Self-organizing Maps for Event-Related Potential Data Analysis

Luk

s Va

reka and Pavel Mautner

Department of Computer Science and Engineering, University of West Bohemia, 22 Univerzitn

ı, Pilsen, Czech Republic

Keywords:

Event-Related Potentials, P300, Self-Organizing Maps, BCI.

Abstract:

Event-Related Potentials (ERPs) and especially the P300 component have been gaining attention in brain-

computer interface design and neurobiological research. The detection of the P300 component in electroen-

cephalographic signal is challenging since its signal-to-noise ratio is very low. Instead of using traditional

supervised pattern recognition, this paper discusses using unsupervised neural networks for the P300 classi-

ﬁcation purposes. To validate the proposed approach, a method for the P300 detection based on matching

pursuit and self-organizing maps is proposed and evaluated. The results may be applied to the design of

brain-computer interfaces.

1 INTRODUCTION

Exogenous event-related potentials (ERPs) are elec-

trical changes recorded from the brain as it makes the

decision or initiates the response. (Picton et al., 1995)

In experiments, event-related potentials are trig-

gered by stimulation while recording electroen-

cephalographic (EEG) signal. There are several ERP

components, differing by latency (time delay after the

stimulus onset) and polarity. Both parameters are typ-

ically reﬂected by naming conventions (e.g. the N100

is an ERP with a negative amplitude located approxi-

mately 100 ms after the stimulus onset). (Luck, 2005)

Event-related potentials with higher latencies tend

to occur in association with speciﬁc experiment

paradigms. For example, oddball paradigm (Luck,

2005) is commonly used for the P300 elicitation.

In this technique, low-probability target stimuli are

mixed with high-probability non-target stimuli. Both

stimuli trigger a reaction which can be measured and

detected shortly after the event in the EEG signal

and consists of multiple ERP components. However,

the target stimuli tend to cause a different reaction,

with the P300 waveform (sometimes referred to as

the P3 component) being most signiﬁcant. This wave-

form (and especially its sub-component P3b (Polich,

2007)) is probably related to the process of decision

making - it is elicited when the subjects classiﬁes

the last stimulus as the target (for example by silent

counting). The P300 is usually the strongest ERP

component and it occurs approximately 250 - 450 ms

after the target stimulus as a positive peak. This ERP

component is frequently used in brain-computer in-

terfaces. (Luck, 2005)

Many different pattern recognition algorithms

have been proposed for brain-computer interfaces

(BCIs). BCIs allow the paralyzed users to communi-

cate with the outside world without using their mus-

cles, i.e. their intention is read from their EEG sig-

nal directly. For example, the P300 speller stimulates

the user with a matrix containing letters. Rows and

columns randomly ﬂash. When the user acknowl-

edges the letter he/she is indented to write, the row

ﬂash is followed by the P300 component that can be

detected in the signal.

Unfortunately, ERP components including the

P300 component are usually hidden in EEG noise,

i.e. signal-to-noise ratio is very low. This problem

is commonly addressed by averaging together subse-

quent trials. Random EEG activity is suppressed and

non-random ERP components time-locked to stimuli

stand out (Luck, 2005). Fig. 1 shows an example of

averaged event-related potentials for target and non-

target stimuli.

This paper discusses the problems of state-of-the-

art P300 BCI systems and proposes a new approach

towards the P300 detection based on self-organizing

maps. The beneﬁts of this approach are discussed.

Furthermore, the validity of proposals was veriﬁed

on an experimental off-line BCI system. Match-

ing pursuit was used for feature extraction and self-

organizing maps were used for classiﬁcation.

State of the art of the ﬁeld is presented in Sec-

tion 1.1. The new approach is proposed in Section

387

reka L. and Mautner P..

Self-Organizing Maps for Event-Related Potential Data Analysis.

DOI: 10.5220/0004885103870392

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2014), pages 387-392

ISBN: 978-989-758-010-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Comparison of averaged EEG responses to non-

target stimuli (Xs) and target stimuli (Os). The ensemble

averaging is necessary, otherwise the event-related poten-

tials could hardly be distinguished from noisy EEG signal.

There is a clear P3b component following the Os stimuli.

Negative is plotted upward (Luck, 2005).

1.2. Section 2 presents an experimental off-line BCI

system to evaluate the previously formulated hypoth-

esis. To describe the method, theoretical background

is introduced in Sections 2.1.1 and 2.1.2. Section 2.2

describes the process of data acquisition. The method

for pattern recognition is described in Section 2.3. In

Section 2.4, the proposed method is evaluated. The

paper is concluded in Section 3.

