Automatic EEG Processing for the Early Diagnosis of Traumatic Brain Injury

Abstract

Traumatic Brain Injury (TBI) is recognized as an important cause of death and disabilities after an accident. The availability a tool for the early diagnosis of brain dysfunctions could greatly improve the quality of life of people affected by TBI and even prevent deaths. The contribution of the paper is a process including several methods for the automatic processing of electroencephalography (EEG) data, in order to provide a fast and reliable diagnosis of TBI. Integrated in a portable decision support system called EmerEEG, the TBI diagnosis is obtained using discriminant analysis based on quantitative EEG (qEEG) features extracted from data recordings after the automatic removal of artifacts. The proposed algorithm computes the TBI diagnosis on the basis of a model extracted from clinically-labelled EEG records. The system evaluations have confirmed the speed and reliability of the processing algorithms as well as the system's ability to deliver accurate diagnosis. The developed algorithms have achieved 79.1% accuracy in removing artifacts, and 87.85% accuracy in TBI diagnosis. Therefore, the developed system enables a short response time in emergency situations and provides a tool the emergency services could base their decision upon, thus preventing possibly miss-diagnosed injuries.

Full text

Available online at www.sciencedirect.com
ScienceDirect
Procedia Computer Science 00 (2016) 000–000
www.elsevier.com/locate/procedia
1877-0509 © 2016 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of KES International.
20th International Conference on Knowledge Based and Intelligent Information and Engineering
Systems, KES2016, 5-7 September 2016, York, United Kingdom
Automatic EEG processing for the early diagnosis of Traumatic
Brain Injury
Bruno Alberta, Jingjing Zhanga, Alexandre Noyvirtb, Rossitza Setchia, Haldor Sjaaheimb,
Svetla Velikovab, Frode Strislandc
aSchool of Engineering, Cardiff University, Cardiff CF24 3AA, UK, Applied Automation, Bridgend, UK, c Smartbrain, Oslo, d SINTEF ICT, Oslo
Norway
Abstract
Traumatic Brain Injury (TBI) is recognized as an important cause of death and disabilities after an accident. The availability a tool
for the early diagnosis of brain dysfunctions could greatly improve the quality of life of people affected by TBI and even prevent
deaths. The contribution of the paper is a process including several methods for the automatic processing of electroencephalography
(EEG) data, in order to provide a fast and reliable diagnosis of TBI. Integrated in a portable decision support system called
EmerEEG, the TBI diagnosis is obtained using discriminant analysis based on quantitative EEG (qEEG) features extracted from
data recordings after the automatic removal of artifacts. The proposed algorithm computes the TBI diagnosis on the basis of a
model extracted from clinically-labelled EEG records. The system evaluations have confirmed the speed and reliability of the
processing algorithms as well as the system’s ability to deliver accurate diagnosis. The developed algorithms have achieved 79.1%
accuracy in removing artifacts, and 87.85% accuracy in TBI diagnosis. Therefore, the developed system enables a short response
time in emergency situations and provides a tool the emergency services could base their decision upon, thus preventing possibl y
miss-diagnosed injuries.
© 2016 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of KES International.
Keywords: Artifact removal, Diagnosis; Electroencephalography (EEG); Portable Medical System; Traumatic Brain Injury (TBI).
1. Introduction
Traumatic brain injury (TBI) is caused by an external force that damages the brain. This brain dysfunction results
as possible physical, cognitive, social, emotional, and behavioral effects on the subject1. The severity of the injury
ranges from mild to severe as well as the associated impacts on the quality of life of the person with TBI2,3,4,5. TBI has

