2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES)

Classification of Resting and Cognitive States

using EEG-based Feature Extraction and
Connectivity Approach
Moona Mazher Ibrahima Faye Abdul Qayyum
Centre for Intelligent Signal and Imaging Centre for Intelligent Signal and Imaging Centre for Intelligent Signal and Imaging
Research (CISIR), Universiti Teknologi Research (CISIR), Universiti Teknologi Research (CISIR), Universiti Teknologi
PETRONAS, Perak, Malaysia PETRONAS, Perak, Malaysia PETRONAS, Perak, Malaysia
[email protected] [email protected] [email protected]

Aamir Saeed Malik

Centre for Intelligent Signal and Imaging
Research (CISIR), Universiti Teknologi
PETRONAS, Perak, Malaysia
[email protected]

Abstract— Classification of resting and cognitive states has some advantages such as (1) solution with easy mobility,
its importance in brain neuroscience for understating the (2) cost effective, (3) able to detect, view and record
underlying behaviors of cognition. The human brain is electrical activities in the brain, (4) non-invasive testing
considered as a complex system having different mental method, which is safe to be used for human body and (5)
states such as resting, active or cognitive states. It is a well- can record EEG data over a long period of time. possesses
understood fact that the brain activity increases with the a high temporal resolution and it is cheaper as compared to
increased demand of cognition. In this paper, the cognitive other technologies like Positron Emission Topography
and resting state classification based on EEG-based feature (PET) and Functional Magnetic Resonance Imaging
extraction and connectivity approaches are described. EEG-
(fMRI). This makes EEG the most suitable medical
based connectivity approaches are a good discriminator for
modality to investigate memory, learning, perception,
different mental states. EEG data were collected from 34
human participants at resting and during a learning state.
emotion, and information processing tasks [7].
After preprocessing, EEG-based feature extraction method The recorded EEG is called electroencephalogram.
and connectivity approach were implemented, and their One of the methods to record EEG signal is by using a 10-
results were classified. Results showed that the connectivity 20 EEG electrode placement technique established by the
approach gave 79.90% accuracy while the highest accuracy International Federation of EEG Societies. The EEG
achieved by feature extraction approach was 78.50%. It is device records and interprets the voltage fluctuations along
concluded that EEG-based connectivity approach
the scalp which are generated by the brain [8]. The voltage
discriminates the resting and cognitive states more
efficiently.
fluctuations are caused by the current that flows during
synaptic excitations of activated neurons in the cerebral
Keywords— Resting states, Cognitive states, Alpha waves, cortex. Generally, there are a few common steps in order
Feature extraction, Connectivity, Classification. to analyze EEG signal. The raw EEG signal that has been
extracted will undergo preprocessing which includes the
acquisition of signal, removal of artifacts, signal averaging,
I. INTRODUCTION
the threshold value of the output, enhancement of the
resulting signal, and edge detection [8]. The next step is to
The human brain has complex dynamics activity of segment the pure signal using discrete wavelet transform
billions of neurons. It provides the understanding of these to determine the coefficients of the signal in different brain
dynamics in the form of interconnected ways that are waves.
formed by the signals from human brain [1]. The human
brain comprises of different states e.g. resting, active and Metal electrodes are used to record the EEG data
cognitive. The investigation and classification of these signals on the scalp surface. EEG provides us the
states to measure the brain subtleties is a big challenge. important information of neural activity that occurs in the
Electroencephalography (EEG) and functional magnetic brain. The brain engaged in various activity during the day
resonance imaging (fMRI) considered as effective time in our daily life [9-10]. By visual inspection, the EEG
modalities to measure the neural activity of the human data acquired from the brain produced the same
brain non-invasively [2] and these modalities are actively information in resting and active states. For accurate
used in neuroscience and neuroimaging. The EEG has information about the states, the different classification

978-1-5386-2471-5 ©2018 IEEE 184

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on January 06,2024 at 15:27:34 UTC from IEEE Xplore. Restrictions apply.
 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES)

