x(n).e

-j2πkn/N

k=0,1,…N-1 (1)

In Equation 1, x(n) represents the discrete samples

of EEG data and N is the length of the EEG data.

After calculation of DFT (X(k)) of EEG samples,

periodogram of EEG is calculated by using Equation

2. Periodogram may express as a basic of PSD.

P(k) = |X(k)|

k=0,1,…,N-1 (2)

BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing

184

For each subject, the EEG signals are recorded for

five minutes. So, there are 75000 samples in five

minutes. We observed P300 signal presence between

200-400 ms after a stimulus is incident on the

computer screen. This interval expresses 50 data

instances. Data separated 1500 segments and each

segment have 50 data instances. Each segment of

PSD is average of own instances of periodograms

and it is calculated Equation 3.

’

(



)







∑





(



)





k=0,1,..,N-1 K=50 (3)

As a result, data sample size is reduced from 75000

to 1500 by PSD method. PSD of target samples and

non-target samples for eight channels are shown in

Fig. 4.

2.4 Classification

MLP-ABC scheme was proposed in this study for

classifying P300 and non-P300 signals. The optimal

weights of MLP were obtained by ABC algorithm

which is an optimization algorithm.

(ANN) model that maps sets of input data onto a set

of appropriate outputs. It was first introduced for the

non-linear XOR, and was then successfully applied

to different combinatorial problems.

In classification, we used MLP which has an input

layer, an output layer, and a hidden layer.

Input layer represents EEG channels which are

defined according to international 10-20 system.

The most commonly used three brain regions

(frontal, vertex and parietal) in the literature is

considered for selecting channels. There are eight

neurons (CFz, CP1, CP2, Fz, Pz, Poz, P7 and P8

channels) in the input layer of MLP which are

shown in Fig. 5 in the order of 4, 21, 41,6, 34, 36, 30

Figure 4: a) PSD of target b) PSD of non – target samples.

Figure 5: Electrode placement.

and 44. Each channel represent The output layer

represents presence of P300.

So, there are two possible outputs: P300 or non-

P300 (see Fig. 6). In this study, activation function

is defined as a log-sigmoid function and it is

calculated by using Equation 4.

()

−

(4)

where x represents the sum of input and weights

multiplication.

MLP is generally trained by back-propagation (BP)

algorithm. The training of a BP neural network is

carried out by the minimization of an error function.

The error function is defined as the difference

between the actual output and the desired output of

the ANN over a set of training patterns. The weights

assigned to the connections between the neurons

within the ANN are updated at each cycle to

minimize the error function. Training an ANN is an

optimization task since the objective is to find the

optimal set of weights of a neural network which

minimizes the error function.

On the other hand, there are some drawbacks of

using BP algorithm. If the multidimensional search

space has many local optima, BP may not reach the

global minimum successfully. Since, ABC algorithm

has the capability of searching the different

unidentified sections in the solution space to find out

the global minimum, it is used in this study to train

MLP (Shah and Gazali, 2011).

The main idea of using ABC algorithm is to

employ bees to search the best combination of

weights of the MLP network. The optimization task

Detection of P300 based on Artﬁcial Bee Colony

185

(a)

(b)

Figure 6: Voltage versus time graphs for a) Target (P300)

and b) Non – target (non – P300) signals.

by the ABC algorithm will involve the “bees”

searching for the optimal values of the weights.

Weights of the MLP structure are expressed as food

sources in ABC algorithm. The detailed pseudocode

of ABC algorithm is given in Section III.

3 ARTIFICIAL BEE COLONY

ALGORİTHM

Artificial Bee Colony algorithm (ABC) is an

optimization algorithm and it simulates the

intelligent foraging behaviour of honey bee colonies.

ABC is more successful and most robust on

multimodal functions included in the set with respect

to Differential Evolution (DE) algorithm (Yu and

He, 2006) and Particle Swarm Optimization (PSO)

(Eberhart et al., 2001).

In ABC algorithm (Karaboga and Akay, 2007,

Karaboga, 2005, Basturk and Karaboga, 2006,

Karaboga and Akay, 2009), the colony of artificial

bees consists of three groups: employed bees,

onlookers and scouts. The number of employed bees

is equal to the number of food sources.

The possible solution of the optimization problem in

ABC corresponds to the position of a food source

and the nectar amount of it is related to the quality of

the possible solution. At the first step, the ABC

generates a randomly distributed initial population

of SN solutions (food source positions), where SN

denotes the size of population. Each solution xi (i =

1, 2, ..., SN) is a D-dimensional vector where D is

the number of optimization parameters. After that,

the employed bees, the onlooker bees and scout bees

search the space for solutions in repeated cycles C =

1, 2, ..., MCN During the search process, an

employed bee finds a new food source (new

solution) and checks its nectar amount (fitness

value). If it has a higher nectar amount than that of

the previous one, the bee memorizes the new

position and erases the old one. If not, the position

of the previous one is retained in her memory. After

all employed bees complete the search process; they

share the information of the solution positions with

the onlooker bees. An onlooker bee evaluates the

nectar information taken from all employed bees and

selects a food source with a probability related to its

fitness value. Each onlooker bee chooses a food

source with the probability as given in (5).









