biocybernetics and biomedical engineering 40 (2020) 649–690

Available online at www.sciencedirect.com

ScienceDirect

journal homepage: www.elsevier.com/locate/bbe

Review Article

A comparative analysis of signal processing and

classification methods for different applications
based on EEG signals

Ashima Khosla *, Padmavati Khandnor, Trilok Chand

Computer Science and Engineering Department, Punjab Engineering College (Deemed to be University),
Chandigarh, India

article info abstract

Article history: Electroencephalogram (EEG) measures the neuronal activities in the form of electric currents
Received 6 August 2019 that are generated due to the synchronized activity by a group of specialized pyramidal cells
Received in revised form inside the brain. The study presents a brief comparison of various functional neuroimaging
30 January 2020 techniques, revealing the excellent neuroimaging capabilities of EEG signals such as high
Accepted 1 February 2020 temporal resolution, inexpensiveness, portability, and non-invasiveness as compared to the
Available online 21 February 2020 other techniques such as positron emission tomography, magnetoencephalogram, func-
tional magnetic resonance imaging, and transcranial magnetic stimulation. Different types
Keywords: of frequency bands associated with the brain signals are also being summarized. The main
EEG applications purpose of this literature survey is to cover the maximum possible applications of EEG
Signal processing signals based on computer-aided technologies, ranging from the diagnosis of various
Classification models neurological disorders such as epilepsy, major depressive disorder, alcohol use disorder,
EEG data collection and dementia to the monitoring of other applications such as motor imagery, identity
Brain rhythms authentication, emotion recognition, sleep stage classification, eye state detection, and
Functional neuroimaging drowsiness monitoring. After reviewing them, the comparative analysis of the publicly
techniques available EEG datasets and other local data acquisition methods, preprocessing techniques,
feature extraction methods, and the result analysis through the classification models and
statistical tests has been presented. Then the research gaps and future directions in the
present studies have been summarized with the aim to inspire the readers to explore more
opportunities on the current topic. Finally, the survey has been completed with the brief
description about the studies exploring the fusion of brain signals from multiple modalities.
© 2020 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish
Academy of Sciences. Published by Elsevier B.V. All rights reserved.

* Corresponding author at: Computer Science and Engineering Department, Punjab Engineering College (Deemed to be University), Sector
12, Chandigarh, India.
E-mail addresses: [email protected] (A. Khosla), [email protected] (P. Khandnor), [email protected] (T. Chand).
https://doi.org/10.1016/j.bbe.2020.02.002
0208-5216/© 2020 Nalecz Institute of Biocybernetics and Biomedical Engineering of the Polish Academy of Sciences. Published by Elsevier
B.V. All rights reserved.
650 biocybernetics and biomedical engineering 40 (2020) 649–690

stage [8] in the humans or less power in alpha and theta bands
1. Introduction
[9] in all the regions (central, occipital, frontal, parietal, and
temporal) are observed for the depressed patients when
EEG transforms Electroencephalogram means it is the contin- compared to the normal subjects.
uous recording of the electrical activity of the brain by placing EEG signals are non-linear and non-stationary in nature.
the metal electrodes over the scalp. The neuronal cells The small variations in the voltage fluctuations of the EEG
spontaneously communicate with each other via generating measurements conclude the happening of some neuronal
the electrical currents and remain active all the time even if a activity. So the visual inspection of these signals varies with
person is sleeping or relaxing. The low cost, high flexibility, the expertise experience. Moreover, the long EEG recordings
high temporal resolution, non-invasiveness, ease of use, require a lot of time for their manual review and sometimes
portability and safe nature make EEG a powerful tool for the the results may be inaccurate due to the presence of artifacts
brain imaging task as compared to the other functional in the signals. So the processing and the analysis of these
neuroimaging techniques such as positron emission tomog- signals can be done with the help of computer-aided
raphy (PET), magnetoencephalogram (MEG), functional mag- technologies in order to get fast and accurate results. The
netic resonance imaging (fMRI), and transcranial magnetic use of computer-aided technologies with EEG signals has
stimulation (TMS). gained a widespread popularity, especially in the diagnosis of
Hans Berger (1873–1941), the German Physician [1] coined various neurological and neuropsychiatric disorders such as
the term 'electroencephalogram' to define the electrical epilepsy [10,11], major depressive disorder (MDD) [12,13],
potentials occurring in the human brain. He took the first alcohol use disorder (AUD) [14,15], and dementia [16,17] such
EEG readings from the patients with ‘‘palliative trepanations’’ as Alzheimer, mild cognitive impairment (MCI), Parkinson,
or having some defects in the skull by making use of metal and dementia with Lewy bodies (DLB). EEG based application
strips and galvanometer [2]. He was able to characterize the a of Motor Imagery has opened a new gateway in the field of
(‘‘slower and larger’’) and b (‘‘faster and smaller’’) rhythms and neuroprosthesis [18,19]. Apart from this, other research
concentrated on the changes occurring in the EEG patterns domains such as identity authentication [20,21], sleep stage
while associating with the mental attention and cerebral classification [22], emotion recognition [23,24], eye state
injuries. His findings opened the gateway for most of the detection [25,26], and drowsiness monitoring [27] are getting
applications that are presently based on the EEG signals as prominent results with the use of physiological data such as
he observed that these brain signals were not irregular and EEG.
inconsistent, instead electrical changes revealed some peri- The EEG signal processing and analysis is basically
odic patterns that help to deduce the occurrence of some performed in four steps: preprocessing the raw signals with
activity, e.g. changing the state from sleep (slow, very low the help of filtering or some other techniques, then extracting
frequency, and high amplitude waves) to wakefulness (fast, the most important information in the form of features from
high frequency, and low amplitude waves). He was also them, further applying the feature selection methods for more
responsible for investigating the effects of anesthesia and optimized results, and finally at the result analysis phase, the
epilepsy on the EEG recordings [1]. disease diagnosis or the recognition of the different functional
EEG can record only those potential changes that occur states of the brain is made through the machine learning
due to the synaptic transmissions. When the action potential models or the statistical tests. So the main aim of this research
reaches at the axon terminal, neurotransmitters are releas- study is to target the maximum number of EEG based research
ed at the synapses site [3], causing the exhibitory post synap- applications available in the literature and to perform the
tic graded potentials (EPSPs) or the inhibitory post synaptic comparative analysis for:
graded potentials (IPSPs) [4,5] to occur at the postsynaptic cells.
These potentials lead to the flow of ionic currents in the (a) Different data collection methods in the form of publicly
extracellular space of the cell membrane, thus generating available EEG datasets and other local data acquisition
the local field potentials with very small magnitudes. Research studies.
studies conclude that the detection of electric fields is possible (b) Various pre-processing methods such as down-sampling,
only due to the synchronized activity by a group of few artifact handling, and feature scaling.
specialized pyramidal neurons that are present in the cortical (c) Different categories of features using various feature
regions of the brain [6,7]. The currents generated inside these extraction methods.
cells do not cancel out the effect of each other due to their (d) Post-processing methods such as feature selection and
unique and stable orientation. The overall electric field dimensionality reduction techniques.
becomes much stronger when the postsynaptic graded (e) Result analysis methods in the form of classification
potentials for such group of neurons are summed up. This algorithms and statistical tests.
summation actually helps in the measurement of EEG signals.
Different types of rhythms such as theta, delta, alpha, beta, The rest of the paper is organized as follows: the taxonomy
and gamma can be observed in the brain waves depending of the proposed study is explained in Section 2, the brief
upon the different functional states of the brain. Any subtle comparison of the various functional neuroimaging techni-
changes in the frequency patterns of these waves help in ques is provided in Section 3, different types of brain rhythms
the diagnosis of certain neurological disorders or to conclude in Section 4, data collection process in Section 5, EEG signal
the occurrence of some neuronal activity in response to some processing and analysis in Section 6, then the other EEG based
external stimuli. The slow brain waves indicate the deep sleep research studies with their findings are explained in Section 7,
 biocybernetics and biomedical engineering 40 (2020) 649–690 651