1.1 State of the Art

The P300 speller has been studied extensively and is

one of the well established BCI systems. However,

a recent review of the ﬁeld (Mak et al., 2011) con-

cludes that more work still needs to be done to opti-

mize the speed and accuracy before the P300 speller is

practical to use with disabled patients. This becomes

even more relevant when considering that paralyzed

patients can display widely varying P300 responses

between subjects. A reliable BCI system must be able

to adapt to the unique ERP responses of each sub-

ject and to handle the variations between trials within

a subject. When using traditional supervised pattern

recognition techniques, it is common to train the BCI

system for each new subject, allowing it to only learn

the characteristics of his/her ERP responses. There-

fore, some approaches might have difﬁculty if they

use a priori information to make assumptions about

the temporal and spatial characteristics of the standard

P300 response, especially when applied to abnormal

ERPs from paralyzed patients. (Cashero, 2012)

Therefore, the universal BCI system should not

only rely on a priori information about expected

event-related response, but should also be able to

adapt and to provide reasonable accuracy for differ-

ent subjects.

1.2 Using Unsupervised Neural

Networks

In the paper, using unsupervised neural networks

(UNNs), e.g. self-organizing maps, will be explored.

When traditional supervised learning methods are

used, all attention is concentrated on separating the

classes using class labels, and any other information

is ignored by the classiﬁer. Instead of using class

labels from a supervisor, unsupervised neural net-

works learn representation of different kinds of data

types that occur in the data sets. Since no assump-

tions of the class structure of the data are made, the

networks may discover new clusters that have not

been apparent before. Therefore, the method may

also contribute to understand the related feature vec-

tors. Self-organizing maps were successfully applied

to recognition of topographic patterns of EEG spectra

in (Joutsiniemi et al., 1995). Six classes in total were

used, for continuous alpha activity, ﬂat EEG, theta ac-

tivity, eye movements, muscle activity and bad elec-

trodes contact. The authors concluded that SOMs

were able to recognize similar topographic patterns in

different EEGs, also in EEGs not used for the training

of the map. According to (Lotte et al., 2007), Learn-

ing Vector Quantization is the closest approach that

has been investigated regarding P300 BCIs. In (Liang

and Bougrain, 2008), supervised LVQ1 has success-

fully been applied to the P300 data. This further sup-

ports the hypothesis that similar models may be ben-

eﬁcial for P300 BCIs.

Unsupervised ANN, e.g. self-organizing maps

can be trained on the data from a simple odd-ball

experiment. At least two clusters and possibly also

a ”noise” cluster should appear after training. One

cluster is expected to be associated with target fea-

tures, another one with nontarget features and in ad-

dition, the rest will probably be undecidable. An

expert can associate the clusters with classiﬁcation

classes, or training features with known classes can

be propagated through the network to create the as-

sociations. For each subject, the clusters will be dis-

tributed differently over the map. The percentage of

training features that will be associated with the un-

decidable cluster may indicate to which extent the

subject is suitable for P300 BCIs. The trained neu-

ral network could be applied to a more complex BCI

HEALTHINF2014-InternationalConferenceonHealthInformatics

388

paradigm, e.g. the P300 speller. If the classiﬁcation

class of a single trial pattern cannot be decided, the

trial can be averaged with the next corresponding trial

to gradually increase signal to noise ratio.

2 EXPERIMENTAL OFF-LINE

BCI SYSTEM

To evaluate the beneﬁts of using UNNs for the P300

detection, an experimental off-line BCI system was

designed and evaluated on P300 datasets. Pattern

recognition was based on matching pursuit and self-

organizing maps.

2.1 Theoretical Background

2.1.1 Signal Decomposition using Matching

Pursuit

Matching pursuit is an algorithm used for continu-

ous EEG signal processing (Durka and Blinowska,

1995). However, its use for event-related potential

data processing has been discussed only in a few pub-

lications (Tomas Rondik and Mautner, 2011).

It decomposes any signal into a linear expansion

of waveforms. These waveforms are selected from a

redundant dictionary of functions. At each iteration,

a waveform is chosen in order to best match the sig-

niﬁcant structures of the signal. Typically, this part is

approximated by a Gabor atom, which has the highest

scalar product with the original signal, and then it is

subtracted from the signal (Mallat and Zhang, 1993).