∑









(5)

where SN is the population size and fit

is the fitness

value of the solution x

(6).

(6)

A new food source is produced by an employed bee

as given in (7).





=



+



(



−



) (7)

where 



is a random number within [-1,1]. After v

is produced, it is compared to x

solution. As can be

seen from (5), as the difference between the

parameters of the x

i,j

and x

k,j

gets smaller, the

perturbation on the position x

i,j

gets decrease, too.

Thus, as the search approaches to the optima in the

search space, the step length is adaptively reduced.

If the solutions can not be improved by employed

bees, they abandon the food source and it is replaced

BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing

186

with a new food source. Employed bees become

scouts. Scouts make random search in search space

by (8 )







=





+

(

0,1

)

(





−





) (8)

After each candidate source position v

i,j

is produced

and evaluated by the artificial bee, it is compared

with the old one. If the new food has equal or better

nectar than the old source, it is replaced with the old

one in the memory.

Detailed pseudo-code of the ABC algorithm is given

below:

1: Initialize the population of solutions xi,j , i = 1

...SN,j = 1 ...D

2: Evaluate the population

3: cycle=1

4: repeat

5: Produce new solutions υi,j for the employed bees

by using (2) and evaluate them

6: Apply the greedy selection process

7: Calculate the probability values Pi,j for the

solutions xi,j by (1)

8: Produce the new solutions υi,j for the onlookers

from the solutions xi,j selected depending on Pi,j

and evaluate them

9: Apply the greedy selection process

10: Determine the abandoned solution for the scout,

if exists, and replace it with a new randomly

produced solution xi,jby (3)

11: Memorize the best solution achieved so far

12: cycle=cycle+1

13: until cycle=MCN

The aim of the bees is to discover the places of food

sources with high nectar amount and finally the one

with the highest nectar. Artificial bees explore the

multidimensional space for food sources depending

on each bee’s experience and adjust their positions.

Every bee produces new solution area for the

network and the Greedy Selection decides the best

food source position. The food area is limited in the

range of [-20, 20]. If the new food source has equal

or better nectar than the old food source, it is

replaced with the new food source. Otherwise, the

old food source is kept in the memory. Employed

bees and onlooker bees continue searching until the

last cycle to find optimum weights. SN parameter is

selected 50 and MCN parameter is selected 100 in

this work.

4 ABC – MLP COMBINATION

The proposed ABC – MLP algorithm used to

optimize weights for training NN. The bees

(employee, onlooker) search until the last cycle for

finding best weights values for network training.

The food source of which the nectar is neglected by

the bees is replacaed by the scout bees. Every bee

(employee, onlooker) produce new solution for the

network and the greedy selection decide the best

foods source position. The food area is limited in

range apply the randomly and is initialized for

evaluation. This operation is defined by the equation

(7). Every bee produce new evaluated solution area

for the network training. The food contains three

parameters such as range, foods number and

dimension which shown in equation 9, 10 and 11

(Shah and Gazali, 2011).

Foods Area = rand(FN,D).* Range + Lower (9)

(10)

(11)

where a

m*n

shows the range of colony which is equal

to upper bound minus lower bound, FN shows the

number of food source equals the half of the colony

size, L shows the lower bound for the foods area.

The upper and lower bounds are optional while the

most commonly used range is [10, -10], D shows

dimension of foods. The agent of ABC select the

value from foods and is assigned to network. The

bees select the weight from the foods area and

evaluate for fitting to input (Shah and Gazali, 2011).

We hope you find the information in this template

useful in the preparation of your submission.

5 RESULTS

In this study, data instances are reduced from 75000

to 1500 using PSD as mentioned in Section II. Then,

samples are separated as target (150x8) and non-

target (1350x8). Before classification, training

(1000x8) and test (500x8) datasets are created.

Training dataset includes 100 target and 900 non-

target instances whereas test dataset include 50

Detection of P300 based on Artﬁcial Bee Colony

187

target 450 non-target instances. The learning and

momentum rate for MLP are chosen to be 0.5. There

are five neurons in the hidden layer. Some

parameters of the MLP network are shown in Table

I. The performance of the proposed BCI system is

demonstrated in Table II. MLP-ABC method is

compared with other classification algorithms such

as MLP-BP, LDA and SVM in order to demonstrate

its performance. Discriminant type of LDA is

Table 1: Architecture of the MLP Network.

Parameters

Neurons in the input layer

Neurons in the hidden layer

Neurons in the output layer

Transfer function in the hidden layers

Logsig

Table 2: Classification results.

Methods Subjects

Subject

ABC + MLP 99.85 99.06 99.59 99.84

MLP 92.75 92.83 92.83 92.83

LDA 93.19 91.27 91.58 91.74

SVM 92.6 92.83 92.83 92.83

‘linear’ which estimates one covariance matrix for

all classes. Kernel function of SVM is ‘polynominal’

which default order is 3.