2 Author name / Procedia Computer Science 00 (2016) 000–000
been recognized as an important cause of death in the US6 as well as in Europe7. Moreover, it leads to a great economic
burden7.
Irreversible brain damages can result from a trauma that is not properly diagnosed, or too late. Hence there is a
need for a reliable tool that can be used by emergency services in order to obtain a quick diagnosis of TBI at the place
of injury. However, current methods and devices that provide TBI diagnosis are limited to clinical environments. In
particular, contrary to other medical imagery technologies, Electroencephalography (EEG) techniques have the
potential for being used in a portable way. In addition, Quantitative Electroencephalography (qEEG) is a sensitive
diagnostic method of brain injury after mild head injury. It has shown over 80% accuracy in discriminating between
normal and traumatic brain-injured subjects2,3,4.
The EmerEEG project addresses this problem by proposing a portable decision support system based on EEG
technology for early diagnosis of TBI at the point of need. This system includes a head device for fast and simple
acquisition of EEG data during emergencies, as well as necessary devices enabling processing power, interfacing and
communication capabilities. This paper focuses on the processing part of the system, which, once integrated to the
rest of the system, provides a tool for the automatic diagnosis of TBI and decision support. The idea is to enable
anyone from the emergency services with minimal training to assess the severity of a brain injury.
The remainder of the paper is organized as follows. Related processing and diagnostic techniques, are reviewed in
section 2. Section 3 outlines the EEG processing method and TBI diagnosis. Section 4 describes the evaluation of the
system in terms of the quality of the EEG pre-processing and TBI diagnostics. Finally, section 5 summarizes the paper
and highlights future work.
2. Literature review
This section reviews methods for EEG data processing and TBI diagnosis.
The clinical criterion most widely used to classify TBI severity is the Glasgow Coma Scale (GCS), which grades
the condition of a patient on a scale from 3 to 15 based on verbal, motor, and eye reactions to stimuli8,9. However, the
GCS is a qualitative method of assessment, which has its limitations. Advanced neuroimaging techniques like
Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) are now widely used in hospitals for the
assessment of neurological damage. The size and non-portability of the equipment, in addition to their limitations in
diagnosing mild TBI10,11,12, however, constrain their use in portable systems. By comparison, the EEG technique
provides a direct measurement of brain activity without the need for external radiation or injected substances.
Rather than only analyzing raw recordings through visual inspection, the extraction of quantitative EEG data, such
as frequency and coherence, has shown its relevance in recent years in supplying relevant and re-producible features
for the development of diagnostic tools13. The discriminant accuracy of qEEG is reported as 95.67% in the detection
of mild head injury3 and 75.8% in predicting the outcome one year after the injury14. Moreover, qEEG demonstrates
96.39% classification accuracy, 95.45% sensitivity and 97.44% specificity in discriminating between groups with
mild and severe TBI4. The EEG discriminant score is also used to measure intermediate severity in moderate TBI
patients. Significant correlations between EEG discriminant scores, emergency admission measures, and post-trauma
neuropsychological test scores have validated the discriminant function as an index of severity of injury and a classifier
of the extremes of severity4.
The procedure for computing a TBI diagnosis using EEG data normally involves pre-processing the raw recording
to reduce the impact of the low signal-to-noise ratio and to obtain a more accurate representation of the pure brain
activity. Artifacts are the most important cause of noise once errors directly due to the instrumentation have been
eliminated. Artifacts are electrical signals detected along the scalp that do not arise from the cerebra. Typical artifacts
include electrocardiography (ECG) artifacts caused by heart beats15, ocular artifacts (EOG) caused by eye blinks or
low-frequency patterns caused by eye movements16,17, and muscle activity (EMG) caused by movements of the head,
body, jaws, or tongue. EOG and EMG activities are unavoidable in EEG recording16,17,18. Conventional clinical
approaches reduce noise by discarding epochs with artifacts through visual inspection by specialists. This manual
process is time-consuming and subject to intra-observation differences, and useful information of the brain activity
embedded in the discarded epochs might be lost.
An effective and popular alternative is the use of Independent Component Analysis (ICA), which separates the
artifacts from the EEG signals without removing epochs19,20. However, components corresponding to the artifacts