This process is repeated until the whole signal is ap-

proximated by Gabor atoms with an acceptable error.

Suppose we have a function g as follows:

g(t) = e

−πt

(1)

The Gabor atom has the following deﬁnition:

s,u,v,w

(t) = g(

t − u

)cos(vt + w) (2)

where s means scale, u latency, v frequency and w

phase. These four parameters deﬁne each individual

atom.

After a given amount of iterations, the signal is

decomposed into a set of Gabor atoms.

Although matching pursuit is traditionally based

on Gabor atoms, it is also possible to apply the same

principal to any other normalized base. For example,

in this paper, we focus on Gaussian base, i.e. Gabor

base with both v and w parameters set to 0. This base

also resembles the shape of event-related potentials

(Luck, 2005). Furthermore, it is less complex and

thus matching with Gaussian base has lower compu-

tational complexity.

2.1.2 Self-organizing Maps

Self-Organizing Maps (SOMs) are neural networks in

the unsupervised-learning category.

SOM converts complex, nonlinear statistical re-

lationship between high-dimensional data items into

simple geometric relationship on a low-dimensional

display. To allow this, a topological structure among

the cluster units is assumed. There are m cluster units,

arranged typically in a two-dimensional array and the

input signals are n-tuples. (Kohonen, 1989)

The weight vector of a cluster unit, commonly

referred to as a codebook vector, serves as an input

pattern associated with that cluster. During the self-

organization process, the cluster unit whose weight

vector matches the input pattern most closely (typi-

cally, by means of the square of the minimum Eu-

clidean distance) is chosen as the winner. The win-

ning unit and typically also its neighboring units (in

terms of topology of the cluster units) update their

weights. The weight vectors of the neighboring units

do not have to be close to the input pattern. The ar-

chitecture and algorithm that follow for the net can be

used to cluster a set of p continuous-valued vectors

x = (x

, x

, .., x

) into m clusters. (Fausett, 1994)

2.2 Data Acquisition

2.2.1 Stimulation Device

The experiments that were used to produce the data

for this paper were based on three-stimulus-paradigm.

The subjects were visually stimulated with three high

power Light-Emitting Diodes (LEDs) differing by

their color: red, green and yellow. Fig. 2 shows the

stimulation LED module. The core of the stimula-

tor is an 8bit micro-controller that generates required

stimuli. It also generates additional synchronization

signals for an EEG recorder. The stimulator is de-

scribed in more detail in (Dudacek et al., Sept).

Figure 2: Stimulation device with ﬂashing diodes.

Self-OrganizingMapsforEvent-RelatedPotentialDataAnalysis

389

2.2.2 Experimental Setup

Six healhy individuals (4 males and 2 females, univer-

sity students, aged 22-29) participated in our experi-

ment. The following setting of the stimulation device

was used: each diode ﬂashed once a second and each

ﬂash took 500 ms. The probabilities of the red, green

and yellow diodes ﬂashing were 75%, 20% and 5%,

respectively. The subjects were sitting 1 m from the

stimulation device for 20 minutes. They were asked

to sit comfortably, not to move and to limit their eye

blinking. They were instructed to pay attention to the

scenario and not to perform another task-relevant cog-

nitive or behavioral activity.

2.3 Pattern Recognition

Pattern recognition algorithm was designed as a pro-

cess consisting of feature extraction and subsequent

clustering of the features. The features were extracted

to correspond to the ERP components. Clustering

and subsequent analysis were used to separate various

P300 component candidates on the map and to tag the

corresponding SOM codebook vectors. This allowed

the SOM to act as a P300 classiﬁer on a testing data-

set.

2.3.1 Preprocessing and Feature Extraction

The input signal was split into epochs using stimuli

markers. For the P300 classiﬁcation purposes, only

the data associated with target and non-target stimuli

(green and red diodes ﬂashing) were extracted. Given

the sampling frequency of 1 kHz and the fact that each

epoch started 100 ms before the stimulus onset and

ended 1 s afterwards, 1100 samples were needed for

the epoch description. The samples corresponding to

the ﬁrst 100 ms were used only for baseline correc-

tion (adjusting these values to average zero). The re-

maining samples were crucial for the procedure itself.

Each sample is a real number, corresponding to its

recorded voltage value in µV. From each stimulus, the

epochs corresponding to the channel Cz were saved.

