A Hierarchical BCI System Able to Discriminate between Non

Intentional Control State and Four Intentional Control Activities

Julio Abascal, Andoni Arruti, José I. Martín and Javier Muguerza

Department of Computer Architecture and Technology, University of the Basque Country (UPV/EHU)

Manuel de Lardizabal 1, 20018 Donostia, Spain

Keywords: Brain-Computer Interface (BCI), Non Intentional Patterns Detection, Electroencephalogram (EEG),

Clustering, Supervised Learning.

Abstract: This paper presents a two-level hierarchical approach to recognising intentional and non intentional mental

tasks on a brain-computer interface. A clustering process is performed at the first recognition level in order

to differentiate Non intentional Control state (NC) patterns from Intentional Control (IC) patterns. At the

second level, the IC detected patterns are classified by means of supervised learning techniques, applied to

the type of movement (left hand, right hand, tongue or foot imagery movement). The objective is to achieve

high correct movement recognition scores, with a low percentage of wrong decisions (that is, low false

positive rates), to avoid user frustration. Offline evaluation of the proposed prototype shows 84.5%

accuracy, with a 6.7% false positive rate.

1 INTRODUCTION

Brain-Computer Interfaces (BCI) based on

Electroencephalography (EEG) enable users to

command computers just by measuring EEG signals

associated with brain activity (Wolpaw et al., 2002).

This kind of BCI requires a system to identify user

brain activity patterns that are later translated into

commands (Lotte et al., 2007).

Most BCI systems are based on synchronous

protocols where the subject must follow a fixed

repetitive scheme to switch from one mental task to

the next (Pfurtscheller and Neuper, 2001) (Wolpaw

et al., 2002). In synchronous BCI systems, the EEG

phenomena to be recognized are time-locked to

diverse cues. A trial typically lasts from 4 to 10 s or

more. In contrast, in asynchronous BCI systems the

subject makes voluntary, self-paced decisions on

when to stop performing a mental task and when to

start the next one (Nooh, Yunus and Daud, 2011).

Designing an asynchronous BCI system requires

continuous analysis of EEG signals. This analysis

should determine whether the user is in an

Intentional Control (IC) state, that is, if (s)he is

producing one of the brain activity patterns used to

control the BCI, or if (s)he is in a Non Control (NC)

state. Finally, if the user is in an IC state, the system

also has to determine which kind of brain activity

pattern is being produced. Therefore, to deal with

asynchronous problems, it is necessary to be able to

differentiate between known and unknown activity

patterns.

This paper presents the preliminary results of a

study dealing with the problem of classifying

patterns between the different types of IC states after

an NC state discarding process. The approach has a

two level hierarchical structure. The first level

determines whether an activity pattern is present or

not by applying a clustering process. The second

level detects which of four mental tasks (left hand,

right hand, tongue and foot imaginary movements)

has been produced by the user. For this level a

supervised classifier based on Support Vector

Machine paradigm is proposed.

The remainder of the paper is organized as

follows. Section 2 explains the experimental

protocol used, the data acquisition process and the

pre-processing carried out. Section 3 presents the

proposed system to classify the EEG signals into the

four types of imaginary movements considered,

discarding the NC states. Results are presented and

discussed in Sections 4 and 5. Finally, some

conclusions and references are presented.

Abascal J., Arruti A., I. Martín J. and Muguerza J..

A Hierarchical BCI System Able to Discriminate between Non Intentional Control State and Four Intentional Control Activities.

DOI: 10.5220/0004723000910097

In Proceedings of the International Conference on Physiological Computing Systems (PhyCS-2014), pages 91-97

ISBN: 978-989-758-006-2

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

2 EXPERIMENTAL PROTOCOL

For our tests, we used the IIIa dataset from the BCI

competition III (Blankertz et al., 2006). It contains

data from 3 subjects: K3b, K6b and L1b, collected

as follows (Schlögl et al, 2005) (see Figure 1). Each

subject, sitting in front of a computer, was asked to

perform imaginary movements of the left hand, right

hand, tongue or foot during a specified time interval

according to a cue. The order of cues was random.

60 electrodes were placed on the subject's scalp

recording a signal sampled at 250 Hz and filtered

between 1 and 50 Hz using a Notch filter. Each trial

started with a blank screen. At t = 2s, a beep was

generated and a cross “+” was shown to attract the

subject's attention. At t = 3s an arrow pointing to the

left, right, up or down was shown for 1s and the

subject was asked to try one of four imaginary

movements until the cross disappeared at t = 7s. This

was followed by a 2s break, and then the next trial

began. The dataset contains 360 instances (cases) for

subject K3b, 240 for K6b and 240 for L1b. Each

instance was labelled as belonging to one of the four

classes. Each dataset contains a balanced distribution

of the classes.