methods are required to classify these states efficiently. were lying in relaxing mode and in cognitive states, the
The brain states classification will be helpful in the brain participants have presented learning sessions three times.
computer interface applications and, could be used for the A total of 34 subjects were used for this experiment. The
prosthesis limbs control with brain signals. Several feature 128-channel EEG device used for EEG signal recording
extraction schemes exist in literature, including time was HydroCel Geodesic Sensor Net (Electrical Geodesic
domain, frequency domain, and entropy-based features. Inc., Eugene, OR, USA. The device impedance was kept
Among these feature extraction methods, entropy-based below 50 KΩ with 250 Hz sampling rate. The frequency
features work better for non-stationary EEG signals, range between 0 and 100 Hz was extracted using band-
making it highly effective. The entropy-based features that pass filter method.
are used for EEG analysis include Spectral-Entropy and Duration of data recordings was about 5 minutes for
Approximate-Entropy. each resting and cognitive states. These recordings were
taken on through ten channels (9, 22, 33, 24, 11, 122, 124,
Other than feature extraction methods in EEG, the 52, 62 and 92) of EEG. These channels belong to frontal
concept of brain connectivity is a well-recognized topic in and parietal areas of the brain which are related to the
neuroscience as it provides the information related to human working memory. For this experiment, we need to
cortical regions’ communication between each node. It is a
follow the standard procedure of an ethical approval
basic understanding of the organized behavior of the approved by the ethical committee of Universiti Teknologi
cortical regions beyond the simple mapping of their Petronas Malaysia. All participants were under 24-30 age
activity [3]. In various studies, it has been noted that how limit. Participants with normal vision and corrected to
promptly fluctuated and extensively distributed neural normal visions were taken and anyone of them has no
activation patterns change during the human recognition neurological disorder. A complete methodology of
and perception [11]. During human recognition process, proposed work is presented in Fig.1.
various cortical, subcortical regions, and their After preprocessing of artifacts removal from EEG
combinations have been activated. Thus, a connectivity- data, different brain waves such as delta, theta, alpha and
based method could yield better results than the widely beta waves were extracted for the further analysis. The
used feature extraction approaches [12]. discrete wavelet transform (DWT) is an excellent
This work focuses on EEG-based feature extraction technique used to decompose the signal in time and
methods and connectivity approach to classifying the frequency domain. The DWT technique provides the better
resting and cognitive states to analyze brain activity time-frequency decomposition of the non-stationary
changes from a resting to cognitive state. signals as compared to the FFT technique. It
simultaneously decomposes the signal into various time -
II. MATERIAL AND METHODS frequency slots. Consequently, DWT can be used to
decompose a signal into various time-frequency
The data used for this experiment was based on resting components [4].
state with the eyes-closed condition, where participants

EEG Data Acquisition

Pre-processing

Brain Waves Extraction

Feature Extraction Effective Connectivity

(Spectral & Approximate Estimation (PDC)
Entropy)

Classification (ELM)

Conclusion

Fig.1. The proposed methodology for classifying the resting and cognitive states using EEG base feature extraction and connectivity approach.

978-1-5386-2471-5 ©2018 IEEE 185

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on January 06,2024 at 15:27:34 UTC from IEEE Xplore. Restrictions apply.
 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES)

It gives the representation of the time scale of a given

signal by using the scaling and translation of a signal based
on the mother wavelet. It decomposes structure into Classification of Spectral Entropy Feature
different octaves in a sequential way by using scaling and
dilation function embedded in mother wavelet. The level
of decomposition of DWT coefficients totally rely on the
sampling frequency of the input signal. In this experiment,
the decimation number goes up-to fifth level as the last
desired wave (delta) was achieved on that level.

III. RESULTS AND ANALYSIS

The raw EEG data comprises of various artifacts such

as heart rate variation, eye blinking, muscle movement and Fig.2 . An ELM based classification accuracy of resting and cognitive
impedance fluctuations. The (BESA) (6) software has been states for spectral entropy.
used to clean the brain electrical signals by removing any
artifacts from raw EEG data and a clean signal was Fig. 2 the classification accuracy plots for delta, theta,
received for the specific resting and cognitive tasks. After alpha and beta brain waves for the combination of EC, L1,
preprocessing, two entropy-based features, spectral L2, and L3. The results show that for classifying the
entropy [5] and approximate entropy were extracted for the resting and all mental states the alpha waves produced the
resting (EC) and three cognitive states (L1, L2, L3). highest accuracy of 78.50% as the alpha waves are the
most prominent brain waves in discriminating the resting
The statistical p-test has been used on the extracted and conscious or active brain states [6]. Table 2 shows the
features for checking the feature discrimination using approximate entropy based ELM classification accuracy
repeated ANOVA software. It provided a p < 0.05 for the score of resting and cognitive states. The highest
comparison between EC and L1-based features. classification is achieved by alpha brain waves in
Comparatively similar results has been obtained in the L2 classifying all resting and learning states as seen in Fig.3.
and L3 states against resting state. The p-values show that Thus, from the results of both feature extraction technique,
significant differences existed between the features of it was observed that alpha waves give the highest
resting and cognitive states, indicating that they could be classification accuracy. Also, it was seen the approximate
further used for the classification of resting and cognitive entropy gives better results than the other entropy based
states of mind using the accuracy of machine-learning technique of spectral entropy.
algorithms.
Table 2. An approximate entropy based ELM classification accuracy
An extreme learning machine classifier was applied to for resting and cognitive states.
classify the extracted features. The classification was
Classification Accuracy (%)
performed for each cognitive (learning) state and resting
state combination and also for a combined grouping of Brain
resting and cognitive states. Table 1 shows the ELM Waves L1EC L2EC L3EC EC L1L2
classification accuracies of resting and cognitive states for L3
five extracted waves.
Delta 80.73 83.76 86.36
Table 1. A spectral entropy based ELM classification accuracy for 77.57
resting and cognitive states.
Theta 85.79 84.89 81.17
75.70
Classification Accuracy (%)
Alpha 86.58 84.89 81.67
78.50
Brain
Waves L1EC L2EC L3EC EC L1L2 Beta 81.23 81.66 88.3
L3 76.16

