Guess the Number - Applying a Simple Brain-Computer Interface to

School-age Children

Luk´aˇs Vaˇreka

, Tom´aˇs Prokop

, Jan

Stˇebet´ak

and Roman Mouˇcek

NTIS - New Technologies for the Information Society, University of West Bohemia, Univerzitni 8, Pilsen, Czech Republic

Department of Computer Science and Engineering, University of West Bohemia, Univerzitni 8, Pilsen, Czech Republic

Keywords:

Electroencephalography, Event-related Potentials, Brain-Computer Interface, P300, Discrete Wavelet Trans-

form, Multi-layer Perceptron.

Abstract:

Although research into brain-computer interfaces is more common in recent years, studies concerning large

groups of speciﬁc subjects are still lacking. This paper describes a simple brain-computer interface (BCI)

experiment that was performed on a group of over 200 school-age children using the technique and methods

of event related potentials. In the ﬁrst phase, experimental data were recorded in various elementary and

secondary schools, mainly in the Pilsen region of the Czech Republic. The task was to guess the number

between 1 and 9 that the measured subject thinks on. Concurrently, a human expert made a decision about

the target number based on averaged P300 waveforms observed on-line. In the second phase, an application

for automatic classiﬁcation was developed for off-line data. A small subset of the data was used for training;

the rest of the data was used to evaluate the accuracy of classiﬁcation. Two feature extraction methods were

compared; subsampling and discrete wavelet transform for feature extraction. Multi-layer perceptron was used

for classiﬁcation. The human expert achieved the accuracy of 67.6%, while some of the automatic algorithms

were able to signiﬁcantly outperform the expert; the maximum classiﬁcation accuracy reached 77.2%.

1 INTRODUCTION

Recent advances in cognitive neuroscience and brain

imaging techniqueshavestarted to provide us with the

ability to interface directly with the human brain. In

brain-computerinterfaces (BCIs), instead of using the

brain’s normal output pathways, users explicitly try to

manipulate their brain activity to produce signals that

can be used to control computers. Any BCI has input

(e.g. electrophysiological activity from the user), out-

put (i.e. device commands), components that trans-

late input into output, and a protocol that determines

the operations. The electroencephalographic signal

is widely used in BCI systems as input because of

low cost of the device and simple use. This signal

is acquired by electrodes on the scalp and processed

to extract speciﬁc signal features (e.g. amplitudes of

evoked potentials) that reﬂect the decision of the user.

These features are translated into commands that op-

erate a device (e.g. a simple word processing pro-

gram). The user must follow the protocol of the BCI

system and maintain attention of focus. Then the BCI

system must select and extract features that the user

can control and must translate those features into de-

vice commands correctly and efﬁciently. (McFarland

and Wolpaw, 2011)

The technique of event-related potentials (ERPs)

uses the electroencephalographic signal enriched by

markers denoting the time points of brain stimulation

precisely synchronized to the brain activity. The P300

(depicted in Fig. 1) waveform is a cognitive event-

related potential with a positive low amplitude. It is

usually obtained with the oddball paradigm, which is

based on random occurrence of rare (target) stimuli

in sequence of frequent (non-target) stimuli. Because

of the fact that the P300 is a cognitive reaction to out-

side events, manybrain-computerinterfaces are based

on the P300 detection. However, the detection of the

P300 is challenging because the P300 component is

usually hidden in the underlying EEG signal. (Luck,

2005)

Many papers report different approaches for the

P300 BCIs, however, it is difﬁcult to compare them

directly because they use data recorded from differ-

ent laboratories and different subjects. Furthermore,

various paradigms that differ in many parameters, in-

cluding inter-stimulus intervals and number of trials

averaged, are used. However, there is a benchmark

reka, L., Prokop, T., Št

ebeták, J. and Mou

cek, R.

Guess the Number - Applying a Simple Brain-Computer Interface to School-age Children.