Comparing the subjects shows that K3b presents

the highest accuracy and K6b the lowest (Lee et al.,

2005). This is attributed to the different amount of

BCI training received by the subjects. K3b was the

most experienced, L1b had less experience and K6b

was a beginner. In this work, we have used the data

corresponding to subjects with extreme skills: K3b

and K6b.

This dataset was designed for a synchronous BCI

system, where the subject is aware of the time period

to imagine the proposed activity. Therefore, this data

was collected while the subjects were performing a

cue-based (synchronized) task. A summary of the

results obtained in this synchronous data logging

exercise can be seen in (AlZoubi, Koprinska and

Calvo, 2008).

In this paper, we extended this experimental data

by adding data associated to NC states to the dataset.

We associated the first three seconds (t = 0-3s) of

each trial with no intentional activity. Therefore, for

subject K3b we get 720 instances: 360 instances that

indicate one of the four different intentional

activities (IC state), corresponding to processing 4-7

seconds of each trial, and another 360 instances with

non intentional brain activity (a new class: NC

state), corresponding to the data obtained from 0-3

seconds of each trial. We processed the data from

subject K6b in the same way, and we obtained 480

instances: 240 instances of intentional activity and

240 instances of non intentional activity.

(a)

Arrow +

Fixation cross

Training

Fixation cross

0 1 2 3 4 5 6 7

time (s)

Blank

screen

Beep

Fixation cross

(b)

Figure 1: (a) EEG electrode placement, and (b) data

acquisition paradigm (BCI Competition III).

We applied the same pre-processing techniques

applied by AlZoubi et al. in (AlZoubi, Koprinska

and Calvo, 2008). Firstly, we applied the Common

Spatial Patterns (CSP) method (Müller-Gerking,

Pfurtscheller and Flyvbjerg, 1999) to the raw EEG

data. The result of applying CSP to the original 60

signals is a new set of 60 signals sorted by their

ability to predict class. We selected the first 5

projections, and then we applied 3 frequency band

filters (for 8-12 Hz, 21-20 Hz and 20-30 Hz).

Finally, we extracted 7 features: max, min and mean

voltage values, voltage range, number of samples

above zero volts, zero voltage crossing rate and

average signal power. This process gives 525

[5x5x3x7] (5 classes, 5 projections, 3 filters and 7

features) discrete numeric values for each case of the

dataset.

After pre-processing the data corresponding to

each subject, we split all the data into three sets: a

training dataset for the clustering process at the first

level (to detect the IC states), another training

dataset for learning a supervised classifier at the

second level (to distinguish between the four types

of mental tasks) and a test dataset for evaluating the

system's performance. Therefore, we created three

datasets randomly for each subject, containing 240

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

cases for subject K3b (120 correspond to NC state

and 120 to IC state, namely 30 cases for each

imaginary movement), and 160 cases (80 for NC

state and 80 for IC state: 20 cases for each imaginary

movement) for subject Kb6.

One of the most important problems in machine

learning is the need to deal with high numbers of

dimensions. This problem is known as the curse of

dimensionality: small numbers of training instances

but highly dimensional. In these cases, it is

necessary to simplify the learning algorithm by

reducing the dimensionality before starting the

learning process. This can be done by selecting the

problem's most informative features and discarding

the most irrelevant and redundant features. In this

work we applied the Correlation-Based Feature

Selection (CFS) method (Hall, 2000) which is the

same feature selection method used by AlZoubi et

al. in (AlZoubi, Koprinska and Calvo, 2008). This

method bases its selection on searching for features

that are highly correlated with a specific class

variable and least correlated with the other variables.

We used the implementation provided by Weka

data-mining platform (Witten and Frank, 2005). We

used the best first (greedy) search option starting

with an empty set of features and adding new

features. It is important to note that feature selection

was only performed using training data, and test data

was not used in any way during feature selection. As

a result, 45 and 40 features were selected for K3b

and K6b, respectively.

3 PROPOSED HIERARCHICAL

SYSTEM: TRAINING PHASE

As mentioned in the Introduction, the system

proposed in this work has a hierarchical structure

that can be seen in Figure 2. The first level

determines the presence or absence of intentional

activity in the EEG signal, applying clustering

techniques. The second level determines whether the

detected intentional activity is a left hand, right

hand, tongue or foot imaginary movement.