From each measured subject, 10% of the target and

nontarget epochs were randomly chosen and merged

into a training data-set. The remaining epochs were

used for evaluation. The preprocessing of the epochs

continued in the following steps:

1. The extracted epochs were randomly shufﬂed.

2. The epochs damaged by eye-blinking artifacts

were automatically removed according to the am-

plitude criterium (Luck, 2005) with the decision

threshold being set to 45 µV .

Figure 3: The ﬁgure depicts how each single feature vector

is extracted from the preprocessed epoch (blue). Matching

pursuit selects the best matching Gaussian atom. The atom

that was selected in the ﬁrst iteration is depicted in red. To

get the feature vector (plotted in black), element-by-element

multiplication of the epoch with the selected Gaussian atom

was calculated.

3. Each epoch was band-pass-ﬁltered with the cut-

off frequencies of 0.2 and 10 Hz to improve

signal-to-noise ratio.

4. The signal was downsampled by factor of 4.

As the next step, matching pursuit with Gaussian

base was applied to the pre-processed signal. Five it-

erations were calculated. However, to extract possible

components, matching pursuit atoms were not con-

sidered ERP components directly. Instead, the ERP

component candidates c were estimating using Equa-

tion 3.

c(t) = abs(a(t))s(t) (3)

In Equation 3, a is the current atom decomposed

by matching pursuit, s is the preprocessed epoch, and

t corresponds to time samples. An example is shown

in Fig. 3.

Finally, the scales of the features were normalized.

2.3.2 Clustering

For the implementation of self-organizing maps,

SOM Toolbox (Vesanto et al., 2000) was used. Self-

organizing map was trained with the training feature

vectors obtained using the procedure described above.

The size of the map was automatically adjusted to

18 x 20. Both rough training phase and ﬁne-tuning

of the network were calculated to optimize cluster-

ing. Fig. 4 shows the codebook vectors of the SOM

network. Note that there appears to be a cluster of

vectors corresponding to the P300 component.

To separate the P300 component candidates,

cross-correlation between the codebook vectors of the

HEALTHINF2014-InternationalConferenceonHealthInformatics

390

Figure 4: All weight vectors from SOM cluster units are

plotted in the upper plot. Note that many of them have

latency and polarity that correspond to the P300 compo-

nent. Unlike the other codebook vectors, these vectors cor-

respond to non-random activity time-locked to stimuli. The

bottom plot only depicts the codebook vectors that were

tagged as the P300 candidates.

Table 1: Accuracy of classiﬁcation using the trained SOM

network.

Subject ID Accuracy (Standard deviation) (%)

86 71.4 (3.6)

93 71.0 (3.6)

94 75.1 (3.4)

98 70.6 (3.6)

99 71.6 (3.6)

100 73.2 (3.5)

SOM and the P300 component (approximated by the

appropriately scaled Gaussian function) was calcu-

lated, but only in the corresponding time intervals

where the P300 component can be located (i.e. 250

- 450 ms after the stimulus (Luck, 2005)). Maximum

correlation coefﬁcient was compared with a thresh-

old. The threshold represents the decision border be-

tween two classiﬁcation classes (the P300 component

detected/not detected). The threshold was empirically

set to 0.03. Any cluster unit that was associated with

a higher maximum correlation coefﬁcient, was tagged

as the P300 candidate cluster unit. The histogram in

Fig. 6 illustrates the threshold decision problem.

2.4 Evaluation of the Off-line BCI

System

To evaluate the proposed classiﬁcation algorithm, the

following procedure was designed.

Feature vectors from the testing data-set were ex-

tracted using the procedure described above (exclud-

ing matching pursuit decomposition). Then, the fea-

ture vectors were applied to the trained SOM. For

each single feature vector, it was veriﬁed, if the SOM

Figure 5: This ﬁgure illustrates how the trained SOM re-

sponds to the testing data-set. To see the difference, only

target epochs (i.e. containing the P300), or only non-target

epochs (i.e. containing smaller P300 components) were ap-

plied to the SOM network. On the left, it is clearly ob-

servable, that more than half of the features are correctly

detected as targets. However, some feature vectors were

misclassiﬁed. In the non-target area, the features are more

equally spread. In absolute values, slightly more feature

vectors were classiﬁed as containing the P300 than in the

target group. However, since there are much more non-

targets than targets, in relative values, the percentage of mis-

classiﬁcation for non-target is lower. If the output matches

the expected output, it is depicted in green, otherwise it is

depicted in blue.

correctly decided whether the feature vector belongs

to the P300 cluster, or not. The decision was based

on comparing the tag of the feature with the tag of the

winning cluster unit. Tab. 1 contains the results. For

all subjects, the accuracy is over 70% for single tri-

als. Fig. 5 illustrates how the SOM responded to the

testing data.

