Classification of Sustained Attention Level Based on Morphological Features of EEG s Independent Components

Abstract— The aim of this study is to investigate the

relations between morphological features of ERP’s independent components and visual sustained attention. Continuous Performance Test (CPT) is used for defining the level of sustained attention. Independent Component Analysis (ICA) is applied on 19-channel recorded EEG and the best component is determined based on time, frequency and spatial specifications of the components. The ERPs are extracted for each group of stimuli and eighteen morphological features (including P3) were extracted. Nineteen subjects were divided into three groups according to their attention level. LDA classifier is then used for discrimination of classes. The results are compared with two other common methods. Classification based on the proposed method yields in accuracy of 81% with the advantage of preserving almost all the data. Outcomes represent a significant correlation between CPT result and some parameters of brain signal’s components which can be used in evaluating the level of attention.

I. INTRODUCTION

VALUATION of attention has various applications such as diagnosis and treatment of attention diseases, learning enhancement in classroom, attention investigation in infants and attention monitoring in critical activities such as aeronautical navigation.

Direct measurement of Attention using brain signals, is important because it can eliminate excessive interfaces. Benefits of achieving such a system are: minimizing individual effects such as responding speed or processing ability, application for subjects with vision/auditory disordersandindependenceofaspecificlanguageor culture.

Since Electroencephalography/Event Related Potential (EEG/ERP) is the most direct noninvasive method for investigating brain activities and consequently attention, thus its evaluation is interesting for neuroscientists in attention studies.

Up to now, different studies have been carried out regarding various types of attention and their relations with

Manuscript received February 14, 2009.

F. Ghassemi is with the Biomedical Engineering Faculty of Amirkabir University of Technology, Tehran, Iran (corresponding author, phone: (+98)912-326-0661; fax: (+98)21-6642-0672; e-mail: [email protected]). M. H. Moradi is with the Biomedical Engineering Faculty of Amirkabir University of Technology, Tehran, Iran (e-mail: [email protected]).

M. Tehrani Doust is with the Psychology Department, University of Tehran, Tehran, Iran, and Institute for Cognitive Science Studies (ICSS), Tehran, Iran (e-mail: [email protected]).

V. Abootalebi is with the Electrical Engineering Faculty of Yazd University, Yazd, Iran (e-mail: [email protected]).

brain activities. One famous kind of attention is "Sustained Attention" which has a key role in attention diseases such as Attention Deficit Hyperactivity Disorder (ADHD) and Attention Deficit Disorder (ADD). Sustained attention is defined as the ability to maintain a consistent behavioral response during continuous and repetitive processing of stimuli whose non-arousing qualities would otherwise lead to habituation and distraction to other stimuli [1], [2].

Unfortunately, there are different cognitive states associated with sustained attention which are in broad use with varied ambiguous meanings such as arousal, alertness and vigilance (Fig. 1). This problem is investigated in [3] and it is mentioned that sustained attention can be used synonymously to vigilance and tonic alertness (not phasic alertness).

One of the most popular tests for evaluating sustained attention is Continuous Performance Test (CPT) [4]. Obtained results on EEG/ERP study during CPT [5] showed that all subjects shared a progressive backshift of alpha rhythm during the performing of the CPT test while beta and gamma activities were stronger in the right hemisphere than in the left. An intense and widespread decrease in EEG spectral power during test performing became visible in many subjects.

Another recent study on a similar test for sustained attention revealed two main ERP components (N2 and P3) investigating whether a central inhibitory mechanism intervenes to prevent the preparation and/or execution of a motor response. N2 was detectable in the No-Go trials. P3 presented a different scalp distribution based on the type of trial [6].

P3 is one of the most obvious cognitive components in ERP. During the process of a series of usual stimuli, when brain meets an unusual stimulus (target stimulus), a P3

.

Classification of Sustained Attention Level Based on

Morphological Features of EEG’s Independent

Components

Farnaz Ghassemi, Student Member, IEEE, Mohammad Hasan Moradi, Member, IEEE,

Mahdi Tehrani Doust, and Vahid Abootalebi, Member, IEEE.

E

wave can be observed in recorded brain signals. Amplitude of P3 is usually about 10-15 µv and its latency for visual stimulus is reported 350-650 ms which may be increased till 1000 ms [7].

Independent Component Analysis (ICA) method has been used in several studies [8]-[11] to separate sources with brain origin from other sources with artifactual origins. ICA is referred to the separation of independent sources which are mixed together with an unknown matrix, i.e. mixing system and source signals are both unknown [12]. Makeig and coworkers applied ICA to investigate early and late ERPs during visual spatial attention and led to new grounds for attention extraction from brain signals [13].