Author name / Procedia Computer Science 00 (2016) 000–000 3
have to be carefully selected for this method to be effective. Therefore, the key to achieving automatic artifact removal
is to find a method that automatically selects artifact components from the brain activity after separation with ICA.
Spatial, spectral, temporal, and statistical features have been combined to identify artifacts15,21,22,23,24. An alternative
approach to automatic artifact removal is a novel technique, named Automatic Wavelet Independent Component
Analysis (AWICA)25. It combines wavelet transform and ICA based on the estimation of kurtosis and Renyi’s entropy.
This is done in a two-step procedure, instead of applying wavelet analysis after ICA26. One important advantage of
this method is that it suppresses artifact components while reducing the loss of residual informative data, since the
components related to relevant EEG activity are mostly preserved25.
3. EEG processing and TBI diagnosis
3.1. Offline and online operation
This section describes the proposed algorithms for automatic detection of traumatic brain injury. The algorithms
are based on data signal processing techniques and a classification approach. The aim is to alert the local operator and
the remote telemedicine personnel when TBI is detected. The flowchart in Fig. 1 shows the processes for computing
the TBI diagnosis online and for training the model offline using pre-recorded data. The two processes share the same
pre-processing and qEEG feature extraction steps. The offline process performs these steps for all recordings from the
clinical database and constructs a model using machine learning, whereas the online process applies these steps on the
continuous recording coming from the portable sub-system and extracts a predicted diagnosis from the trained model.
Fig. 1. Offline model training and online EEG processing and TBI diagnosis.
The continuous EEG acquisition and online processing starts after the montage of the electrodes on the patient’s
head is completed and the electrical contact is assured. The raw EEG recording is first pre-processed by filtering high
frequency noise and removing artifacts. Next, qEEG features within four frequency bands are calculated from this
‘clean’ recording. In particular, 16 features that have been proven discriminant in the detection of TBI3 are used in the
diagnosis step. A classification prediction is performed using discriminant analysis and the model extracted from the
previously recorded data. Detailed information about data processing and classification methods is given in the
following sections for both online and offline operations.
3.2. Continuous EEG acquisition
Clinical best practice recommends the use of at least 60 seconds of artifact-free EEG27,28. This has been confirmed
by a systematic analysis using the EmerEEG system, which has shown that one minute long epochs are sufficient for
obtaining reliable diagnosis results. Files containing segments of one minute EEG data are stored according to the
European Data Format (EDF) standard29 and loaded for processing. The diagnosis process starts when the first
segment of one minute EEG data is collected. In addition to montage verifications, a fault alert mechanism in the

4 Author name / Procedia Computer Science 00 (2016) 000–000
processing algorithm detects faults in signals for each channel to ensure that the diagnosis is based on reliable data.
The electrodes are positioned following the 10-20 rule employed in EEG best practice and the montage used here is
linked-ear.
Consider, as an example, the piece of raw EEG data shown in Fig. 2. According to the annotations made by a
clinical specialist, this one minute segment contains eye blinking and electrode movement artifacts. Eye blinking
artifacts mainly appear around 1, 24, 37, 43, and 56 seconds. Electrode movements occur around 2, 12, 40, and 50
seconds. These artifacts have higher amplitude and frequency compared to the brain signals. This points to the need
for a pre-processing method. The next section describes the algorithm employed to remove automatically such
artifacts.
Fig. 2. Example of a one minute segment of raw EEG data.
3.3. EEG Pre-processing
In EEG data, the signal-to-noise ratio is usually low and the noise frequency and amplitude are higher than the
brain signals. On the other hand, since the analysis of the frequency bands in the range between 0 and 30 Hz can be
sufficient for detection of TBI3,4, the recording is first filtered by a low pass filter with high limit of 30 Hz.
The next step is the removal of artifacts, including those resulting from eye blinks, eye movements, electrode
movements, muscle activity, drowsiness, and head movements. The method adopted for performing the automatic
removal of artifact is the Automatic Wavelet Independent Component Analysis (AWICA)25. The flowchart of the
algorithm21,25 is shown in Fig. 3.
The algorithm consists of five steps:
(1) Wavelet component (WC) extraction. Each channel of the filtered recording is divided into four frequency
bands (delta, theta, alpha and beta) using a four-level Discrete Wavelet Transform (DWT). Each band of each channel
is represented using a Wavelet Component.
(2) Critical WC selection. WCs corresponding to artifacts are automatically identified through a quantitative
measure as critical WCs. The selection is based on kurtosis and Renyi’s entropy, which measure randomness and
peakyness of the signals. Given a scalar random variable x, kurtosis is expressed as   
 , where  is the
nth order central moment of the variable     and  is the mean value. The kurtosis values are first
normalized to zero mean and unit variance, and those with values larger than the threshold ±1.5 (the value is found by
trial and errors) are selected as critical WCs. The recommended value of order for Renyi’s entropy is 2. WCs with
entropy larger than the threshold ±1.5 after normalization are also selected as critical WCs. Fig. 4 shows an example
of critical WC selection from the recording shown in Fig. 2. A total of 76 WCs are generated from this 19 electrode
recording, following four frequency bands (delta, theta, alpha, and beta). With a threshold of ±1.5, the red bars in Fig.
4 indicate the identified critical WCs.
(3) Wavelet independent component (WIC) extraction. The selected critical WCs are then passed to ICA to separate
the artifactual WICs. This system adopts the FastICA 2.530 algorithm for Matlab. Fig. 5 shows an example of WICs
obtained after application of ICA on the selected critical WCs. Artifacts with similar pattern as those annotated by the
specialist in Fig. 2 can be identified independently.