3 CONCLUSIONS AND FUTURE

WORK

The suggested method has proven to be suitable for

the P300 component detection in single EEG chan-

nels. In the future, improving the method using spa-

tial ﬁltering may be further explored and tested. In

addition, a seed corresponding to the expected target

response may be inserted into the network with the

expectation that the target patterns will be associated

with the neurons positioned nearby. This would allow

the researcher to control the process of clustering.

ACKNOWLEDGEMENTS

This work was supported by the UWB grant SGS-

2013-039 Methods and Applications of Bio- and

Medical Informatics.

Self-OrganizingMapsforEvent-RelatedPotentialDataAnalysis

391

(a) Cross-correlation histogram (b) The P300 cluster in the SOM

Figure 6: The histogram illustrates how the normalized Gaussian function scaled to resemble the P300 component correlates

with the SOM codebook vectors. The threshold was set to 0.03 and any cluster units that were associated with higher

correlations were tagged as the P300 component candidates (in red). Fig. 6(b) shows the distribution of the P300 candidates

on the map. The cluster units in red belong to the P300 component cluster.

REFERENCES

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

Methods to Improve the Performance of the P300

Speller. Proquest, Umi Dissertation Publishing.

Dudacek, K., Mautner, P., Moucek, R., and Novotny, J.

(Sept.). Odd-ball protocol stimulator for neuroinfor-

matics research. In Applied Electronics (AE), 2011

International Conference on, pages 1–4.

Durka, P. and Blinowska, K. (1995). Analysis of EEG

transients by means of matching pursuit. Annals of

Biomedical Engineering, 23(5):608–611.

Fausett, L., editor (1994). Fundamentals of neural net-

works: architectures, algorithms, and applications.

Prentice-Hall, Inc., Upper Saddle River, NJ, USA.

Joutsiniemi, S. L., Kaski, S., and Larsen, A. T. (1995). Self-

organizing map in recognition of topographic patterns

of EEG spectra. IEEE Transactions on Biomedical

Engineering, 42:1062–1068.

Kohonen, T. (1989). Self-organization and associative

memory: 3rd edition. Springer-Verlag New York, Inc.,

New York, NY, USA.

Liang, N. and Bougrain, L. (2008). Non-identity Learning

Vector Quantization applied to evoked potential de-

tection. In Deuxi

eme conf

erence franc¸aise de Neuro-

sciences Computationnelles, ”Neurocomp08”, Mar-

seille, France. ISBN : 978-2-9532965-0-1.

Lotte, F., Congedo, M., L

ecuyer, A., Lamarche, F., and Ar-

naldi, B. (2007). A review of classiﬁcation algorithms

for EEG-based brain-computer interfaces. Journal of

neural engineering, 4(2).

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

tial technique. Cognitive neuroscience. MIT Press.

Mak, J. N., Arbel, Y., Minett, J. W., McCane, L. M., Yuk-

sel, B., Ryan, D., Thompson, D., Bianchi, L., and

Erdogmus, D. (2011). Optimizing the P300-based

brain-computer interface: current status, limitations

and future directions. Journal of Neural Engineering,

8(2):025003+.

Mallat, S. and Zhang, Z. (1993). Matching pursuit with

time-frequency dictionaries. IEEE Transactions on

Signal Processing, 41:3397–3415.

Picton, T. W., Lins, O. G., and Scherg, M. (1995). The

recording and analysis of event-related potentials. In

Boller, F. and Grafman, J., editors, Handbook of Neu-

ropsychology, volume 10, pages 3–73. Elsevier, Ams-

terdam.

Polich, J. (2007). Updating P300: an integrative the-

ory of P3a and P3b. Clinical neurophysiology,

118(10):2128–2148.

Tomas Rondik, Jindrich Ciniburk, R. M. and Mautner,

P. (2011). Erp components detection using wavelet

transform and matching pursuit algorithm. In Applied

Electronics 2011, pages 1–4.

Vesanto, J., Himberg, J., Alhoniemi, E., and Parhankangas,

J. (2000). Self-organizing map in matlab: the som

toolbox. In In Proceedings of the Matlab DSP Con-

ference, pages 35–40.

HEALTHINF2014-InternationalConferenceonHealthInformatics

392