Section 8 summarizes the research gaps and the future rized on the basis of applications for which the EEG signals are
directions in the given studies, then a brief description about used – 14 are for epilepsy and seizures, 10 for depression, 9 for
the multi-modal fusion of brain signals has been given in MI, 8 for emotion state recognition, 8 for eye state recognition,
Section 9, and finally the paper has been concluded with its 6 for sleep stage classification, 5 for alcoholism, 5 for dementia,
future scope in Section 10. The list of acronyms and 3 for driver drowsiness, 2 for identity authentication, 1 for
abbreviations as summarized in Table 21 are provided in multi-class task recognition, and 4 for others (visual comfort
Appendix A. level of images, BCI based spelling interface, autism, and one
study based on reduced number of electrodes using the
concept of ERP), the total actually makes 75 because two
2. Taxonomy of the proposed study
studies are based on multiple applications such as one study
based on epilepsy with alcoholism and another on epilepsy
The main aim of the present study is to inspire the readers with autism, so are counted twice. Also all the application
with the excellent neuroimaging capabilities of the EEG signals studies are classified on the basis of source of datasets used -
and how the computer-aided technologies are used to monitor 38 are based upon using the publicly available datasets and 35
the numerous applications by using a variety of signal pro- are acquiring data locally, out of total, 2 are based upon the
cessing and classification algorithms. The idea is to explode both. So, this makes a total of 71 studies. Out of 73, the
the market with the automated machine learning recognition information about the datasets is not given in the two studies.
systems based on brain signals. This analysis is completed in After reviewing all the studies, firstly a brief comparison of
the following manner (as shown in Fig. 1): A total of 131 various functional neuroimaging techniques has been given,
research papers have been surveyed to complete the current then the description about the different brain rhythms has
study. Except four, all the studies are from the years 1999 to been summarized briefly, and finally, the available EEG
2019. 3 handbooks, 103 journal papers, 9 book chapters, 13 application studies based on computer-aided technologies
conference papers, 1 Ph.D. thesis, 1 web-link, and 1 tutorial file are analysed from signal processing and classification per-
constitute the total number of studies. Apart from that, 15 spective, with the target to extract the maximum information
web-links are included in the study for accessing the publicly in the form of four stages- preprocessing techniques, catego-
available datasets for various applications. Further categoriz- ries of features, post-processing methods, and finally the
ing the studies on the basis of the explained topics – there are result analysis using classification algorithms and statistical
12 research studies for the comparison of various functional tests. For preprocessing stage, 44 studies are mentioned in
neuroimaging techniques, 90 studies for the EEG applications Table 5 for artifact handling and 7 have been explained in
based on computer-aided technologies, 16 studies for brain separate section of preprocessing details for public datasets for
rhythms (out of 16, 8 are already included in the applica- artifact handling, 8 for downsampling (out of 8, 3 are based
tion studies), 5 studies based upon multi-modal fusion of upon preprocessing for public datasets), and 8 in Table 6 for
brain signals, and the remaining 16 are based on the basic feature scaling. For feature extraction, 28 studies are based
knowledge related to the EEG signals such as history, on statistical features (Table 8), 19 for spectral (Table 9), 21
generation of electric currents inside the brain, types of for non-linear (Table 10), and 7 for functional-connectivity
artifacts, comparison of feature extraction methods, and non- based features (Table 11). For postprocessing, 23 studies
linear analysis studies for EEG signals. The main purpose of are mentioned in Table 12 using different types of feature
the present literature review is to target the maximum number selection techniques and 12 for dimensionality reduction in
of EEG applications and explore a variety of signal processing Table 13. For result analysis, 51 are based on the classification
and classification algorithms used in those studies. Out of 90 techniques (Table 16) and 14 on statistical tests (Table 17).
application studies, 17 give the description about the public Apart from that, there are about 13 studies that have been
datasets that are available online and 73 are further catego- chosen separately and explained in Table 18 with their

Fig. 1 – Taxonomy of the proposed study.