This paper examines the correlation between sustained attention level and morphological features of independent components of brain signals. Significant related features are then used to classify sustained attention levels.

II. METHODS AND MATERIALS A. CPT Task

This study is based on the Conners’ CPT II test [4] which is a "No-Go" CPT task. The subjects seated on a comfortable chair with a place for relaxing the head. The test was performed in a quiet and dimly lit room. The distance between the subjects’ eyes and monitor (19 inch) was 75±5 cm depending on their height. Different letters of English alphabet were presented randomly on the screen and subjects were asked to click the left mouse button with the index finger of their dominant hand when any letter except the target "X" appeared. Subjects were instructed to respond as fast as they can, but also as accurate as possible.

The letters were 7.5 cm high and 7 cm wide which results in a 7° visual angle. They appeared white colored on a black background. The inter-stimulus intervals (ISIs) were 1, 2 or 4 seconds with a display time of 250ms. There were 6 blocks, with 3 sub-blocks each containing 20 trials. The order in which the different ISIs were presented varied between blocks. The experiment involved 360 stimuli: 36 X letters (No-Go stimuli) and 324 other letters (Go stimuli). Each test took approximately 14 minutes to complete.

B. EEG Recordings

19-channel EEG was recorded with Ag/AgCl electrodes mounted in an electrode cap placed according to the international 10-20 standard (Fp1/Fp2, F3/F4, F7/F8, T3/T4, C3/C4, T5/T6, P3/P4, O1/O2, Fz/Cz/Pz). Average of A1 and A2 was used as reference. Three additional bipolar channels were used for vertical EOG recording, synchronization of CPT system with EEG signals and recording of subject’s responses. For the last two pairs of channels, two independent isolator circuits were used for providing the subjects’ safety and reducing noise. A 32-channel AC/DC amplifier (Walter) was used for data recording and Pl-Winsor 3.0 for data acquisition. Amplifier band pass was

0.05–100 Hz and a 50 Hz notch filter was used for line noise reduction. The A/D sampling rate was 200 Hz. Impedance of all electrodes were kept below 5 KΩ.

C. Subjects

Nineteen volunteers (9 male) with average age of 28.5 ± 6.3 years participated in the experiment. Since financial incentives affect the sustained attention level [3], subjects were not paid to participate and they were just thanked with a small gift at the end of the experiment. All subjects were tested for handiness with Edinburgh test (1 left-handed). They had normal or corrected to normal vision. For evaluating generalization (as the test is in W/B), all subjects were checked for color-blindness with Ishihara test and one male was color-blind. For evaluating subject’s Neuropsychological history SARs test was used and subjects were free of neurological disorders and were not taking any medication. All subjects gave their informed consent. The experiment was conducted in accordance with the Declaration of Helsinki.

D. ERP Extraction

A suitable band pass filter (0.1 – 85 Hz) was used to eliminate line noise and movement artifacts.

Three different methods were performed. In the main method (A), ICA was applied to the data. Combination of Efficient Variant of Fast ICA (EFICA) and Efficient Weights Adjusted SOBI (EWASOBI) was used as the ICA method [14]. Time, frequency and spatial characteristics of the resulted independent components were studied and the best component with brain origin (not artifact) which was physiologically plausible was selected manually.

In method B, after applying ICA, only the independent component which was considered as EOG artifact was eliminated and then other independent components were transferred into initial electrode space, thus only the effects of EOG on the signals were eliminated.

In method C, portion of signals which contained EOG artifacts, were deleted. In the last method 52% of data was eliminated and only 48% was remained for further processing. Therefore the first two methods have the advantage of saving almost all epochs even if the subject blinked or had eye movements. This is an important advantage specially when the task is long and the probability of blinking is high so a big portion of data is polluted with EOG artifact (As in this case).

The P3 window was considered at 350–650ms after stimulus onset based on the prior researches [6], [7], [15]. Amplitude of P3 is defined as the difference between the mean pre-stimulus baseline voltage and the largest positive-going peak of the ERP waveform within the time window [7].

Four groups were considered for each subject: Target stimuli (X), Non-target stimuli (nX), Correct answered

target stimuli (CX, which is the X that subject did not respond) and Non-correct answered target stimuli (nCX).

Then in each mentioned method, periods of 200 ms before till 1000 ms after stimuli onset were considered for all groups. Extracted epochs which were time-locked to stimuli onset, were averaged to calculate the ERPs.

Based on previous studies, P3 component is more clear in central channels specially Pz [7], [15], thus in this paper only Pz channel signals were considered.