Author name / Procedia Computer Science 00 (2016) 000–000 5
Fig. 3. AWICA artifacts removing algorithm (adapted2 1,25)
(4) Artifactual WIC selection. This step concentrates on removing one or more WICs remained after using ICA.
The selection of these artifactual WICs is also based on kurtosis and entropy. The difference in this step is that the
WIC dataset is first divided into 0.5s non-overlapping windows (trials), and then the kurtosis and entropy are
calculated based on these trials. WICs with more than 20% of the trial’s kurtosis and entropy above the threshold ±1.5
are rejected.
(5) Reconstruction. The remaining WICs are then used to project back artifact-free WCs with an inverse ICA and
combined with the non-critical WCs. The result is WCs cleaned from artifacts. Performing an inverse DWT enables
the reconstruction of an artifact-free EEG recording.
Fig. 4. Example of critical WC selection (threshold=±1.5).

6 Author name / Procedia Computer Science 00 (2016) 000–000
Fig. 5. Example of EEG data after application of ICA.
3.4. qEEG Feature Extraction
qEEG is a numerical analysis of EEG data using signal analysis techniques such as wavelet analysis and Fourier
analysis. The commonly used features are EEG coherence, phase, power, and amplitude. The features are calculated
based on artifact-free recordings for eight frequency bands: delta (δ, 1 to 4 Hz), theta (θ, 4 to 8 Hz), alpha (α, 8 to 12
Hz), beta (β, 8 to 25 Hz), hi-beta (hi-β, 25 to 30 Hz), beta1 (β1, 12 to 15 Hz), beta2 (β2, 15 to 18 Hz), and beta3 (β3,18
to 25 Hz). The Fast Fourier Transform (FFT) converts signal  from the time to the frequency domain:
   (1)
The auto-spectrum AP at frequency f is:
      , (2)
where   and   are the cosine and sine coefficients at frequency  for signal . The amplitude A can
then be obtained as the square root of the auto-spectrum:   
. The amplitude asymmetry AA31 for two signals
 and  is calculated as
   , (3)
where          and     
   
Next, the coherence CO is computed for pair-wise combination of electrodes x and y,
  

 , (4)
The phase difference PH is then computed:
   
 . (5)
These calculations allow the 16 features identified by Thatcher4 as discriminant in TBI diagnosis to be employed
in the algorithm. The labels used in the features follow the 10-20 coding system used in EEG research and practice
that indicate the position of the electrodes on the sculp. The letters F, T, C, P and O stand for frontal, temporal, central,
parietal, and occipital lobes, respectively. Even numbers refer to electrode positions on the right hemisphere, whereas
odd numbers denote those located on the left hemisphere. The number zero represents an electrode placed on the
midline. The features employed in the research are the those selected by Thatcher3.