652 biocybernetics and biomedical engineering 40 (2020) 649–690

important findings. These number of studies are also depicted Comparison of functional neuroimaging techniques
in Fig. 3. There are 5 studies focusing on the multi-modal
fusion of brain signals, that have been summarized in Table 20. (a) Measuring the neuronal activity indirectly/directly
The idea is to make the readers aware that what level of PET involves the generation of cross-sectional 2D and
research has already been done on EEG signals, from signal 3D images of the brain giving the measurement of
processing and classification perspective and how more radiations emitted by the ‘‘radiotracers’’ that are injected
number of opportunities can be explored on this topic. into the blood stream [31,32]. The use of tracers helps to
map the blood flow differences in order to visually
represent the pathological conditions and functions of
3. Comparison of functional neuroimaging
the brain. fMRI aims at tracking the amount of blood
techniques
oxygen levels in the brain [35] and gives information about
the active brain regions during certain activities. It is
Functional neuroimaging is a brain imaging method that performed by using a signal method called as BOLD (Blood
encompasses a wide variety of technologies that are directly or Oxygenation Level Depredent signal) [33,29,36,32,30], i.e.
indirectly used to investigate the functional information of the the variations occurring in the intensities of the signal as a
central nervous system. It is used to represent the metabolic result of changing oxygen levels in the blood due to the
and physiological processes occurring inside the brain. It occurrence of some neuronal activity. In order to under-
includes various techniques: positron emission tomography stand the complex brain functionality, the subjects are
(PET), magnetoencephalogram (MEG), functional magnetic made to perform certain tasks in response to stimuli
resonance imaging (fMRI), electroencephalogram (EEG), and (visual, auditory and so on) while present in the scanner.
transcranial magnetic stimulation (TMS) [28–30]. These tech- Thus, both fMRI and PET give an indirect measurement of
niques are used to detect the various biomarkers that are the brain's electrical activities. MEG [36,30] is used to
helpful in the diagnosis of various neuropsychiatric and directly measure the brain functionality by recording the
neurological disorders [31–34,29] such as Alzeihmer, Parkin- magnetic fields with the help of highly sensitive magnet-
son, alcoholism, schizophrenia, Huntington's disease, Tour- ometers called as SQUIDs (superconducting quantum
ette syndrome, stress and mood disorders, etc. This is done by interference devices). EEG directly records and interprets
identifying the dysfunctioning in the release of certain the electrical activity of the brain by using the metal
substances like serotonin, dopamine, etc. in the brain. The electrodes placed over the scalp. The electric currents are
comparison of various functional neuroimaging techniques generated by the synchronization of millions of active
has been graphically represented through Fig. 2. neurons. Then, TMS is also based upon the direct

Fig. 2 – Comparison of functional neuroimaging techniques.

 biocybernetics and biomedical engineering 40 (2020) 649–690 653

measurement of the neuronal activity by stimulating the TMS is having its own drawbacks such as, it is not suitable for
nerves cells of the brain with the electric simulator. The studying the functions in the temporal lobe because the
functionality of the specific brain areas can be tested temporalis muscle is contracted painfully with the TMS
through the process of ‘‘virtual lesions’’ in which the stimulation [37] and secondly, headaches or seizures may
disruptions to the different brain areas is made temporari- occur if the safety measures are not followed while undergoing
ly and in a reversible manner [37]. TMS [30]. Thus, the low cost, high flexibility, high temporal
(b) Based on invasive/non-invasive resolution, non-invasiveness, ease of use, portability, and high
In PET, it is required to inject a radioactive and positron density recording capability make EEG a powerful tool for the
emitting contrast elements into the subject's body, thus brain imaging task, especially for studying the dynamically
making it an invasive methodology while all other occurring complex processes of the brain.
techniques are based upon non-invasive methodology.
(c) Based on spatial/temporal resolution
4. Brain rhythms
fMRI has high temporal resolution than that of PET but
less than that of EEG, MEG, and TMS [38,30,37]. fMRI and
PET can never match the temporal resolution [36,38] of Due to the extremely complex patterns generated by the firing
electrophysiological signals because it involves the indi- of billions of neurons, the signals consist of a mixture of
rect measurement of the brain activity through the various base frequencies. The researchers have classified
hemodynamic changes. The temporal resolutions for the those varied frequency ranges into some sub-groups, known
methods PET, fMRI, MEG, EEG, and TMS are of the order of as the frequency bands. Each of these frequency bands
>10 s [36,30,37], >1 s [36,30], <1 ms [36,30,37], <1 ms represents a different cognitive or attentional state of a brain.
[36,30,37], and <1/2 s [37], respectively. The history of these waves [5,8] can be briefly traced as – in
fMRI has the highest spatial resolution [30,35] (order of 1929, Berger introduced the term ‘‘alpha’’ and ‘‘beta’’ [1]. In
millimetres) than that of PET, MEG, EEG, and TMS, thus 1934, Adrian and Matthews gave a significant contribution that
helps to do the good quality anatomy analysis of images. the alpha rhythms of 10–12 Hz are dominant in the human
MEG has high spatial resolution than EEG [30]. The spatial brain when a person is in the resting state with his eyes closed
resolutions for the methods PET, fMRI, MEG, EEG, and TMS and are mainly identified from the occipital region of the
are of the order of 5–20 mm [37], 1–5 mm [36,37], >5 mm cerebral cortex [41]. In 1938, the term ‘‘gamma’’ was intro-
[37], >10 mm [37], and 5–10 mm [37], respectively. duced by Jasper and Andrews with the frequencies of greater
MEG and EEG offer the highest temporal resolution than than 30–35 Hz. In 1936, Walter introduced the term ‘‘delta’’ for
other techniques that is highly demanded in measuring all those frequencies which lie below the alpha band. He also
the complex functional events occurring dynamically in defined the frequency range of 4–7.5 Hz for ‘‘theta’’ waves. The
the brain. idea of theta was clearly introduced in 1944 by Walter and
(d) Cost Dovey [42]. The fuzzy upper and lower frequency limits are
Except EEG and TMS, the other functional imaging available for these frequency bands in the literature. So,
techniques are very expensive [36,30]. Like for MEG, the various brain rhythms with their frequency ranges [8], regions
strength of magnetic field inside the brain is very weak, so of occurrence and characteristics [5,8,43,44] have been
it is important to have magnetically shielded devices and generalized in Table 1.
rooms in which those are measured thus, making the MEG The other research studies have used some other variants
devices highly expensive (in millions of dollars) whereas of the above brain rhythms depending upon the requirement
EEG and TMS are comparatively inexpensive (in thou- of certain frequency bands for the analysis of particular
sands). application. These are summarized in Table 2.
(e) Portability
EEG and TMS are portable techniques while PET, fMRI,
and MEG have to be performed in some confined place [30]. 5. Data collection studies

Unlike PET and fMRI, EEG directly measures the electrical The different research applications based on EEG signals have
activity of the brain and does not require the subject to be either used the publicly available datasets or have themselves
exposed to any magnetic fields or injected with any radio- created their own datasets but have not released them
tracers, thus making the EEG tests to be very safe. The excellent publicly. So, the data collection process can be divided as:
temporal resolution and the reasonable spatial resolution make publicly available datasets and other locally collected data
EEG a perfect candidate for recording the complex neuronal acquisition methods. These are shown in Fig. 3.
activities occurring dynamically on a temporal scale of
milliseconds [39]. Although MEG has comparable temporal 5.1. Publicly available EEG datasets
resolution as that of EEG but it has its own limitations. MEG is
insensitive to some of the sources that are radially oriented Table 3 gives the description about the publicly available
while EEG has higher sensitivity than MEG so it is equally able to datasets including – the name with which that dataset is
detect all the source orientations very well. EEG has the available online, their web links, number of classes in which
capability to deeply analyse the sources while MEG detects the the EEG signals have been distinguished, number of subjects
sources superficially. EEG is less expensive and portable in involved in the experiments, devices used for collecting the
nature as compared to the other techniques except TMS. But EEG signals, and the sampling rate with which the data has
654 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 1 – Brain rhythms.