DOI: 10.5220/0005801402630270

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 4: BIOSIGNALS, pages 263-270

ISBN: 978-989-758-170-0

263

Figure 1: Comparison of averaged EEG responses to non-

target stimuli (Xs) and target stimuli (Os). There is a clear

P300 component following the Os stimuli. Negative is plot-

ted upward (Luck, 2005).

P300 speller dataset from the BCI Competition 2003

(Blankertz et al., 2004) and some papers report results

achieved on this dataset. Several approaches were

able to reach 100% accuracy using only 4-8 averaged

trials on the BCI Competition 2003 data (Cashero,

2012). To the best knowledge of the authors of this

paper, the BCI Competition 2003 datasets as well as

other datasets used in P300 BCI publications are rel-

atively small or not publicly available.

The aim of this paper is to propose a simple BCI

and to compare different P300 detection techniques

by testing them on a large dataset containing more

than 200 records from school-age children. The data

and metadata used in this article are freely avail-

able on the EEG/ERP portal in the package named

“PROJECT DAYS P3 NUMBERS” (Moucek and

Jezek, 2009). A “guess the number” experiment was

used for this purpose. The goal of the experiment is to

ﬁgure out the number chosen by a subject. Typically,

different classiﬁcation methods in BCI are evaluated

and comparedin terms of classiﬁcation accuracy. Fur-

thermore, because the numbers were also guessed by

an expert in the ﬁeld, a unique opportunity arises to

compare both accuracies.

2 EXPERIMENTAL DESIGN

The guess the number experiment was originally de-

veloped to demonstrate the beneﬁts of using BCI to

public. The experiment is based on visual stimula-

tion. The subject is asked to secretly choose a num-

ber between 1 and 9 and to concentrate on this number

(i.e. the target stimulus). The subject told the experi-

menters the number they were thinking of at the end

of the experiment.

2.1 Environment

The experiments were conducted in elementary and

secondary schools mainly in the Pilsen region, the

Czech Republic between autumn 2014 and spring

2015. The measurements were taken at the time of

regular school hours, typically in the morning. Each

experiment was performed in a classroom that was

dedicated for health entertaining and educating pro-

gramme, also including Neurosky brain games, ECG

monitoring, modeling of body muscles, etc. Unfortu-

nately, the environment was usually quite noisy since

many children were in the room at the same time.

2.2 Stimulation Protocol

The participants were stimulated with numbers ﬂash-

ing on the monitor in random order. The interstimu-

lus interval was 1500 ms. The ﬂashing number were

white on the black background as shown in Figure

2. The subjects were sitting approximately 1 m from

the monitor for as long as needed (approximately 10

minutes on average, stopped when the experimenters

were convinced that they are able to guess the num-

ber thought). They were asked to sit comfortably, pay

attention to the stimulation, not to move, and to limit

their eye blinking. To increase alertness, the subjects

were instructed to silently count the occurrences of

target stimuli.

2.3 Hardware and Software

The mobile EEG laboratory was used. It was nec-

essary to have equipment that was easy to unpack, in-

stall, and pack. Consequently, the following hardware

devices were used: a standard small or medium 10/20

EEG cap, the BrainVision standard V-Amp ampli-

ﬁer, standard electrodes, electro gel, conductivepaste,

and degreasing gel. To speed up the guessing task,

only three electrodes were active: Fz, Cz, and Pz.

These electrodes are signiﬁcant for the P300 detec-

tion (Luck, 2005).

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

264

Figure 2: Numbers 1 - 9 were randomly shown on the mon-

itor.

2.4 Measured Subjects

Most subjects were school-age children (average age

13.2): 135 males and 104 females. The total number

of subjects of different age with their gender distribu-

tion is shown in Figure 3.