For all subjects, ABC+MLP approach gives a better

result than other methods.

6 CONCLUSION

The aim of this study is to detect P300 signals by

employing PSD for feature extraction and MLP-

ABC scheme as a classifier. BP is a common

approach for training MLP. The ABC algorithm

combines the exploration and exploitation processes

successfully, which proves the high performance of

training MLP for P300 classification. It has the

powerful ability of searching global optimal

solution. The simulation results show that the

proposed MLP-ABC algorithm can successfully

classify P300 data comparing with the traditional BP

algorithm and some classification algorithms that

include LDA and SVM. It was shown that MLP-

ABC approach shows significantly higher accuracy

in classification than the other methods.

ACKNOWLEDGEMENTS

This research has been supported by Yildiz

Technical University Scientific Research Projects

Coordination Department. Project Number is 2014-

04-03-KAP01.

REFERENCES

Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,

Pfurtscheller, G., Vaughan, T. M., 2002. Brain-

computer interfaces for communication and control,

Clin. Neurophysiol. Vol. 113, Issue 6, pp. 767-791.

Farwell L.A., Donchin E., 1988. Talking of top of your

head: Toward a mental prosthesis utilizing event-

related brain potentials, Journal of Electroensephalogr.

Clin. Neurophy,, Vol. 70, pp. 510–523.

Rakotomamonjy, A., Guigue, V., 2008. BCI competition

III: dataset II-ensemble of SVMsfor BCI P300 speller,

IEEE Trans. Biomed. Eng., vol. 55, no. 3, pp. 1147 –

1154.

Serby H., Yom-Tov, E., Inbar G.F., 2005. An improved

P300-based brain-computer interface, IEEE Trans. on

Neural Systems and Rehabilitation Eng., Vol. 13,

Issue 1, pp. 89–98.

Polikoff, J., Bunnell, H., Borkowski, W., 1995. Toward a

P300-based computer interface, Proceedings of the

RESNA '95 Annual Conference.

Bayliss, J. D., 2003. Use of the evoked P3 component for

control in a virtual apartment, IEEE Trans. Neural

Syst. Rehab. Eng., Vol. 11, Issue 2, pp. 113-116.

Hoffmann, U., Garcia, G. N., Diserens, K., Vesin,J.-M.,

Ebrahimi, T., 2005. A boosting approach to P300

detection with application to brain-computer

interfaces, Proceedings of the IEEE EMBS Neural

Engineering Conference, pp. 97-100.

Kaper, M., Meinicke, P., Grosskathoefer, U.,Lingner, T.,

Ritter, H., 2004. Support vector machines for the P300

speller paradigm, IEEE Trans. Biomed. Eng. Vol. 51,

Issue 6, pp. 1073-1076.

Rakotomamonjy, A., Guigue, V., Mallet, G., Alvarado, V.,

2005. Ensemble of SVMs for improving brain-

computer interface P300 speller performances,

Proceedings of International Conference on Neural

Networks (ICANN), pp.45-50.

Thulasidas, M., Guan, C.,Wu, J., 2006. Robust

classification of EEG signal for brain-computer

interface, IEEE Trans. Neural Syst. Rehab. Eng. Vol.

14, Issue 1, pp. 24-29.

Xu, N., Gao, X., Hong, B., Miao, X., Gao, S., Yang, F.,

2004. BCI competition 2003 Data Set IIb: Enhancing

P300 wave detection using ICA-based subspace

projections for BCI applications, IEEE Trans. Biomed.

Eng., Vol. 51 Issue 6, pp. 1067-1072.

Howard, R. M., 2002. Principles of Random Signal

Analysis and Low Noise Design: The Power Spectral

BIOSIGNALS 2016 - 9th International Conference on Bio-inspired Systems and Signal Processing

188

Density and its Application, Wiley-IEEE Press, pp.

59-60.

Shah, H., Gazali R., 2011. Prediction of Earthquake

Magnitude by an Improved ABC-MLP, Developments

in E-systems Engineering (DeSE), pp. 312-317.

Yu, B., He, X., 2006. Training Radial Basis Function

Networks with Differential Evolution, IEEE

International Conference, pp. 369–372.

Eberhart, R.C., Shi, Y., Kennedy, J., 2001. Swarm

Intelligence, Morgan Kaufmann.

Karaboga, D., Akay B., 2007. Artificial Bee Colony

(ABC) Algorithm on Training Artificial Neural

Networks, Signals Processing and Communications

Applications.

Karaboga, D., 2005. An Idea Based On Honey Bee Swarm

For Numerical Optimization, Technical Report-TR06,

Erciyes University, Engineering Faculty, Computer

Engineering Department.

Basturk, B., Karaboga, D., 2006. An Artificial Bee Colony

(ABC) Algorithm for Numeric function Optimization,

IEEE Swarm Intelligence Symposium.

Karaboga, D., Akay, B., 2009. A comparative study of

Artificial Bee Colony algorithm, Applied Mathematics

and Computation, Vol. 214, Issue 1, pp.108–132.

Detection of P300 based on Artﬁcial Bee Colony

189