3.1 First Level: Detecting IC Activity

We used the K-means algorithm (Weka

implementation) with the Euclidean distance for the

first level of the system. One key issue in this phase

is how good the system is at rejecting the NC state.

Hence, to design this level we transformed the five-

class dataset into a two-class dataset.

It is critical to minimize the False Positive Rate

of the NC class (FPR = FP / (FP+TN)). FPR

depends on False Positives (FP, acceptance of an NC

state as an IC state) and on True Negatives (TN,

rejection of a true NC state). A high FPR tends to

cause excessive user frustration making the resulting

BCI unusable (e.g. if the system is used to control a

wheelchair, an FP would imply undesirable chair

movements implying high risk for the user).

According to the literature, we have selected a

maximum threshold of 10% for the system's first

level FPR. This value is similar to the FPR used in

(Lotte, Mouchère and Lécuyer, 2008) (Scherer et al.,

2008). After the clustering process, each cluster is

labelled as belonging to one of the two classes, NC

or IC, taking into account that the FPR (for the NC

class) must be less than the selected threshold

(10%). This has been done by establishing a

minimum number of IC class patterns for the

generated clusters (IC-threshold) to be labelled as an

IC cluster.

Figure 2: Structure of the proposed BCI system.

Therefore, for this first stage of the system, we had

to select the IC-threshold and the K value for the K-

means algorithm. We analysed 10 different values

for the parameter K: 5, 10, 15, 20, 25, 30, 35, 40, 45

and 50. This estimation was made by applying a 10-

fold cross-validation methodology using the first set

of training data with patterns from the two-class

system: NC state and IC state (grouping all patterns

of each imaginary movement). Only clusters

exceeding the IC-threshold were labelled as IC class.

We calculated the nearest cluster for each unused

pattern of the cross-validation fold using average

linkage distance. As shown in Table 1, to maintain

the FPR level under 10%, a 70% IC-threshold was

necessary. The best results were obtained with K =

Electrode

array

clustering

NC state: no movement

supervised

classification

IC state

IC state 1: left hand movement

IC state 2: ri

ht hand movement

IC state 3: ton

ue movement

IC state 4: foot movement

AHierarchicalBCISystemAbletoDiscriminatebetweenNonIntentionalControlStateandFourIntentionalControl

Activities

Table 1: FPR (NC class) and accuracy (10-CV) depending on the value of the K parameter and the IC-threshold. Grey

shading for FPR under 10%, * shows the highest accuracy, and bold highlights the best option for each subject.

IC-

threshold

10-CV

K value

5 10 15 20 25 30 35 40 45 50

90%

K3b

FPR (%) 0.0 5.8 6.7 4.2 6.7 4.2 5 5.8 4.2 5.8

Accuracy (%) 51.7 62.1 62.9 65.8 65 70 70.8 67.1 72.1* 70.8

K6b FPR (%) 0.0 6.3 5.0 5.0 8.8 7.5 10.0 13.8 13.8 13.8

Accuracy (%) 50.0 56.9 63.8 70.0* 64.4 61.3 59.4 63.8 66.3 66.9

80%

K3b FPR (%) 10.8 12.5 10.0 8.3 10 8.3 8.3 8.3 6.7 10

Accuracy (%) 74.6 75.0 74.6 77.9 72.9 80.4 79.2 79.2 81.3* 82.9

K6b FPR (%) 7.5 11.3 6.3 6.3 10.0 12.5 12.5 18.8 16.3 16.3

Accuracy (%) 67.5 68.8 73.8* 73.8* 72.5 70.0 68.1 73.1 73.1 73.8

70%

K3b FPR (%) 10.8 12.5 10.8 11.7 15.8 8.3 10.8 13.3 12.5 13.3

Accuracy (%) 75.4 79.6 75.4 80.0 76.3 83.8 80.4 79.2 80.4 81.3

K6b FPR (%) 15.0 16.3 7.5 11.3 16.3 13.8 13.8 20.0 16.3 16.3

Accuracy (%) 78.1 76.3 77.5 76.3 74.4 76.9 70.0 73.8 731 73.8

60%

K3b FPR (%) 15.8 19.2 10.8 14.2 19.2 15.8 18.3 20.0 15.8 19.2

Accuracy (%) 75.8 79.6 76.7 82.5 79.2 83.3 80.0 78.8 81.7 82.1

K6b FPR (%) 18.8 17.5 10.0 13.8 20.0 18.8 22.5 20.0 20.0 20.0

Accuracy (%) 81.9 79.4 80.0 78.1 75.0 80.0 75.0 75.6 76.3 74.4

Table 2: Accuracy (10-CV) of the classifiers.