3 GUESS THE NUMBER -

APPLICATION FOR ON-LINE

AND OFF-LINE BCI

CLASSIFICATION

An application for analysis of the experiments previ-

ously described was developed. It is a desktop ap-

plication written in Java language using Swing for its

graphical user interface (a screenshot is shown in Fig-

ure 4). The purposeof this applicationis to enable off-

line (experimental data are collected and analyzed af-

ter an experiment is performed) and on-line (data are

streamed into the application during an experiment)

classiﬁcation. Off-line classiﬁcation allows users to

test preprocessing, feature extraction, and classiﬁca-

tion algorithms. Subsequently, suitable combinations

can be selected for on-line classiﬁcation.

4 PATTERN RECOGNITION

Traditional pattern recognition was used: feature ex-

traction was followed by classiﬁcation. Two com-

monly used methods for feature extraction, band-

pass ﬁltering with subsampling of the feature vector

and discrete wavelet transform, were tested. In both

cases, data from three EEG channels were included.

A multi-layer perceptron was used for classiﬁcation.

The workﬂow of pattern recognition is shown in Fig-

ure 5.

4.1 Preprocessing and Feature

Extraction

The raw input data were measured with the sampling

frequencyof 1 kHz. However, continuous EEG (with-

out synchronization marks) is not suitable for ERP

detection. Therefore, the data were split into epochs.

Each epoch started 100 ms before the stimulus and

ended 750 ms after the stimulus. The pre-stimulus

interval was used only to correct the baseline (by sub-

tracting the mean in the pre-stimulus interval from the

whole epoch). Subsequently, the reduction of dimen-

sionality was needed.

The input for further feature extraction methods

involves only a part of each epoch because the P300

component can occur only in a speciﬁc interval after a

stimulus. The number of samples in each epoch (see

Table 1) was skipped (refered to as Skip samples).

The next 512 samples (suitable for DWT and related

to the area of occurence of the P300 component) were

selected for feature extraction. Subsequently, sub-

sampling or discrete wavelet transform (DWT) was

used to reduce input data dimension and extract fea-

tures. The subsampling method reduces data dimen-

sion by a given subsampling factor; this was set to

32. Before skipping out samples, in half of the clas-

siﬁcation experiments, high (16 Hz and higher) and

low (0.4 Hz and lower) frequencies were ﬁltered from

baseline corrected epochs using Chebyshev 1 IIR ﬁl-

ter (Chebyshev ﬁlters minimize the error between the

idealized and the actual ﬁlter characteristic). Sub-

sampled epochs were stored into 1-dimensional fea-

ture vector f. The second method used for feature

extraction was DWT since the method was sucess-

fully used for P300 processing (Quiroga and Garcia,

2003). The basic idea of wavelet transform is to de-

compose the input signal into a set of basis functions

called wavelets (Letelier and Weber, 2000). This is

done by scaling and dilatation of a prototype wavelet

called the mother wavelet by the following equation:

a,b

(t) =

√

ψ(

t−b

)

where ψ is the analyzing wavelet, a is the scaling

factor, and b is the time shift. In DWT the dilated and

translated version of the wavelet function performs a

high-pass and low-pass ﬁltering respectively. Hence

the convolution of the original signal with the high-

pass ﬁlter produces the Detail coefﬁcients. Similarly,

approximation coefﬁcients are computed by a con-

volution of the signal with a low-pass ﬁlter.(Markazi

et al., 2006)

Daubechies 8 mother wavelet was used to ex-

Guess the Number - Applying a Simple Brain-Computer Interface to School-age Children

265

Figure 3: The total number of subjects of different age with their gender distribution.

Figure 4: Front end of the Guess the number application after successful classiﬁcation. The left column represents the

stimulus, the stimulus average score is in the center column and the last column represents the number of epochs for each

stimulus. Rows are descendingly ordered by the stimulus score. The higher the score the more likely the stimulus number is

the number thought. The stimulus with the highest score is displayed in the right panel.

tract features from the input signal. The Daubechies

wavelets have been recently used for EEG signal pro-

cessing (Raﬁee et al., 2011). After DWT is per-

formed, 16 approximation coefﬁcients of level 5 for

each channel is stored in 1-dimensional array f and

used for classiﬁcation. Figure 6 shows how DWT co-

efﬁcients are obtained from input data.