Rhythm Frequency Regions Characteristics
range (in Hz)
Gamma (g) >30 Fronto-central areas [5]  Sometimes known as fast beta waves.
 Rarely occur in the brain area, so can be used as an indicator to diagnose
certain neurological disorders.
 Having the highest frequencies with the lowest amplitudes.
 Functional integration of activities occurring transiently.
 Help to get the brain locations responsible for voluntary movements
(such as movement of left and right index finger, right toes, etc.)
Beta (b) 14–30  Mostly found in normal adults.
Parietal, somatosensory,  Increased attention and alertness.
frontal, and motor areas [43]
 Increasing levels of beta may be obtained in
panic condition.
 May be increased due to the bone defects or in the regions
with tumours.
 Amplitudes generally less than 30 mV [5,8].
Alpha (a) 8–13 Occipital and parietal regions [44]  Relaxed with closed eyes and wakefulness state.
 Greatly affected during the menstrual cycle
 Attenuation in alpha frequencies can be used for assessing the anxiety
and emotional tension of the subjects.
 Generally have an amplitude of less than 50mV [5].
Theta (u) 4–7.5 Hippocampus region [8]  Abnormal in adults, common for young children below 13 years.
 Sub-conscious activity of the brain.
 Deeply relaxed and meditated mind state.
 An increase in the theta band power under testing condition depicts the
memory demands [43].
 Arise of these waves in normal subjects shows the change from
conscious to the drowsy state.
Delta (d) 0.1–3.5 Mostly in thalamus region [8]  Predominate in new born babies, rarely present in adults with normal
state during waking condition.
 Observed at third sleep stage known as slow wave sleep (SWS).
 Artifacts generated from the muscles of the jaw and the neck are
generally confused with the delta rhythms.
 ‘‘Slow’’ waves having the ‘‘highest’’ amplitudes (75–200 mV) [43].

Table 2 – Other brain rhythms.

Purpose Sub-bands with frequency limits (in Hz) References
Sleep stage classification  Sigma waves (12–16) and k-complex waves [45]
 b1 (14–22) and b2 (22–31) [46]
Emotion recognition  Low alpha (8–10) [47]
Depression or stress recognition  Low beta (13–15), beta (15–20) and high beta (20–38) [48]
 Theta (Fc 6) to (Fc 2) and beta (Fc + 2) to (Fc + 26) where Fc is the [49]
central maximum frequency in alpha band
Alcoholism  Beta (12–25) and high beta (25–30), gamma (30–40) and high gamma [50]
(40–50)
Dementia (Alzeihmer, MCI and DLB)  Beta 1 (13–19) and beta 2 (19–30) [51]
 Alpha 1 (8–10) and alpha 2 (10–13) [52]

been stored. The last column gives the references of different have not released them publicly. The description includes-
research studies that have used these datasets for different purpose of EEG data collection, details of subjects involved in
purposes such as alcoholism, eye state detection, epilepsy and the experiments, devices used, number of EEG channels used
seizure detection, motor imagery, emotion recognition, sleep for data acquisition, sampling rate/recording time/electrode
stage classification, identity authentication, multi-task recog- impedance (EI), state in which the data has been collected (can
nition, and drowsiness detection. be relaxed or task based), references of the research studies that
have carried that process. The reviewed studies for these locally
5.2. Local data acquisition studies collected datasets include the applications such as dementia,
depression, motor imagery, emotion recognition, identity
Table 4 gives the description about the various data acquisition authentication, alcoholism, eyes state detection, driver drows-
studies that have themselves created their own datasets and iness/fatigue detection, and multi-class task recognition.
 biocybernetics and biomedical engineering 40 (2020) 649–690 655

Fig. 3 – EEG data collection, signal processing, and result analysis.

6.1. Pre-processing
6. EEG signal processing and analysis
The pre-processing methods applied in the research studies
The dynamically changing functional states of the brain are can be broadly divided into three categories- downsampling,
easily captured in the small variations of the EEG readings. artifact handling and feature scaling, as explained below:
Also, the EEG of the normal person differs from the abnormal
one. So it is very important to identify those changes by (a) Downsampling
applying a series of signal processing and analysis mecha- The data collected by the different EEG devices is
nisms with the help of computer-aided technologies. This is downsampled to certain ranges depending upon the
carried out in four stages – pre-processing, feature extraction, requirement of an application. Downsampling to 16 Hz
post-processing and result analysis. These are shown in Fig. 3. [105], 64 Hz [21], from 512 to 64 Hz [18], 256 to 16 Hz [104],
The raw signals can be directly fed to the result analysis phase and 2400 to 600 Hz [20] is done in the research studies
for classification or statistical analysis, i.e. any of the first three depending upon the requirement of an application.
stages can be skipped in the processing mechanism depending (b) Artifact handling
upon the requirement of the application. Like in some of the Artifacts are not generated from the cortical activity but
research studies where the deep neural architectures are used due to the errors originated from the experimental
for the data analysis, any feature extraction or feature settings, environment noise or biological artifacts. The
selection is not required to be performed manually. Also in artifact signals can be broadly categorized into two classes
other cases, if the data is already pre-processed or the features depending upon their origin [40,115] [116]: (i) Technical or
are already extracted, especially for the publicly available extrinsic artifacts: These arise due to the technical issues or
datasets, then any of those steps can also be skipped. Again if some external factors in the data collecting environment
the extracted feature set is small, then there is no need for such as due to the misplacement of electrodes, powerline
post-processing that involves the feature selection or di- interference (50–60 Hz) or the electromagnetic interference
mensionality reduction methods. Finally, at the result analysis in the cables or other devices, etc., and (ii) Physiological or
phase, the classification or the statistical analysis is made with intrinsic artifacts: Different types of physiological signals
the aim of diagnosing some abnormality or recognizing the generated inside the human body act as the artifacts
different functional states of the brain to monitor some during the EEG data collection process.
application. The values inside the parenthesis in Fig. 3 gives These include eye movements or blinks (or electro-
the information about the number of application studies that oculogram (EOG) artifacts), muscle activities (or electro-
have worked upon the corresponding techniques. myogram (EMG) artifacts), cardiac activities (or
 656
Table 3 – Public datasets.
Dataset Web link No. of classes Subjects Device Sampling References
rate (in Hz)
UCI dataset for Alcohol [53] http://kdd.ics.uci.edu/databases/eeg/ 2 (Control and alcohol) 122 61 electrode cap (ECI, Electrocap 256 [14,54,55]
eeg.data.html International) + 3 reference channels
+ horizontal and vertical EOG
UCI dataset for Eye State [56] https://archive.ics.uci.edu/ml/datasets/ 2 (Eyes open and close) 1 Neuroheadset EmotivEpoc 14 128 [57–59,26]
EEG+Eye+State channels
Bonn university database http://www.meb.unibonn.de/ 3 (Control, interictal and ictal/ 10 128-Channel amplifier system 173.61 [55,61–68,10,11]