Author name / Procedia Computer Science 00 (2016) 000–000 7
3.5. Dataset Construction and Model Training
When new field data is recorded, the TBI diagnosis is obtained by performing a classification prediction based on
a comparison of the extracted vector of qEEG features defined above with the trained model. The discriminant analysis
uses two classes: TBI and normal. The model is trained with the same qEEG features, extracted from EEG data
previously recorded in a clinical setting. The dataset is composed of EEG data recorded from 21 electrodes (2
electrodes used as references) at a sample rate of 256 Hz using a BrainMaster device32. Recordings include data with
patients’ eyes open and closed. The recordings are annotated by specialists with labels corresponding to the
International Statistical Classification of Diseases (ICD-10) system33, thereby providing a ground truth for the
classification.
The recordings used for training the model in this study have labels F07.2 and Avrg. The label F07.2 corresponds
to a post-concussional syndrome, i.e., patients diagnosed with TBI. The label Avrg stands for average healthy subjects.
In total, 288 recordings from 14 patients (8 female and 6 male) have been used, including 251 recordings labelled as
F07.2 and 37 regarded as Avrg. After pre-processing and extraction of the discriminant features, the training dataset
was constructed as an Nx16 matrix with N being the total number of clinical samples. A model with good
generalization performance was obtained by splitting this dataset into a training dataset for model training and a
validation dataset for evaluating the model.
As mentioned, the proposed method builds the model by performing a discriminant analysis. The relation between
the selected qEEG features and the classes is assumed to follow a multivariate normal distribution. The mean of each
feature is calculated for each class. The covariance is also calculated, after first subtracting the mean. Considering a
linear discriminant analysis, the model has the same covariance for each class, only the means vary. No prior
probabilities or costs are used to compute the model, as the labels define the class to which each sample belongs. The
trained model is then used to predict the classification of newly acquired data. The principle is to find the class with
the highest probability that the new sample belongs to. The obtained classification is then returned for each segment
considering an online situation. This classification of multiple segments allows better precision to be achieved in the
diagnosis.
4. Evaluation
This section evaluates the developed system in terms of its EEG processing and TBI diagnosis.
The proposed algorithm for TBI diagnosis has two main functions: pre-processing of EEG data, including an
automatic artifact removal, and the diagnosis of TBI itself, based on discriminant qEEG features and the comparison
with a model trained on previously collected and annotated data. The performance of the artifact-removing algorithm
has a great impact on the final diagnosis due to the high signal-to-noise ratio. Thus, in this evaluation, the performance
of the pre-processing algorithm is presented first, followed by the performance of the classification algorithm for TBI
diagnosis.
4.1. Evaluation of the pre-processing method
For the purpose of evaluating the performance of the pre-processing algorithm in removing different types of
artifacts, 20 randomly selected recordings were first visually inspected by a highly qualified specialist, who annotated
each artifact with its type. In total, 225 artifacts of 6 different types were found by the specialist (Table I); these include
eye blinking (138 instances), eye movement (48), electrode movement (14), muscle activity (3), drowsiness (3), and
head movement (19). The dataset is representative as the total number of annotated eye blinks in these recordings is
more than half of all marked artifacts, while the number of muscle activities and drowsiness periods is very low as
these are easier to control and eliminate during the recordings.
Table 1. Artifact Annotation by a specialist and percentage of removal by the pre-processing algorithm
Type of artifact
Eye Blinking
Eye Movement
Electrode Movement
Muscle Activity
Drowsiness
Head Movement
Total
Total
138
48
14
3
3
19
225
Removal %
84.8%
83.3%
85.7%
0%
66.7%
36.8%
79.1%

8 Author name / Procedia Computer Science 00 (2016) 000–000
The artifact-removal algorithm was applied to the same 20 recordings. Table I shows the results. The results show
that 79.1% of the annotated artifacts have been successfully removed. Most of the eye blinks (84.8%), eye movements
(83.3%), and electrode movements (85.7%) have been eliminated. These three groups represent the majority of
artifacts in the dataset (88.88%). The percentage of successful removal of artifacts related to drowsiness and head
movement is lower, at 66.7% and 36.8%, respectively. The algorithm has failed to remove any artifacts related to
muscle activity; experiments with more data is needed to improve this parameter.
Fig. 6 shows signals obtained after pre-processing the raw segments shown in Fig. 2. The comparison between Fig.
6A, B and C shows that the success rate in removing artifacts depends on the threshold value. Small threshold values
may result in rejection of some elements of the brain signals (see Fig. 6A, threshold ±1) while high threshold values
may result in low success rate in artifact removal (Fig. 6C, threshold ±2). Fig. 6B is obtained with the optimal threshold
of ±1.5.
Fig. 6. Artifact-free epochs: A, threshold = ±1; B, threshold=±1.5; C, threshold = ±2
5. Conclusion
This paper proposes a method for an automatic early diagnosis of TBI in emergency situations. The system is based
on state-of-the-art standards and advanced technologies for processing and intelligent diagnosis. The system has been
specifically developed in response to real needs: fast and reliable assessment of possible brain injury where the
accident occurred.
The development of the automatic TBI diagnosis algorithm is based on advanced EEG signal processing and
machine learning techniques. The pre-processing step of the algorithm enables the automatic removal of artifacts and
noise, avoiding the need for a time-consuming manual inspection and removal of data segments. The diagnosis is
computed using supervised machine learning based on clinical data. The system operator is then provided with an
assessment of the possible patient’s traumatic brain injury.
The evaluation of the proposed algorithms has shown it to be fast and reliable, with a good generalization
performance of the model. The result of the automatic diagnosis, coupled with a decision support within the EmerEEG
system, provides the operator with an effective basis for the early application of an adapted treatment in an emergency
situation.
Currently, the data stream from the head device is simulated with previously recorded data. The actual testing with
humans is beyond the scope of this project. Future work includes integration of the head device with the portable
system and clinical evaluations once medical approval is obtained.
Acknowledgement
The authors thank the European Commission for funding this research (Grant 605103, FP7-SME-2013). They also
thank their partners Maytec, neuroConn, and University Hospital Gottingen from Germany, Primasil, UK and Tallinn University,
Estonia.