Each DWT coefﬁcient or output sample of the

subsampling method f

is then normalized by the fol-

lowing formula:

= f

∗

m−1

∑

i=0

where m is the size of the feature vector. Figure 7

shows the structure of the feature vector.

4.2 Classiﬁcation

A multi-layer perceptron (MLP) was used for classi-

ﬁcation. The number of input neurons corresponded

to the dimension of the feature vector (it means 48).

The number of neurons in the middle layer was em-

pirically optimized to eight. One output neuron was

taughted to classify input patterns into two classes (0

- non-target, 1 - target).

4.2.1 Training

The ﬁrst idea was to use the data from three-stimulus

pattern that is another odd-ball paradigm (Vareka

et al., 2014). Only the epochs containing a small per-

centage of artifacts and signiﬁcant P300 components

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

266

Start

Channel

separation

Epoch

extraction

Baseline

correction

Selection of

epoch

subinterval

Bandpass

filtering

Feature

extraction

method

selection

Subsampling

DTW

computation

DWT

Vector

normalization

Feature

Vector

Classification

Raw EEG

data

Subsampled

data

DWT

coefficients

Score computation

and averaging

YES Filter data

Subsampling

Figure 5: Diagram of EEG signal processing, feature extraction and classiﬁcation.

Guess the Number - Applying a Simple Brain-Computer Interface to School-age Children

267

cA5

(16)

cD5

(16)

cD4

(32)

cD3 (64)cD2 (128)cD1 (256)

Epoch (512 samples)

cA1

cD1

cD2

cA5cA2

cD4

cA4cA3

cD3 cD5

Figure 6: Discrete wavelet transform coefﬁcients. Input EEG signal has 512 samples. The number of coefﬁcients obtained by

DWT is in brackets. 5-level DWT was performed. cA1 - cA5 represent approximation coefﬁcients of different levels, cD1-

cD5 represent detail coefﬁcients. Bold marked (cA5) coefﬁcients form the feature vector.

Pz Cz Fz

features

Feature

extraction

Preprocessing

Figure 7: Feature vector.

were added to the training set. Surprisingly, the ac-

curacy on the testing set was below 50% with these

training data, the reasons requirefurther investigation.

It seems, however, that differences in the latencies be-

tween three-stimulus datasets and datasets described

in this paper may explain this to some extent.

Instead, the training set containing 13 datasets was

selected from the guess the number data (255 target

epochs and 255 non-target epochs were used in total)

according to the following procedure. First, the data

were split into several groups using visual inspec-

tion: high impedanceat scalp electrodes (8 datasets

severely distorted by high frequency (12 datasets),

moderately distorted by high frequency (30 datasets),

clean (107 datasets), damaged by biological artifacts

(82 datasets), and missing metadata (5 datasets - not

included in the statistics). From the group contain-

ing clean data, 13 datasets (more precisely all target

epochs and the same number of randomly selected

non-target epochs from 13 datasets) formed the train-

ing set.

Before the training phase, the weights of the neu-

ral network were randomized. The feature vectors

were shufﬂed to mix target and non-target patterns.

For training, backpropagation was used. 20% of the

training features (derived from epochs) were used as

a validation set. The training was stopped when the

accuracy peaked on the validation set.

4.2.2 Testing

The testing set was formed by excluding the datasets

with high impedance, severely distorted datasets, and

datasets with missing metadata. Then the testing set

contains 206 datasets in total. The testing set is used

for testing the multi-layer perceptron (MLP) when

the training phase is ﬁnished. The MLP receives

an epoch and its expected classiﬁcation class (target,

non-target). The MLP calculates a score in range be-

tween 0 and 1. The higher score the more likely the

dataset means all the data from one measured subject

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

268

epoch belongs to the target number. Scores to each

number are summed. At the end of the classiﬁcation

of each number, all scores are averaged. The number

with the highest averaged score is the winner.