E. Features and Classifier

Amplitude and latency were both computed for four components: P3, the positive peak after P3 (named P4 in this article), positive peak before P3 (P2) and the negative peak before P3 (N2) in all four groups. In addition, these features could be considered absolutely or relative to each other, thus 64 features (2×4×4×2) were calculated.

Subjects divided to three classes according to their attention level, 7 in high (H), 5 in medium (M) and 7 in low (L) attention class.

For investigating the relation between these features and CPT result, Pearson correlation was calculated between attention level and the morphological features. Eighteen features had significant correlation (P-value<0.05) with attention level which are characterized in Table. 1. Symbol "A" is used for amplitude and "L" for latency. The phrase in the parenthesis explains the group (X, nX, CX, nCX) and subscript indices express the related component of the waveform (P2, P3, P4 and N2). Features could be absolute or relative. For example, feature No.2, AP3(nCX), indicates the P3 amplitude on non-correct X ERP, or feature No.15, LN2P2(X),indicates the difference in latencies of N2 and P3 on X ERP.

In the next step, Linear Discriminant Analysis (LDA) classifier was used for discrimination of each two classes. According to Leave One Out (LOO) method, one subject's data was considered as test data and the classifier was trained based on others, then it was evaluated on the test data. This procedure was repeated for all subjects and accuracy was achieved by the ratio of correct classifications

to total number of subjects. This test was performed for each feature and then repeated for all combinations of two features.

III. RESULTS

Grand average ERPs for all subjects are illustrated in Fig. 2. As it can be seen, amplitude of P3 is obviously larger in target (X) than non-target (nX) stimuli. These curves are smoothed with a 12 Hz low-pass filter. Fig. 3 demonstrates the topographic maps of calculated independent components for one subject.For this subject, component 4 was considered as the best component with brain origins based on its time, frequency and spatial characteristics, thus in method A, only this component was preserved. For method B, component 1 which was considered as EOG component was eliminated and all other components were transferred to electrode space. These maps are obtained using EEGLABsoftware [16]. It should

be noted that achieved components by ICA can be a scaled estimation of original sources (because of ICA limitations) so components shown in Fig.3 are normalized.

Calculated correlations between CPT result and defined sixty four features showed a significant relation (P-value < 0.05) in eighteen features (Table I). Three of these features are amplitude and fifteen are latency. Four features are on X, six on non-X, three on correct X and five on non-correct X groups.

Accuracy for classification on test and train data with one and two features is shown in Table II. Four groups of results are presented on this table. The leftmost group (in pink) shows the results achieved for discrimination between high and low classes of attention. The middle left group (in green) shows the results achieved for classification between high and medium classes of attention. The middle right group (in orange) shows the results achieved for segregation between medium and low classes of attention. The rightmost group (in gray) shows the average of accuracy for all three classifications. The upper part of the table indicates the results for one feature while the beneath part is related to the results of two features. Each part contains the best three features for all three methods, i.e. the first three rows indicate the best three features and their accuracy for method A (the best independent component of brain signals) while the accuracy is achieved with one feature. For example the best feature for H/L discrimination is feature no. 8 which yields in accuracy of 92.9% for both test and train data. This feature can be identified from table 1 which is the difference of latencies between P3 and P2 in non-X group (LP3P2(nX)). Second three rows demonstrate the best three features and their accuracy in method B (EOG Elimination with ICA) and similarly next three rows express the best three features and their accuracy in method C (EOG Deletion). The accuracy for all classes is summarized in Fig. 4. TABLEI FEATURE CHARACTERISTICS Features No. Features No. LP3P2(nCX) 10 AP3 (X) 1 LN2(X) 11 AP3 (nCX) 2 LN2(nX) 12 LP3(X) 3 LN2(CX) 13 LP3(nX) 4 LN2(nCX) 14 LP3(CX) 5 LN2P2(X) 15 LP3P4(AX) 6 LN2P2(nX) 16 AP3P2(nCX) 7 LN2P2(CX) 17 LP3P2(nX) 8 LN2P2(nCX) 18 LP3P2(nX) 9

Features which had significant correlation (P-value<0.05) with sustained attention level are defined. For example feature No.15, LN2P2(X), indicates the difference in latencies of N2 and P3 on X ERP.

IV. DISCUSSION

The P3 peak on target stimuli (X) was clearly observed for all subjects (irrespective of answers correctness) and this is in complete agreement with previous studies [5], [6], [14]. This peak was obviously larger on target than non-target stimuli as can be seen in Fig. 2. This evident could be due to characteristics of task, complete perception of task for subjects and/or the proper ratio of target stimuli to total stimuli (1 to 10).