10-CV

Accuracy (%)

1R DT 1-NN 5-NN NB RBF SVM LR AdaB Bag RF

K3b 45.8 60.8 77.5 82.5 80.8 77.5 84.2 74.2 72.5 74.2 77.5

K6b 31.3 58.8 53.8 58.8 60.0 63.8 65.0 53.8 62.5 56.3 56.3

30 for subject Kb3, and K = 15 for subject K6b. For

these K values, the FPR (NC class) was 8.3% and

7.5%, respectively.

3.2 Second Level: Classifying the Type

of Imaginary Movement

We used supervised learning algorithms to

implement this level. From the great variety of

algorithms that have been applied in BCI systems

(Lotte et al., 2007), we selected 11 algorithms: 1R

rule, Decision Tree (DT), k-NN (1-NN and 5-NN),

Naive Bayes (NB), Radial-bases Network (RBF),

Support Vector Machine (SVM), Logistic

Regression (LR), Ada Boost (AdaB, combining 10

decision trees), Bagging (Bag, combining 10

decision trees) and Random Forest (RF). We chose

these algorithms because they represent different

paradigms (rule-based, tree-based, distance-based,

probabilistic, function-based, ensemble of

classifiers) and they are state of the art in data

mining. We use their Weka (Witten and Frank,

2005) implementation by applying the default values

for the parameters.

All classifiers were trained using the second

training set (120 cases, 30 cases of each imagery

mental task). The best algorithm was estimated by

applying a 10-fold cross-validation methodology.

Table 2 shows the accuracy achieved by each

classifier for both subjects. The best overall

classifier was the SVM algorithm with an accuracy

of 84.2% for subject K3b and 65.0% for subject

K6b.

Summarizing, the proposed hierarchical BCI

system consists of a first level that differentiates

between NC and IC states, based on clustering

techniques; and a second level, based on a SVM

classifier, that discriminates between the four types

of mental tasks considered. The optimal number of

first level clusters is different for each subject: 30 for

subject K3b and 15 for subject K6b.

4 EXPERIMENTAL RESULTS:

EXPLOITATION PHASE

Once the system has been trained, it can be used for

classifying new EEG patterns, so far unknown to the

system. The newly-designed system's performance

was tested using the test dataset. As previously

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

Table 3: Classification performance for the first level using the test dataset.

Confusion Matrix FPR (%) Accuracy (%)

K3b

(K-means, K=30)

IC-estimated NC-estimated

6.7

81.7

IC-real 84 36

NC-real 8 112

K6b

(K-means, K=15)

IC-estimated NC-estimated

2.5

80.0

IC-real 50 30

NC-real 2 78

Table 4: Classification performance for the second level with the test dataset.

Confusion Matrix Accuracy (%)

K3b

(SVM)

Left-

estimated

Right-

estimated

Tongue-

estimated

Foot-

estimated

84.5

Left-real 6 7 0 2

Right-real 2 26 0 0

Tongue-real 1 0 18 1

Foot-real 0 0 0 21

K6b

(SVM)

Left-

estimated

Right-

estimated

Tongue-

estimated

Foot-

estimated

64.0

Left-real 2 9 0 0

Right-real 1 11 1 0

Tongue-real 3 3 1 0

Foot-real 0 0 1 18

explained, the test set for subject K3b comprises 240

cases (120 corresponding to NC state and 120 cases

for IC state), while the test set for subject K6b

consists of 160 cases (80 of each state). Considering

only the IC patterns, subject K3b's test set consists

of 30 cases of each of the 4 types of mental tasks,

whereas, subject K6b's test set consists of 20 cases

of each type of movement.

Table 3 summarizes the results obtained for the

first level of the system when the new patterns were

processed. This table shows the confusion matrix

obtained, as well as the FPR (NC class) and the

accuracy of this first level. There are 8 cases of NC

patterns misclassified into clusters labelled with

some kind of movement (i.e. intentional activity)

leading to 6.7% FPR (NC class) for subject K3b. 84

cases (from the initial 120) corresponding to activity

patterns (IC state) will be classified in the second

level of the system. On the other hand, they are only

2 cases of misclassified NC patterns; yielding a

2.5% FPR for subject K6b, 50 cases (from the initial

80) correspond to activity patterns.