5 RESULTS

Table 1 shows average and maximum classiﬁcation

accuracy of Subsampling and DWT feature extraction

methods with different settings of preprocessing. The

classiﬁcation accuracy of the MLP classiﬁer highly

depends on the initial settings of weights and it may

differ in each trial (the whole testing set is used). Fur-

thermore, since 20% randomly selected epochs were

excluded from the training set for validation, the train-

ing set was different for each training phase. We aver-

aged at least 50 trials (classiﬁcations) for each feature

extraction settings to get average classiﬁcation accu-

racy.

5.1 Comparison with Human Expert

Classiﬁcation

All the data from the testing sample were evaluated

by a human expert with neuroinformatics background

- 67.61%accuracywas achieved. The expert observed

epochs as they were gradually averaged for each stim-

ulus (guessed number) in the BrainVision Recorder

(BrainProducts, 2012) software and searched for the

most likely target. Each experiment ended either

when the expert was convinced about the number

guessed or was unable to make the decision. As it can

be seen in Table 1, some of the developed classiﬁca-

tion algorithms were able to outperform the expert.

6 DISCUSSION

The parameter settings affected the resulting accu-

racy. For example, the selected time intervals need to

be in the time domain in which the P300 component

is expected. The latency of the P300 varied greatly

- it was typically around 450 ms but for some sub-

jects, it was 700 ms. Consequently, increasing the

skipped samples was associated with higher accuracy.

As Table 1 shows, the best results were achieved for

skipped samples between 150 and 200.

Filtering does not seem to lead to higher classiﬁ-

cation accuracy. This could be caused by a fact that

ﬁltering distorts the signal and could damage some

signiﬁcant features. Neural network seems to be able

to ignore irrelevant frequencies by itself.

In all cases, the empirically set dimensionality of

feature vectors was 48 (16 features for each channel).

Both higher and lower dimensionality did notimprove

classiﬁcation accuracy.

7 CONCLUSION

We proposed a BCI system based on visual stimula-

tion; the system was adjusted to school-aged children.

The aim of this BCI is to detect the P300 component

in the EEG signal and decide which number the sub-

ject thinks on. The overall dataset consists of 239

measurements from different subjects with average

age 13.2. The training set is formed from 13 datasets

and testing set contains 206 datasets. The rest of the

collected data were excluded because the signal was

highly damaged. For feature extraction we used sub-

sampling and DWT feature extraction methods that

we tested with different settings. The selection of the

optimal epoch subinterval has a high effect on clas-

siﬁcation accuracy (increased by up to 13.8%) while

the band-pass ﬁltering has not. Using the multi-layer

perceptron classiﬁer and feature extraction method,

68.9% average and 77.2%maximumclassiﬁcation ac-

curacy was achieved. The classiﬁcation accuracy of

human expert is 67.6%. It means that in compari-

son with human we achieved approximately1.3% bet-

ter average and 9.6% maximum classiﬁcation accu-

racy. Our next step will be implementation, testing

and comparison of more classiﬁers (e.g. SVM, LDA,

deep learning) and feature extraction methods (more

mother wavelets, HHT, matching pursuit). We want

to determine which combination of a feature extrac-

tion method and classiﬁer is the best solution for BCI

systems based on P300 component detection having

school-aged children as measured subjects. We will

also test artifact rejection methods which can further

improve classiﬁcation accuracy.

ACKNOWLEDGEMENTS

The work was supported by the UWB grant SGS-

2013-039 Methods and Applications of Bio- and

Medical Informatics and by the European Regional

Development Fund (ERDF), Project ”NTIS - New

Technologies for Information Society”, European

Centre of Excellence, CZ.1.05/1.1.00/02.0090.

Guess the Number - Applying a Simple Brain-Computer Interface to School-age Children

269

Table 1: Classiﬁcation accuracy obtained for different parameter settings.