Independent components for EEG signals were calculated and best component was selected based on its time, frequency and spatial characteristics. Calculated correlations confirmed the significant relations between CPT result and some ERP features. About 83% of these

features were related to latency which also is in agreement with previous studies [13] and 33% of them were on non-X group.

Best result of classification in all classes on test data (81%) was obtained with proposed method using only one feature (LN2P2(X)), as can be seen in Fig. 4. The best result for classification between H/L classes on test data (92.9%) was obtained with proposed method using feature 8, 15 or 17 (LP3P2(nX), LN2P2(X), LN2P2(CX)). It should be mentioned that in this case increasing of features to two did not improve the best result for any of the methods.

In classification between H/M classes, it could be seen that with one feature, accuracy of method A was better than the others, but the best result on test data (88.9%) was obtained with method C using combination of features 1

.

Fig. 3. Topographic maps for one subject. Component 4 is considered as the best component with brain origins based on its time, frequency and spatial characteristics and component 1 is considered as EOG component.

Fig. 2. Grand average ERP for all subjects. ERPs for the proposed method are demonstrated in solid line, whereas ERPs for methods B and C are shown with dash-dotted and dotted line respectively. Four different groups of ERP (X, nX, CX, nCX) are demonstrated on distinct subplots. Amplitude of P3 is obviously larger in target (X) than non-target (nX) stimuli.

and 11 which are related to AP3(X) a amplitude and one latency both on X gr The best result for classification betw test data (100%) was obtained wit combination of features 2 and 12 whic (nCX) and LN2(nX), which were again one latency, but the first one were on and the other on non-X group.

It is noteworthy that the data used in than two other methods, so it seemed increasing the number of epochs and individual characteristics, the accuracy In fact the only difference between resu C was the preserved data, thus for conditions, the results of proposed compared with method B where in m method had a better performance.

In addition, although in 2 by 2 classi method C showed better results, but classes was best obtained with metho means in this method same featu performance in all 2 by 2 classificat methods (B and C), the best feature(s) f 2 classifications differed.

Consequently results represent a between CPT result and some parame which can be used in evaluating the lev ACKNOWLEDGMENT

We thank “Institute for Cognitiv (ICSS) for using their EEG laboratory tests.

Fig. 4. Chart of achieved accuracy for all cla The front row shows the accuracy achieved fo

0 20 40 60 80 100 A B 1 Feature 81 68.3 78.7 6

and LN2(X), i.e. one roup.

ween M/L classes on th method C using ch are related to AP3

n one amplitude and non-correct X group n method C was less reasonable that with thus more different had been decreased. ults for method B and r considering same

method should be most cases proposed ifications, sometimes

the accuracy for all od A (ig. 4), which ure(s) had a good

tions while in other for each of these 2 by

significant relation eters of brain signals vel of attention.

T

ve Science Studies” y for performing the

REFER [1] E. Kandel, J. Schwartz, & T. Je Neural Science.” 4th _{ed. Vol. 1:} [2] I. H. Robertson, T. Manly, J. A

“Oops!: Performance correlate traumatic brain injured and n 24(5), pp. 636–647, 1997. [3] B.S. Oken, M.C. Salinsky, &

sustained attention:physiologic Neurophysiology, 117, pp.1885 [4] H. E. Rosvold, A. F. Mirsky, I. Beck, “A continuous performa Consulting Psychology, 20, 343 [5] E. Molteni, A. M. Bianchi, M.

of the dynamical behaviour o sustained attention,” Proceedin Conference of the IEEE EMB August 23-26, 2007. [6] L. Zordan, M. Sarlo, & F. Stablu

“GO!” and “WITHHOLD!” Attention to Response Task.” 2008.

[7] J. Polich, “Updating P300: An Clinical Neurophysiology, 118 [8] S. Makeig, A. J. Bell, T. -P component analysis of ele Touretzky, M. Mozer, & M.

information processing system

MA: The MIT Press, 1996. [9] M. Potter, N. Gadhok, & W.

ICA On Simulated EEG And E Proceedings of the 2002 IEEE Computer Engineering, 1099:1 [10] A. Delorme, T. Sejnowski,& artifacts in EEGdata using hi component analysis. NeuroIma [11] J. Onton, S. Makeig, “Informat Brain Dynamics, “ Neuper & Research, Vol. 159, Ch. 7, 200 [12] A. Hyvarinen, J. Karhunen,

Analysis-Theory and Applicati [13] S. Makeig, M. Westerfield, J. T T. J. Sejnowski, “Functionally asses. These accuracy is calculated by averaging the results for 2 by or test data and the back row demonstrates the results of train data.