Delta 81.8 81.17 83.12 77.57 Classification of Approximate Entropy

Theta 84.4 85.77 87.34 76.64

Alpha 89.5 84.67 80.34 78.17

Beta 83.7 82.87 87.67 76.26 Fig.3 . An ELM based classification accuracy of resting and cognitive
states for approximate entropy.

978-1-5386-2471-5 ©2018 IEEE 186

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on January 06,2024 at 15:27:34 UTC from IEEE Xplore. Restrictions apply.
 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES)

Table 3. PDC connectivity results of alpha waves for resting and L3) to further classify the resting and cognitive states. A
cognitive states. connectivity density factor was calculated for all resting
and cognitive states for all extracted brain waves like delta,
Resting Cognitive states theta, alpha, and beta. The connectivity among the frontal
state and parietal EEG channels was calculated using PDC. In
Channels
PDC analysis, greater numbers of connections mean more
EC L1 L2 L3 brain areas are interconnected.
22 6 13 11 13 Table 5. PDC connectivity results of theta waves for resting and cognitive
states.
9 6 13 9 16
Resting Cognitive states
24 5 7 12 15 state
Channels
124 3 15 10 16 EC L1 L2 L3

33 0 14 13 15 22 0 13 7 5

122 0 14 13 17 9 0 12 3 13

11 5 13 14 14 24 0 6 4 7

52 7 7 15 17 124 6 5 4 11

62 0 6 8 16 33 4 9 12 9

92 1 12 15 13 122 0 9 8 7

11 5 9 7 6

Table 4. PDC connectivity results of beta waves for resting and cognitive 52 7 4 3 2
states.
62 0 4 8 3
Resting Cognitive states
state 92 0 11 2 3
Channels
EC L1 L2 L3

22 8 4 8 15 Table 6. PDC connectivity results of delta waves for resting and cognitive
states.
9 6 4 7 16
Resting Cognitive states
24 7 4 8 10 state
Channels
124 9 12 10 14 EC L1 L2 L3

33 4 11 10 14 22 9 16 16 15
122 0 11 9 13 9 1 16 17 15
11 5 9 8 12 24 0 9 10 9
52 3 3 3 12 124 2 7 12 15
62 5 3 5 6 33 0 14 15 15
92 9 3 6 14 122 0 14 12 16

11 6 15 9 15

After traditional feature extraction method 52 5 13 16 7

implementation, the connectivity approach was deployed
on the clean EEG data. In the connectivity approach, the 62 0 6 16 10
partial directed coherence method was used to analyze the
differences between resting and cognitive states. 92 0 13 16 15
Coherence values were extracted from the clean EEG
signals for resting (EC) and cognitive states (L1, L2, and

978-1-5386-2471-5 ©2018 IEEE 187

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on January 06,2024 at 15:27:34 UTC from IEEE Xplore. Restrictions apply.
 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES)

IV. CONCLUSION
Table 3, Table 4, Table 5 and Table 6 show the PDC Feature extraction and connectivity based classification
coherence results for resting (EC) and cognitive states (L1, of resting and cognitive states have been discussed in this
L2, and L3) based alpha, beta, theta and delta brain waves, work. The work demonstrated which EEG frequency
respectively. An ELM based classification was performed bands, and which methods were best for classifying the
on the PDC generated results for resting and cognitive resting and cognitive states. The results showed that alpha
states. Table 7 shows the classification score of resting and waves generated highest classification accuracy in both
cognitive states for four abovementioned brain waves on implemented methods. The results presented that PDC
the PDC results. Highest classification accuracy achieved generated higher classification scores than other traditional
by alpha brain waves of 79.9. %. Which shows that alpha feature extraction approaches like approximate entropy
waves are a better discriminator of resting and cognitive and spectral entropy. These findings indicate that
states. connectivity-based PDC methods are more reliable
Table 7. An ELM based classification accuracy scores for PDC results of approaches to classify the resting and cognitive states.
resting and cognitive states for all brain waves. Future work will explore the other machine learning and
deep learning methods and will also expand the number of
Classification Accuracy (%) electrodes or channels over different brain regions to
Brain Waves classify the diverse mental states.
ECL1L2L3
Delta 71.31 ACKNOWLEDGMENT
Theta 74.47 This research work was supported by the URIF grant
(015LB0-004), Universiti Teknologi PETRONAS,
Alpha 79.90 Malaysia