biocybernetics and biomedical engineering 40 (2020) 649–690

[60] epileptologie/science/physik/eegdata. epileptic)
html
Bern Barcelona Database [69] https://www.upf.edu/web/mdm-dtic/ 2 (intercranial signals with 5 Intracranial strip along with depth 512 [70]
datasets categories- ‘‘focal’’ and ‘‘non- electrodes manufactured by AD-TECH
focal’’) (Racine, WI, USA)
Seizure-Prediction data [71] https://www.kaggle.com/c/ 2 (pre-ictal and inter-ictal) 5 dogs and 16 sub-dural conductors inserted into 400 (for dogs), [72]
seizure-prediction/data 2 patients canine's head (for dogs) Patient 1 = 8 5000 Hz
depth electrodes, patient 2 = 3 * 8 sub- (2 patients)
dural electrodes
CHB-MIT Scalp EEG Database https://archive.physionet.org/pn6/ 2 (seizure and non-seizure) 23 An array of 23-EEG channels 256 [67,74]
[73] chbmit/
Autism dataset by KAU, https://malhaddad.kau.edu.sa/ 2 (normal and autistic) 19 16 Ag/Agcl electrodes, g.tec EEG cap, g. 256 [67]
Saudi Arabia [75] Pages-BCI-Datasets-En.aspx tec GAMMAbox, g.tec USBamp and
BCI2000
BCI competition II Dataset-III http://www.bbci.de/competition/ii/ 2 (Left or Right hand imagery 1 G.tec amplifier and a set of Ag/Agcl 128 [77,78]
[76] movement) electrodes
DEAP database [79] http://www.eecs.qmul.ac.uk/mmv/ Labels- valence, arousal, 40 Biosemi ActiveTwo with 32 EEG 512 [80,47,81,21]
datasets/deap/ dominance and liking channels
BCI Competition III dataset http://www.bbci.de/competition/iii/ 2 (Right foot and right hand) 5 BrainAmp amplifier system with 128 100 [83,84]
IV-a [82] channel electrode cap
Physionet EEG Motor https://physionet.org/pn4/eegmmidb/ Contains the two baseline tasks 109 BCI 2000 system consisting of 64 160 [87,25,88]
Movement/Imagery [85] and http://www.bci2000.org and four motor imagery tasks channels
[86]
Sleep EDF https://physionet.org/physiobank/ Sleep stages according to rules 8 PSGs Four channel cassette recorder and 100 [45,22,46,89]
database/sleep-edf/ by Rechtschaffen and Kales (sleep edf) telemetric system and contains two EEG
(Wake, REM, Stage 1, Stage 2, channels (FpzCz and PzOz), horizontal
Stage 3, and Stage 4) EOG data and EMG signals
Sleep EDF Database https://physionet.org/physiobank/ 197 PSGs
[Expanded] [90,86] [91] [92] database/sleep-edfx/ (expanded)
MIT-BIH Polysomnographic https://physionet.org/physiobank/ Sleep stages defined according 16 3 channels – C3-O1, C4-A1 and O2-A1 250 [94,95]
Database [93] [86] database/slpdb/ to rules by Rechtschaffen and
Kales
Sleep Heart Health Study https://sleepdata.org/datasets/shhs/ Sleep stages defined according 5,793 Gold cup electrodes for two EEG 125 [97]
(SHHS)-1 [96] pages to rules by Rechtschaffen and channels (C4/A1) and (C3/A2)
Kales-Wake, N1, N2, N3, N4,
and REM
Table 4 – Other data acquisition studies.
Purpose Subjects Device Channels Sampling rate/ Recording state Refs.
Recording time/EI
Dementia (Alzheimer, HC: 37 (M: 12, F: 25, age = 76), 19-channel system (XLTEK®, Natus 19 200 Hz/5 min Relaxed awake with [51]
MCI, Parkinson, and MMSE = 29, MCI: 37 (M: 16, F: 21, Medical, Plesanton, CA, USA) eyes closed
DLB) age = 76.6), MMSE = 27.25, AD: 37 (M: 12,
F: 25, age = 81.55), MMSE = 21.5, Total:
111
HC: 30 (age = 70–76), MMSE = 28–30, AD: Symtop amplifier 16 + 2 Ref (linked 1024 Hz/30 min/ EC, EO [17]
30 (age = 74–78), MMSE = 12–15 earlobes A1 and A2) EI ≤ 3 KV

biocybernetics and biomedical engineering 40 (2020) 649–690

ADhall+: 36 (M:17, F: 19, age: 69.41 OSG digital equipment (Brainlab and 21 500 Hz/20 min/ Resting state awake [52]
 7.74), MMSE = 19 (n = 33), ADhall-: 108 BrainRT®; OSG B.V.Belgium) EI < 5 KV
(M: 51, F: 57, age: 69.41  7.37),
MMSE = 21 (n = 108), DLBhall+: 29 (M: 20,
F: 9, age: 70.76  9.51), MMSE = 23
(n = 24)
AD: 10 (M: 5, F: 5, age = 69.4  9.2), TrueScan 32 21 200 Hz EC [98]
MMSE = 16.2, HC: 10 (M: 4, F: 6,
age = 68.7  7.7), MMSE = 30
PD: 20 (59.05  5.64, M: 10, F: 10, MMSE: Emotiv Epoc 14 128 Hz/5 min EC [16]
26.90  1.51), HC: 20 (58.10  2.95, F: 11,
M: 9, MMSE: 27.15  1.63)
Depression or stress HC: 18 (M: 18, F: 0, college aged) Emotiv Epoc neuroheadset 14 + 2 Ref 128 Hz Mild stress is induced [23]
in response to stimuli
presentation module
under congruent and
incongruent
conditions
HC: 13 (M: 5, F: 8, age = 38.7  15.8), Neuroscan synamps2 (compumedics, 30 + avg (M1, M2) taken 1000 Hz/2 min (EO) and EO and EC [49]
MDD: 13 (M: 5, F: 8, age = 38.7  15.8) NC, USA) as reference 30 min (EC)/EI < 10 KV
+ Horizontal EOG
+ Vertical EOG
HC: 30 (age = 38.227  15.64), MDD: 33 EEG cap attached to Brain Master 19 (1 Linked ear ref) 256 Hz/5 min each EO EO and EC [9]
(age = 40.33  12.861) systems and EC
7 subjects out of 11 with high cortisol Emotiv Epoc+ 14 128 Hz Under working [48]
levels (M: 11, F: 0, age = 37.9  8.8) condition at onsite and
offsite construction
site
HC: 45 (M: 20, F: 25, age = 33.7  10.2), 19 electrodes placed over the scalp 19 256 Hz/5 min EC [99]
MDD: 45 (M: 22, F: 23, age = 33.5  10.7)
HC: 15 (age = 20–50), MDD: 15 (age = 20– 1 bipolar EEG channels from left and 2 256 Hz/5 min (EO and EO and EC [100,13]
50) right hemisphere each EC each)
HC: 30 (M: 21, F: 9, age = 38.227  15.64), EEG sensor cap with amplifier from 19 (1 Linked ear ref) 256 Hz/5 min each EO EO and EC [12]
MDD: 34 (M: 16, F: 18, age = 40.33 Brain Master system and EC
 12.861)
HC: 204 (M: 68, F: 136, age = 27.46  9.61), MUSE EEG headband 5 (FPz as ref) 220 Hz Task based on [101]
MDD: 144 (M: 58, F: 86, age = 27.65 watching 8 short