C A B C

2 Features

77.1 78.6 78.6

69.7 _{79.5 83.3 79.9}

RENCES

essell, Visual Attention. “Principles of : Mc Grow Hill, 2000.

Andrade, B. T. Baddeley, & J. Yiend, es of everyday Attentional failures in normal subjects.” Neuropsychologia, S.M. Elsas, “Vigilance, alertness, or cal basis and measurement.” Clinical 5–1901, 2006.

. Sarason, E. D. Jr. Bransome, & L. K. ance test of brain damage.” Journal of

3–350, 1956.

Butti, G. Reni, & C. Zucca, “Analysis of the EEG rhythms during a test of ngs of the 29th Annual International BS Cité Internationale, Lyon, France um, “ERP components activated by the

conflict in the random Sustained Brain and Cognition, 66, pp. 57–64, n integrative theory of P3a and P3b,”

, pp.2128–2148, 2007.

P. Jung, T. J. Sejnowski, Independent ectroencephalographic data. In D.

Hasselmo Eds., Advances in neural ms Vol. 8, pp. 145–151. Cambridge, Kinsner, Separation Performance Of ECG Signals Contaminated By Noise. Canadian Conference on Electrical & 104, 2002.

S. Makeig, Enhanced detection of gher-order statistics and independent age 34, 1443:1449, 2007.

tion-Based Modeling of Event-Related & Klimesch (Eds.), Progress in Brain

06.

E. Oja, “Independent Component ions,” John Wiely & Sons, 2001. Townsend, T-P. Jung, E. Courchesne, & y independent components of the early y 2 classifications for one feature.

Test Train 78.6

event-related potential in a visual spatial attention task.” Philosophical Transactions of the Royal Society: Biological, 1999. Vols. 354 1135-44.

[14] http://read.pudn.com/downloads86/sourcecode/math/333232/combi.

m__.htm

[15] V. Abootalebi, M. Moradi, M. Khalilzadeh, A comparison of methods

for ERP assessment in a P300-based GKT, International Journal of

Psychophysiology,62, pp.309–320, 2006.

[16] http://www.sccn.ucsd.edu/eeglab/

TABLEII

THE ACCURACY ACHIEVED WITH LDA CLASSIFIER

High/Low

High/Medium

Medium/Low

ALL

Feature No.

Test Train

Test Train

Test Train

Test Train

1 Feature

A

15

8

92.9 92.9

92.9 92.9

A

15 75.0 71.3

17 75.0 74.5

A

10 75.0 71.1

15 75.0 69.9

A

15 81.0 78.7

17 81.0 78.0

17

92.9 92.9

12 66.7 61.8

17 75.0 68.7

18 78.2 77.7

B

14 78.6 78.7

4 85.7 87.4

B

16 66.7 69.3

9 75.0 78.3

B

11 75.0 71.6

4 75.0 72.1

B

18 68.3 65.9

15 65.5 69.7

9 71.4 77.0

6 58.3 61.3

12 75.0 71.6

16 65.5 68.3

C

2 84.6 83.9

5 84.6 83.9

C

11 66.7 66.7

5 66.7 66.7

C

2 100 99.3

7 90 92.9

C

13 77.1 79.5

11 73.8 76.0

11 84.6 83.3

13 66.7 60.7

1 80.0 76.2

15 70.8 67.8

2 Features

A

8 9 92.9 92.9

A

15 17 66.7 72.2

A

15 17 83.3 84.7

A

15 17 78.6 83.3

8 12 92.9 92.9

8 17 58.3 65.4

10 15 75.0 77.6

8 17 70.2 76.2

9 15 92.9 92.3

8 18 58.3 62.8

8 15 66.7 76.5

12 15 69.8 74.0

B

4 9 85.7 86.2

4 12 85.7 85.7

B

9 14 83.3 82.3

9 11 75.0 85.7

B

4 15 75.0 74.7

4 18 75.0 74.7

B

9 14 78.6 79.9

9 11 73.0 79.8

4 16 85.7 85.7

4 9 66.7 77.9

6 13 75.0 72.9

9 12 73.0 79.4

C

1 5 84.6 91.7

2 3 84.6 90.9

C

11 14 77.8 85.2

1 11 88.9 91.7

C

2 5 100 100

2 12 100 99.3

C

1 11 78.6 85.3

2 11 77.9 84.5

2 5 84.6 86.9

1 3 66.7 0.3

1 2 90.0 100

11 14 77.5 85.6

The accuracy is calculated for test and train data for classification of each two classes with 1 feature (first nine lines) and 2 features (last nine lines). Last column is the average of all classifications (3 cases) for each feature or combination of features.