Feature Extraction method settings Classiﬁcation

Name Skip samples

Band-pass

ﬁlter [Hz]

Average

classiﬁcation

accuracy ±

standard

deviation [%]

Maximum

classiﬁcation

accuracy ±

standard

deviation [%]

Human

- - - 67.61 ± 3.26

Subsampling (down sampling

factor 32)

0 0.4 - 16 56.69 ± 3.45 69.41 ± 3.21

Subsampling (down sampling

factor 32)

0 - 55.14 ± 3.46 66.5 ± 3.29

Subsampling (down sampling

factor 32)

100 0.4 - 16 63.4 ± 3.36 72.33 ± 3.12

Subsampling (down sampling

factor 32)

100 - 58.82 ± 3.43 74.27 ± 3.05

Subsampling (down sampling

factor 32)

150 0.4 - 16 67.66 ± 3.26 74.75 ± 3.03

Subsampling (down sampling

factor 32)

150 - 67.74 ± 3.26 75.72 ± 2.99

Subsampling (down sampling

factor 32)

175 0.4 - 16 64.59 ± 3.33 73.78 ± 3.06

Subsampling (down sampling

factor 32)

175 - 68.62 ± 3.23 73.78 ±3.06

Subsampling (down sampling

factor 32)

200 0.4 - 16 68.24 ± 3.24 75.72 ± 2.99

Subsampling (down sampling

factor 32)

200 - 68.89 ± 3.23 77.18 ± 2.92

DWT (Daubechies 8)

0 - 56.87 ± 3.45 70.39 ± 3.18

DWT (Daubechies 8) 100 - 63.46 ± 3.36 74.75 ± 3.03

DWT (Daubechies 8)

175 - 68.94 ± 3.22 76.7 ± 2.95

DWT (Daubechies 8) 200 - 66.82 ± 3.28 74.76 ± 3.03

REFERENCES

Blankertz, B., Muller, K., Curio, G., Vaughan, T., Schalk,

G., Wolpaw, J., Schlogl, A., Neuper, C., Pfurtscheller,

G., Hinterberger, T., Schroder, M., and Birbaumer, N.

(2004). The bci competition 2003: progress and per-

spectives in detection and discrimination of eeg single

trials. Biomedical Engineering, IEEE Transactions

on, 51(6):1044–1051.

BrainProducts (2012). Brain vision recorder.

www.brainproducts.com/productdetails.php?id=21.

Cashero, Z. (2012). Comparison of Eeg Preprocessing

Methods to Improve the Performance of the P300

Speller. Proquest, Umi Dissertation Publishing.

Letelier, J. C. and Weber, P. P. (2000). Spike sorting based

on discrete wavelet transform coefﬁcients. Journal of

Neuroscience Methods, 101(2):93 – 106.

Luck, S. (2005). An introduction to the event-related poten-

tial technique. Cognitive neuroscience. MIT Press.

Markazi, S., Qazi, S., Stergioulas, L., Ramchurn, A., and

Bunce, D. (2006). Wavelet ﬁltering of the p300 com-

ponent in event-related potentials. In Engineering in

Medicine and Biology Society, 2006. EMBS ’06. 28th

Annual International Conference of the IEEE, pages

1719–1722.

McFarland, D. J. and Wolpaw, J. R. (2011). Brain-computer

interfaces for communication and control. Commun.

ACM, 54(5):60–66.

Moucek, R. and Jezek, P. (2009). Eeg/erp portal.

http://eegdatabase.kiv.zcu.cz/home.html.

Quiroga, R. and Garcia, H. (2003). Single-trial event-

related potentials with wavelet denoising. Clinical

Neurophysiology, 114(2):376–390.

Raﬁee, J., Raﬁee, M., Prause, N., and Schoen, M. (2011).

Wavelet basis functions in biomedical signal process-

ing. Expert Systems with Applications, 38(5):6190 –

6201.

Vareka, L., Bruha, P., and Moucek, R. (2014). Event-related

potential datasets based on a three-stimulus paradigm.

GigaScience, 3(1):1–5.

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

270