Author name / Procedia Computer Science 00 (2016) 000–000 9
References
1. Saatman K. E., Duhaime A., Bullock R., Maas A. I., Valadka A., Manley G., Classification of traumatic brain injury for targeted therapies,
Neurotraum, vol. 25, no. 7, pp. 719-738, Nov. 2010.
2. Rimel R. W., Giordani B., Barth J. T., Boll T. J., Jane J. A., Disability caused by minor head injury, Neurosurgery, vol. 9, no. 3, pp. 221-228,
Sep. 1981.
3. Thatcher R. W., Walker R. A., Gerson I., Geisler F. H., EEG discriminant analyses of mild head trauma, Electroenc. Clin. Neuro., vol. 73, no.
2, pp. 94-106, Aug. 1989.
4. Thatcher R. W., North D. M., Curtin R. T., Walker R. A., Biver C. J., Gomez J. F., Salazar A. M., An EEG severity index of traumatic brain
injury, J. Neuropsych. Clin. N., vol. 13, no. 1, pp. 77-87, Feb. 2001.
5. Injury prevention and control: Traumatic brain injury - TBI data & statistics, Centers for Disease Control and Prevention, Available:
http://www.cdc.gov/traumaticbraininjury/data/index.html. [Accessed 21 July 2015].
6. Tagliaferri F., Compagnone C., Korsic M., Servadei F., Kraus J., A systematic review of brain injury epidemiology in Europe, Acta Neurochir.,
vol. 148, no. 3, pp. 255-268, Mar. 2006.
7. Gustavsson A., Svensson M., Jacobi F., Allgulander C., Alonso J., Beghi E., Dodel R., Ekman M., Faravelli C., Fratiglioni L., Gannon B., Jones
D. H., Jennum P., Jordanova A., Cost of disorders of the brain in Europe 2010, Eur. Neuropsychopharmacol., vol. 21, no. 10, pp. 718-779, Oct.
2011.
8. Jennett G. T. B., Assessment of coma and impaired consciousness: a practical scale, Lancet, vol. 304, no. 7872, pp. 81-84, Jul. 1974.
9. Pal J., Brown R., Fleiszer D., The value of the Glasgow Coma Scale and Injury Severity Score: predicting outcome in multiple trauma patients
with head injury, J. Trauma., vol. 29, no. 6, pp. 746-748, Jun. 1989.
10. Bigler E. D., Maxwell W. L., Neuroimaging and neuropathology of TBI, NeuroRehabilitation, vol. 28, no. 2, pp. 63-74, 2011.
11. Bardin J. C., Fins J. J., Katz D. I., Hersh J., Heier L. A., Tabelow K., Dyke J. P., Ballon D. J., Schiff N. D., Voss H. U., Dissociations between
behavioural and functional magnetic resonance imaging-based evaluations of cognitive function after brain injury, Brain, vol. 134, no. 3, pp. 769-
782, Feb. 2011.
12. Lee B., Newberg A., Neuroimaging in traumatic brain imaging, NeuroRx, vol. 2, no. 2, pp. 372-383, Apr. 2005.
13. Thatcher R. W., Electroencephalography and mild traumatic brain injury, in Foundations of Sport-Related Brain Injuries, Springer, 2006, pp.
241-265.
14. Thatcher R. W., Cantor D. S., McAlaster R., Geisler F., Krause P., Comprehensive Predictions of Outcome in Closed Head-Injured Patients,.
Ann. NY Acad. of Sci., vol. 620, no. 1, pp. 82-101, Apr. 1991.
15. Winkler I., Haufe S., Tangermann M., Automatic classification of artifactual ICA-components for artifact removal in EEG signals, Behavi.
Brain Func., vol. 7, no. 1, p. 30, Aug. 2011.
16. Anderer P., Roberts S., Schlogl A., Gruber G., Klosch G., Herrmann W., Rappelsberger P., Filz O., Barbanoj M. J., Dorffner G., Saletu B.,
Artifact processing in computerized analysis of sleep EEG - a review, Neuropsychobiology, vol. 40, no. 3, pp. 150-157, Sep. 1999.