657
 9.50) videos
 658
Table 4 (Continued )
Purpose Subjects Device Channels Sampling rate/ Recording state Refs.
Recording time/EI
HC: 23 (M: 23, age = 18–28) Mistar system, MCScap-26 hat for EEG, 5 (FPz as ref) 256 Hz/1 min Recording during EC [102]
30 Ag/Agcl electrodes and EO, before,
immediately after, and
after 20 min of the test

biocybernetics and biomedical engineering 40 (2020) 649–690

(TSST – trier social
stress test)
Motor imagery HC:10 (M: 5, F: 5, age = 26.3  5.4) EEG cap with four g.USBamp amplifiers 60 + 3 EOG, Ref on left 512 Hz/EI < 5 KV Paradigm where the [103]
(g.tec, Graz, Austria) ear and ground on participants have to
right ear look at goal on screen
and do movements
accordingly
HC:3 (M: 1, F: 2 age = 22–24) EEG cap with 5 g.USBamp amplifiers (g. 68 + 3 EOG, Ref on left 512 Hz Based on self-paced [18]
tec, Graz, Austria) mastoid and ground on center out reaching
right mastoid task
HC:9 (M: 5, F: 4, age = 26.1  4.3) EEG cap with 5 g.USBamp amplifiers (g. 68 + 3 EOG, Ref on left 256 Hz Based on imagining [104]
tec, Graz, Austria) mastoid and ground on rhythmic movements
right mastoid of the right arm in
horizontal and vertical
plane in response to
arrow on the screen
HC: 15 EEG cap with g.USBamps (g.tec, Austria) 61 channels – Based on performing [105]
the hand movements
according to the
presented cues on the
screen
Emotion recognition HC: 28 (M: 13, F: 15, avg age = 23.62) MUSE headband 4 220 Hz (raw data), Based on video stimuli [24]
10 Hz (features)
HC: 15 (M: 8, F: 7, age = 29.42  4.02), EasyCap with BrainAmp DC (Brain 16 250 Hz/15 min (3 min Resting state + In [106]
MDD:15 (M: 8, F: 7, age = 30.92  3.65) Products, Munich, Germany) each for different response to audio
condition)/EI < 10 KV stimuli (noise and
music)
HC: 24 Biopac provided electrode cap with 10 3 electrodes (Fz, Cz, Pz) 500 Hz The emotions have [107]
Ag/Agcl electrodes and MP 150 system been evoked by using
images from IAPS
belonging to the four
quadrants (LVHA,
HVHA, HVLA, LVLA).
HC: 10 (M: 8 and F:2, mean age = 30.6 Emotiv Epoc+ 14 128 Hz images from MUG [108]
years and SD = 13.73 years) database from
categories of happy,
angry, and neutral
were shown to the
participants
Table 4 (Continued )
Purpose Subjects Device Channels Sampling rate/ Recording state Refs.
Recording time/EI
Identity authentication HC: 45 (age = 22.4  2.1, User: 15, g.USBamp amplifier with electrodes 16 2400 Hz, EI < 5 KV Based on rapid visual [20]
Imposter: 30) presentation of images
of fare and non-fare
users
Alcoholism HC: 30 (M: 18, F: 12, age = 42.67  15.90), Discovery 24E system 19 256 Hz/5 min for each EO and EC [15]
AUD: 30 (M: 16, F: 14, age = 55.4  12.87) EO and EC
HC: 15 (age = 42.67  15.90), Alcoholics: (1) Discovery 24E (1) 19 + 1 grd + 2 ref + 2 256 Hz and 500 Hz for EO and EC [50]

biocybernetics and biomedical engineering 40 (2020) 649–690

18 (age = 46.80  9.29), Alcoholic for event (1) and (2) resp./5 min
abusers: 12 (age = 56.70  15.33) synchronization each for EO and EC/
EI < 10 KV for (1)
(2) Enobio System (2) 19 + 1 External + 2
mastoid ref
Eyes state detection HC: 2 (M: 1, F: 1) Mindwave headset 1 16 min (two 4 min for EO and EC [109]
each EO and EC)
HC: 20 (M: 8, F: 12, age = 18–30) NeuroScan SynAmps2 30 + 1 electrode A2 + 1 sampled DC to 70 Hz EOand EC [110]
ref (A1) + Vertical EOG and digitized at 1 KHz/
+ Horizontal EOG 2 min each for EO and
EC/EI < 5 KV
Driver drowsiness/ HC: 13 (M: 13, F: 0, age = 22–25) Nihon-Koden EEG 2110 9 + 2 Earlobe channels 256 Hz To make some mental [27]
fatigue detection (A1, A2) + 2 EOG calculation during the
experiment
HC: 8 (age = 20–24) Neuroscan device 30 + 2 Ref + Horizontal 1000 Hz/60 min Simulated driving [111]
and vertical EOG Environment
Multi-class task HC: 5 (M: 3, F: 2, age = 24–30) Emotiv Epoc+ 14 128 Hz Tasks in response to [87]
recognition mark on the screen
Others HC: 5 with normal vision (age = 21.6 BioAmp2 wireless recording system Oz-Cz 250 Hz Subjects were asked to [112]
 5.32) look at the RVS source.
HC: 28 (M: 25, F: 3 age = 19–37 years and Neurosky Mindwave headset 1 EEG channel + eye 300 Hz Data collected under [113]
SD = 4) tracking data the effect of visual
stimuli, images
selected from IVY LAB
stereoscopic 3D image
database.
50 evoked potentials from healthy Mobile device, LiveAmp (by Brain 32 channels and 3 axis 500 Hz Visual, auditory, and [114]
participants (age: 28–53 years, all males) vision) sensor for body somatosensory
movement detection stimulations
HC: 10 64 channel g.Gamma cap with g.tech 60 256 Hz/5 min EO [67]
software