Table 4 shows the confusion matrix for the

second level of the system. The patterns are

classified in the 4 possible imaginary movements

(right, left, tongue, foot) using a Support Vector

Machine (SVM) classifier. The accuracies obtained

from the two subjects are 84.5% and 64.0%,

respectively.

Analysing the confusion matrix shows that it is

more difficult for both subjects to detect the left

hand movement. The number of patterns reaching

the second level of the system for this movement is

clearly lower than for all other movements, and,

there is a bias in the system that classifies these

patterns as belonging to the right imaginary

movement. In general, subject K6b obtained worse

results and presented specific difficulties with the

tongue movement (only 7 patterns reached the

second level, and, only one was correctly classified).

Analysing the overall system's performance in

terms of classifying the five different patterns (NC

class + 4 imaginary movements), the accuracy for

subject K3b is 76.3% and 68.8% for the other

subject. The differences in results obtained for each

subject confirmed the description of subject K6b as

a less trained (beginner) user, and, as a consequence,

the system had greater difficulty dealing with this

subject's EEG patterns.

Although it is difficult to compare the results

obtained with other works, mainly because we have

included a class for the Non Intentional Control

state, we can say that the results obtained are similar

to the work presented by AlZoubi et al. in (AlZoubi,

Koprinska and Calvo, 2008). They obtained a 78.5%

average result for both subjects, whereas in our case

the accuracy was 72.6% (taking into account the

difficulty of introducing the NC class).

AHierarchicalBCISystemAbletoDiscriminatebetweenNonIntentionalControlStateandFourIntentionalControl

Activities

Table 5: Accuracy (10-CV) of the classifiers (one-level system).

10-CV

Accuracy (%)

1R DT 1-NN 5-NN NB RBF SVM LR AdaB Bag RF

K3b 40.4 63.9 75.8 79.6 75.7 77.1 81.1 69.9 76.0 72.5 73.3

K6b 37.0 50.0 44.0 46.5 52.0 50.5 60.0 47.5 56.0 57.0 56.0

5 COMPARISION WITH AN

ONE-LEVEL SYSTEM

In order to test the validity of the proposed

hierarchical system, we also implemented a one

level system without the clustering phase. We

analysed the performance of the same classifiers

used in Section 2 to develop the second level

classifier for the five classes (NC, left hand, right

hand, tongue or foot). All the classifiers were now

trained using the two previously defined training

data sets: 300 cases (60 cases of each class) for

subject K3b and 200 cases (40 cases of each class)

for subject K6b. The best algorithm was estimated

again by applying a 10-fold cross-validation

methodology. Table 5 shows the accuracy achieved

by each classifier for both subjects. The best overall

classifier was again the SVM algorithm, with an

accuracy of 81.1% for subject K3b and 60% for

subject K6b.

The system performance was tested with the

same test data set used for the hierarchical proposal:

240 cases for subject K3b and 160 cases for subject

K6b. The obtained accuracy using the SVM

classifier was lower than the one obtained with the

two-level system: 74.5% for subject K3b and 66.9%

for subject K6b. These results confirm that the two-

level approach has higher performance than the one-

level system.

6 CONCLUSIONS

In this paper we proposed a two-level hierarchical

approach to recognise mental tasks including

intentional and non intentional states on a brain-

computer interface. At the first level, the proposal

performs a clustering process in order to

differentiate patterns of Non intentional Control

state (NC) from patterns of Intentional Control (IC).

At the second level, the IC detected patterns are

classified by movement type (left hand, right hand,

tongue or foot imaginary movement) by a supervised

learning classifier.

After a pre-processing phase and reducing the

number of dimensions of the problem, we applied

the K-means algorithm for the first level of the

system, obtaining the best results with K = 30

(accuracy of 81.7%) and K = 15 (accuracy of 80.0%)

for subjects K3b and K6b, respectively, using BCI

III Competition dataset IIIa. The best results

obtained for the second level were achieved with the

Support Vector Machine classifier with 84.5% and

64.0% overall accuracy, respectively. These results

were obtained maintaining the False Positive Rate

for the NC class under 10% (achieving 6.7% and

2.5% rates for the subjects participating in the

experiment). The classification phase results

encourage us to apply Support Vector Machine

based algorithms in the clustering phase.