17. Fatourechi M., Bashashati A., Ward R. K., Birch G. E., EMG and EOG artefacts in brain computer interface systems: A survey, Clin.
Neurophysiol., vol. 118, no. 3, pp. 480-494, Mar. 2007.
18. McFarland D. J., McCane L. M., David S. V., Wolpaw J. R., Spatial filter selection for EEG-based communication, Electroen. Clin. Neuro.,
vol. 103, no. 3, pp. 386-394, Sep. 1997.
19. Hyvarinen A., Oja E., Independent component analysis: algorithms and applications, Neural networks, vol. 13, no. 4, pp. 411-430, Jun. 2000.
20. Jung T.-P., Makeig S., Humphries C., Lee T.-W., Mckeown M. J., Iragui V., Sejnowski T. J., Removing electroencephalographic artifacts by
blind source separation, Psychophysiology, vol. 37, no. 2, pp. 163-178, Mar. 2000.
21. Viola F. C., Thorne J., Edmonds B., Schneider T. E. T., Debener S., Semi-automatic identification of independent components representing
EEG artifact, Clin. Neurophysiol., vol. 120, no. 5, pp. 868-877, May 2009.
22. Mognon A., Jovicich J., Bruzzone L., Buiatti M., ADJUST: An automatic EEG artifact detector based on the joint use of spatial and temporal
features, Psychophysiology, vol. 48, no. 2, pp. 229-240, Feb. 2011.
23. LeVan P., Urrestarazu E., Gotman J., A system for automatic artifact removal in ictal scalp EEG based on independent component analysis and
Bayesian classification, Clin. Neurophysiol., vol. 117, no. 4, pp. 912-927, Apr. 2006.
24. Halder S., Bensch M., Mellinger J., Bogdan M., Kubler A., Birbaumer N., Rosenstiel W., Online artifact removal for brain-computer interfaces
using support vector machines and blind source separation, Comput. Intell. Neurosci., Apr. 2007.
25. Mammone N., La Foresta F., Morabito F. C., Automatic artifact rejection from multichannel scalp EEG by wavelet ICA, IEEE Sensors Journal,
vol. 12, no. 3, pp. 533-542, Feb. 2012.
26. Castellanos N. P., Makarov V. A., Recovering EEG brain signals: artifact suppression with wavelet enhanced independent component analysis,
J. Neurosci. Meth., vol. 158, no. 2, pp. 300-312, Dec. 2006.
27. Duffy F. H., Hughes J. R., Miranda F., Bernad P., Cook P., Status of quantitative EEG (QEEG) in clinical practice, Clin. EEG Neurosci., vol.
25, no. 4, pp. 6-22, Oct. 1994.
28. Hughes J. R., John E. R., Conventional and quantitative electroencephalography in psychiatry, J. Neuropsych. Clin. N., vol. 11, no. 2, pp. 190-
208, May 1999.
29. European Data Format, [Online]. Available: http://www.edfplus.info/. [Accessed 21 July. 2015].
30. Gavert H., Hurri J., Sarela J., Hyvarinen A., Fastica 2.5, [Online]. Available: http://research.ics.aalto.fi/ica/fastica/. [Accessed 12 Jan. 2015].
31. Neuroguide help manual, Applied Neuroscience, [Online]. Available: http://www. appliedneuroscience.com/Tutorials.htm. [Accessed 12 Jan.
2015].
32. Brainmaster Technologies, Inc., [Online]. Available: http://www.brainmaster.com/. [Accessed 21 July 2015].
33. International statistical classification of diseases and related health problems 10th revision, World Health Organization, [Online]. Available:
http://apps.who.int/classifications/icd10/browse/2015/en. [Accessed Jan. 2015]