659
 660
Table 5 – Artifact removal methods.
Application Depression Alcoholism Dementia MI/MT Emotion Sleep stage Identity Epilepsy Drowsiness Eyes open/closed Others
Artifact removal recognition authentication
technique

Notch filter (50 Hz [101,12,9,100,13,99]

[15] [51] [103,18,104] [108]
powerline
interference

biocybernetics and biomedical engineering 40 (2020) 649–690

removed)
     [112]
Notch filter (60 Hz [48]    [108]      
powerline
interference
removed)
Bandpass filter 2–36 Hz [101], 0–70 Hz [15] 0.5–60 Hz [98],  0.5–50 Hz [106], 0.3–45 Hz [45] 0.1–55 Hz [21] 0–50 Hz [68] 1–50 Hz [111]  
0.5–70 Hz [12,9], 0.5–30 Hz [17], 1– 4–45 Hz [47], 0.2–
0.5–64 Hz [48], 70 Hz [51] 45 Hz [108]
0.5–46 Hz [49]
Chebyshev filter Transition   8th order, 0.01–   Passband    
width = 1 Hz, 200 Hz [103,18] ripple = 40 Hz,
passband Stopband
ripple = 1 dB, ripple = 49 Hz [20]
Stopband
ripple = 80 dB
[23]
Elliptic band pass    0.5–50 Hz [78]    0.5–60 Hz [67]   0.5–60 Hz [67]
filter
Bandstop filter   49–51 Hz [98]        
Low pass filter   70 or 100 Hz [52]  40 Hz [107]      
High pass filter     0.5 Hz [107]      
Butterworth filter   6th order, 1– 4th order, 1– 3th order, 0.5– 20th order, 1–   2nd order, 0.5– 5th order, 40 Hz 5–80 Hz [112],
49 Hz (bandpass) 70 Hz [103], 4th 15 Hz [108] 60 Hz [22] 8th 60 Hz [94] (low pass) [25], 4th order,
[16] order, 0.3–100 Hz order, 0.5–35 Hz 5th order, 0.5 Hz 0.01–45 Hz
[18], 4th order, [89] (high pass) [26] [113]
0.3–3 Hz [18], 4th
order, 0.3–3 Hz
[105], 8th order,
0.01–100 Hz, 4th
order, 0.2–70 Hz,
2nd order, 0.4–
0.6 Hz [104]
Accelerometer To track and          
module along 3- reject artifacts
axis due to head
movements [101]
Multiple source [12,9] [15]         
modelling
technique
Manual removal by [100,13,99,49]  [52,17,51] [103,87] [106]    [27] [26] 
experts after
visual inspection
or visual
selection of
artifact free data
Table 5 (Continued )
Application Depression Alcoholism Dementia MI/MT Emotion Sleep stage Identity Epilepsy Drowsiness Eyes open/closed Others
Artifact removal recognition authentication
technique
Semi-automatic  Physiological         
detection noise removed
method with
contribution of
both user and
machine,
segments with
SHR score > 0.90
selected [50]

biocybernetics and biomedical engineering 40 (2020) 649–690

PCA    [18]       
ICA [48]  [51] [103,18,105,104]     [111]  
Automatic removal     [106,47]      
Thresholding   Segments with Median absolute Epochs with peak      
technique voltage > 150 mV deviation [18,104] value > 70 mV
[17], Segments were rejected
with [108]
voltage > 80 mV
[16]
Wavelet method      DWT ('db4') with    Threshold 
adaptive denoising using
threshold and heuristic method
soft threshold [57]
function [22]
Revised aligned          [110] 
artifact average
Regression based    [103,104]       
correction
method
Adaptive filters     FIR filter with    Biological noise  
LMS and line
optimization interference
algorithm [117] handled [94]
Wavelet modified           [113]
ICA method

661
662 biocybernetics and biomedical engineering 40 (2020) 649–690

electrocardiogram (ECG) artifacts), etc. All these kinds of v. For BCI Competition II Dataset-III, the readings for the
artifacts are generally handled by applying a variety of electrode channels were filtered with the bandpass filter of
artifact removal/reduction methods. These have been frequency range 0.5–30 Hz.
tabulated in Table 5 – the type of artifact removal technique vi. For BCI Competition III dataset-IVa, the sampling rate was
has been mentioned in the first column, while the rest of 1000 Hz which was downsampled to 100 Hz. The signals
the columns denote the type of application in which the were filtered with the band pass filter of frequency range
corresponding method has been applied for filtering. The 0.05–200 Hz.
symbol () denotes that the particular artifact removal vii. For autism dataset given by King Abdulaziz University,
method has not been used in any of the studied research Saudi Arabia, band pass filters with frequency range 0.1–
papers. 60 Hz and notch filter at 60 Hz stop band frequency are
(c) Feature scaling used for preprocessing the data.
It is very important to scale the features of the dataset in
order to exhibit the symmetrical behaviour. Normalization
is one of the most commonly used feature scaling methods 6.2. Feature extraction
in the studied papers. The scaling is performed either on
the raw/filtered values or on the extracted features. Table 6 A number of feature extraction methods are used to perform
gives the description about the feature scaling methods. time domain, frequency domain, and time-frequency domain
analysis of the signals. Some of them are empirical mode
Pre-processing on publicly available datasets decomposition (EMD), fast Fourier transform (FFT), wavelet
The pre-processing was also performed on some of the transform (WT), wavelet packet decomposition (WPD), and so
publicly available datasets during their data collection process on. The feature extraction methods have been studied in three
before they were publicly released. These involve: categories- spectral estimation methods, family of transforms,
and time decomposition methods (see Fig. 4). A comparative
i. In UCI dataset for alcohol, the trials with amplitudes analysis of different feature extraction methods has been
greater than 73.3 mV, indicating extreme eye and body briefly summarized in Table 7. Different types of band pass
movements were discarded. filters are also used to decompose the signals into various
ii. In Bonn University dataset for epileptic seizure, only those frequency sub-bands, from which then the features are
segments of the collected data were chosen through the extracted for more detailed analysis.
visual inspection that were free from the eye movement Four categories of features can be reviewed from the
and muscle artifacts. Bandpass filters with frequency studied applications that are named as – statistical/wavelet,
range 0.53-40Hz were also applied. spectral, non-linear, and functional connectivity based fea-
iii. In Bern Bercelona database for epilepsy, the signals with a tures. These have been precisely summarized in Fig. 4. These
sampling rate of 1024 Hz were downsampled to 512 Hz and features can be directly extracted from the raw or prepro-
then were filtered with Butterworth bandpass filters of 4th cessed signals. But for more deeper analysis, a number of
order with the frequency range of 0.5–150 Hz. feature extraction methods (as summarized in Table 7) are
iv. For DEAP dataset for emotion recognition, a preprocessed applied to obtain various detailed sub-bands that are used to
form of the original data is also made available online by extract the different types of features from them. A complete
the authors in which the downsampling of data to 128Hz survey for different category of features using different feature
had been performed followed by the removal of EOG extraction methods for various applications has been
artifacts by blink source separation technique. Then the explained as:
band pass filtering with frequency range of 4–45 Hz had (a) Statistical/wavelet features
been applied, followed by some segmentation and Table 8 gives the description about the statistical/
reordering steps. The pre-processed dataset can be wavelet features – the name of features, reference of the
acquired from the same web-link as mentioned in Table 3 studies which have worked upon those features, name of
for the original dataset. the feature extraction methods (if applied), sub-bands