Beta 69.44

REFERENCES
This study focuses on the classification of resting and [1] R. E. Mayer and R. Moreno, “Animation as an Aid to Multimedia
cognitive states. In experiment design, the three different Learning,” Educ. Psychol. Rev., vol. 14, no. 1, pp. 87–99, 2002.
learning and eye close states were used to collect the EEG [2] P. Ayres, “Impact of reducing the intrinsic cognitive load on
learning in a mathematical domain,” Appl. Cogn. Psychol., vol. 20,
data from the participants by showing multimedia learning no. 3, pp. 287–298, Apr. 2006.
content. The DWT has been used to extract four different [3] D. Levy, “How Dynamic Visualization Technology can Support
frequency bands (alpha, beta, theta, and delta) of the brain Molecular Reasoning,” J. Sci. Educ. Technol., vol. 22, no. 5, pp.
signals. The spectral features (i) spectral entropy and (ii) 702–717, Oct. 2013.
approximate entropy have been extracted for the [4] F. Paas, J. E. Tuovinen, H. Tabbers, and P. W. M. Van Gerven,
classification. ELM classifier was applied to the extracted “Cognitive Load Measurement as a Means to Advance Cognitive
feature values. Based on the classification results, the alpha Load Theory,” Educ. Psychol., vol. 38, no. 1, pp. 63–71, Mar.
2003.
waves produced highest classification accuracy for all
[5] S. E. Gathercole and A. D. Baddeley, “Phonological memory
spectral features. The alpha waves shows more deficits in language disordered children: Is there a causal
discrimination in the feature space indicate that alpha connection?,” J. Mem. Lang., vol. 29, no. 3, pp. 336–360, Jun.
waves produced best results for classifying the cognitive 1990.
and rest states. These results also indicate that the [6] J. Virtanen, J. Ruohonen, R. Näätänen, and R. J. Ilmoniemi,
classification models based on the proposed features “Instrumentation for the measurement of electric brain responses to
provided better classification performance in rest and transcranial magnetic stimulation,” Med. Biol. Eng. Comput., vol.
37, no. 3, pp. 322–326, May 1999.
cognitive states. Hence the proposed method could be used
to classify specific resting and cognition-based tasks. [7] M. Othman and A. Wahab, “Affective face processing analysis in
autism using electroencephalogram,” in Proceeding of the 3rd
The second part of the analysis focused on determining International Conference on Information and Communication
Technology for the Moslem World (ICT4M), pp. E23–E27, 2010.
the effective connectivity for each of the resting and
cognitive states using PDC. PDC adjacency matrices were [8] J. Lamberts, P. L. C. van den Broek, L. Bener, J. van Egmond, R.
Dirksen, and A. M. L. Coenen, “Correlation Dimension of the
calculated for each of the participants for all ten selected Human Electroencephalogram Corresponds with Cognitive Load,”
channel for above-mentioned brain waves. The findings Neuropsychobiology, vol. 41, no. 3, pp. 149–153, 2000.
show that alpha waves produced high classification [9] Amin, Hafeez Ullah, et al. "Feature extraction and classification
accuracy in the case of alpha waves. This method for EEG signals using wavelet transform and machine learning
generated better classification accuracy scores than techniques."Australasian Physical & Engineering Sciences in
traditional feature-based methods. These findings indicate Medicine, vol. 38, pp. 139-149, 2015.
that PDC approach successfully classifies the resting and [10] Faust, Oliver, et al. "Wavelet-based EEG processing for
computer-aided seizure detection and epilepsy diagnosis." Seizure,
cognitive states. Thus, EEG-based effective connectivity pp. 56-64, 2015.
method is a better solution for classifying the resting and [11] Nicholson, Charles. "Diffusion and related transport mechanisms
cognitive states as well as different mental states. in brain tissue." Reports on progress in Physics, vol. 64 pp. 815,
2001.
[12] Lotte, Fabien, et al. "A review of classification algorithms for
EEG-based brain–computer interfaces." Journal of neural
engineering vol. 4, 2007.

978-1-5386-2471-5 ©2018 IEEE 188

Authorized licensed use limited to: University of the Phillippines Diliman. Downloaded on January 06,2024 at 15:27:34 UTC from IEEE Xplore. Restrictions apply.