This work takes advantage of the good results

obtained in synchronous experiments to apply them

in a more realistic but more demanding

asynchronous environment. In the asynchronous

case, the data includes inactivity periods along with

activity states. Our work firstly distinguishes activity

from non-activity status. Once voluntary activity is

found, we detect the type of virtual movement

associated with it. In this way, our research proposes

a step forward towards practical asynchronous

detection.

Nevertheless, this work did not deal with the

problem of detecting IC states in a continuous EEG

signal (asynchronous or self-paced BCI). We just

introduced the NC state as a new class. This may

explain why our results are generally better than

presented in other works (Scherer et al., 2008) (Satti,

Coyle and Prasad, 2009). We plan to apply our

proposal to the asynchronous BCI problem in order

to be able to compare both approaches.

ACKNOWLEDGEMENTS

This work was funded by the University of the

Basque Country (Aldapa, GIU10/02), by the Science

and Education Department of the Spanish

Government (ModelAccess project, TIN2010-

15549), and by the Department of Education,

Universities and Research of the Basque

Government (IT395-10 research group grant).

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

REFERENCES

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

G., Vaughan T., 2002. Brain-computer interfaces for

communications and control. Clinical

Neurophysiology, 113 (6): 767-791.

Lotte, F., Confgedo, M., Lécuyer, A., Lamarche, F.,

Arnaldi, B., 2007. A review of classification

algorithms for eeg-based brain-computer interfaces.

Journal of Neural Engineering, 4:R1-R13.

Pfurtscheller, G., Neuper, C., 2001. Motor imagery and

direct brain-computer communication. Proc. IEEE,

vol. 89, pp.1123-1134.

Nooh, A. A., Yunus, J., Daud, S.M., 2011. A review of

Asynchronous Electroencephalogram-based Brain

Computer Interface Systems. International Conference

on Biomedical Engineering and Technology,

Singapore, vol. 11.

Blankertz, B., Müller, K. R., Krusienski, D., Schalk, G.,

Wolpaw, J. R., Schlögl, A., Pfurtscheller, G., Millan,

J., Schröder, M., Birbaumer, N., 2006. The BCI2000

Competition III: Validating Alternative Approaches to

Actual BCI Problems. IEEE Transactions on Neural

Systems and Rehabilitation Engineering, vol. 14, no.

2, pp. 153-159.

Schlögl, A., Lee, F., Bischof, H., Pfurtscheller, G., 2005.

Characterization of four-class motor imagery EEG

data ffor the BCI competition 2005. Journal of Neural

Engineering, 2: L14-L22.

Lee, F., Sherer, R., Leeb, R., Neuper, C., Bischof, H.,

Pfurtscheller, G., 2005. A Comparative Analysis of

Multi-class EEG Classification for Brain-computer

interface. 10th Computer Vision Winter Workshop

(CVWW). Technical University of Graz, Austria.

AlZoubi, O., Koprinska, I., Calvo, R.A., 2008.

Classification of Brain-computer Interface Data. 7th

Australasian Data Mining Conference (AusDM),

Adelaide (Australia), pp. 123-132.

Müller-Gerking, J., Pfurtscheller, G., Flyvbjerg H., 1999.

Designing optimal spatial filters for single-trial EEG

classification in a movement task. Clinical

Neurophysiology, vol. 110, no. 5, pp. 787-798.

Hall, M., 2000. Correlation-based feature selection for

discrete and numeric class machine learning. 17th

International Conference on Machine Learning

(ICML), 359-366, Morgan Kaufmann.

Witten, L. H., Frank, E., 2005. Data Mining: Practical

Machine Learning Tools and Techniques. Morgan

Kaufmann, San Francisco.

Lotte, F., Mouchère, H., Lécuyer, A., 2008. Pattern

Rejection Strategies for the Design of Self-Paced

EEG-based Brain-Computer Interfaces. 19th

International Conference on Pattern Recognition,

Florida (USA), pp. 1-5.

Scherer, R., Lee, F., Schlögl, A., Leeb, R., Bischof, H.,

Pfurtscheller, G., 2008. Towards self-paced brain-

computer communication: Navigation through virtual

worlds. IEEE Transactions on Biomedical

Engineering, vol. 55, no. 2, pp. 675-682.

Satti, A., Coyle, D., Prasad, G., 2009. Continuous EEG

Classification for a Self-paced BCI. 4th International

IEEE EMBS Conference on Neural Engineering,

Turkey, pp. 315-318.

AHierarchicalBCISystemAbletoDiscriminatebetweenNonIntentionalControlStateandFourIntentionalControl

Activities