Table 6 – Feature scaling methods.

Normalization method Before/after extracting features References Purpose
Max-Min Before [87] Multi-class task recognition
After [24] Emotion recognition

Z-score (mean = 0, Std Dev = 1) Before [87] Multi-class task recognition

[62] Seizure/epilepsy
[13] Depression
After [9,12] Depression
[15,50] Alcoholism

Unity before [87] Multi-class task recognition

 biocybernetics and biomedical engineering 40 (2020) 649–690 663

Fig. 4 – Types of features extracted with various extraction methods.

from which the corresponding features have been analysis, therefore, no sub-band divisions are performed
extracted after applying the corresponding feature extrac- for them.
tion method (if applied), and the purpose or the application (d) Functional connectivity based features
of the EEG study. In some of the applications, no feature By applying the network theory on the collected EEG
extraction method has been used, instead the raw or the data, the functional networks of the brain are constructed
preprocessed signals have been directly used for the result and the connectivity inside them is measured with the
analysis, therefore, no sub-band divisions are performed help of various features. The functional connectivity based
for them. metrics that are derived from the EEG data are described in
(b) Spectral features Table 11, along with the application study and their
The spectral analysis of the features can be done by references in which they are computed.
using either parametric (Yuler-walker or Burg's Method) or
non-parametric approach (Welch method with FFT)
[118,55]. These features have been explained in Table 9 6.3. Postprocessing
giving the information about the- name of the spectral
parameters, any decomposition method if applied, name Post-processing can be done either through the feature
of the sub-bands from which the respective spectral selection methods or the dimensionality reduction methods.
features can be extracted, and the purpose of the study.
(c) Non-linear features (a) Feature selection
Different categories of non-linear features [125,126] A number of feature selection methods are available in
are available that are important to understand the non- the studied applications that aim to improve the quality of
linear features and complex nature of the EEG signals. the result analysis phase. Table 12 shows the feature
Table 10 gives an explanation about those features as- selection methods along with the application and refer-
category of the non-linear features, name of the features ence of the study in which they have been used.
extracted from the particular category, reference of the (b) Dimensionality Reduction
studies which have worked upon those features, name of EEG application studies using the different dimension-
the feature extraction methods (if applied), sub-bands ality reduction methods have been provided in Table 13.
from which the features have been extracted after
applying the corresponding feature extraction method
(if applied), and the purpose or the application of the EEG 6.4. Result analysis
study. In some of the applications, no feature extraction
method has been used, instead the raw or the prepro- A variety of machine learning algorithms are used for the
cessed signals have been directly used for the result diagnosis of neurological disorders (such as epilepsy/seizure,
664 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 7 – Comparison of feature extraction methods.

Feature Name Description/Advantages Disadvantages/challenges
extraction
method type
Spectral Non-  Estimation of auto-correlation sequence is done from a  Concept of windowing creates the spectral
estimation parametric given set of data and then applying the Fourier leakage effects.
methods or classical transformation to it.
[40,118,55] approach
 Welch method is one of the most popular spectral  Not good for the spectral estimation of short
estimation method: EEG segments.
- The sequences are allowed to overlap and the data  Suffers from a problem of noise sensitivity.
windows are applied.
- Modified periodiograms are generated for each sequence.
- Their average gives an estimation of power spectrum.
Parametric  Model based power spectrum estimation method.  Suffers from a problem of selecting an optimal
or non- model order as:
classical
approach
 Overcomes the problem of spectral leakage thus leading - Higher order induces false peaks and
to better frequency resolution
 Estimation of the parameters of the linear system is - Lower order leads to the smoothing of spectra.
done.
 AR models, linearly stochastic in nature, are considered  Highly vulnerable to heavy biases and larger
as the most preferred ones for estimating the PSD graphs. variability.
 The parameters such as sampling rate, model order, etc.
can be estimated by using two methods:
- Yulk-Walker
- Burg's method
 Best for spectral estimation of shorter data segments.

Family of Fourier  Transforms the raw time domain signals into frequency  Has zero temporal resolution.
transforms Transform domain by using exponential function of varying
[119,120,118] (FT) frequencies as an analysing function for transformation.
 Captures the different frequency components of the  Not good for non-stationary signals i.e. FT is not
signals. a good tool for real time applications where
representation of time varying spectra is
required.
 Appropriate for stationary signals.  Discontinuities in the signals cannot be
represented appropriately.
 Achieves high frequency resolution.
STFT  Windowed version of FT.  Choosing a width of the window is a very
difficult problem as:
 Concept of 'windows' is used, i.e. short time stationary - Shorter time window leads to low frequency
data segments from the non-stationary ones are chosen and high temporal resolution.
and FT is applied to those segments.
 Gives both time and frequency representation of the - Longer time window leads to high frequency
signals. and low temporal resolution.
 Also the window size is fixed, thus the fixed
resolution throughout the time.
CWT  Alternative method of transformation to STFT.  Parameters, a and b, vary continuously over the
whole time during the calculation of the wavelet
coefficients.
 Time-varying window sizes according to the different  Thus, involves a lot of redundancy and effort
spectral components. for analysing and reconstructing the signal, thus
wastage of computational time and resources.
 Time localization of various frequencies can be obtained.
 Good for non-stationary signals.
 Signal is multiplied with the mother wavelet to get the
transformed signal.
 Transformed signal is represented as a function of two
parameters, a and b, called as the ‘‘scale’’ and
‘‘translation’’ respectively.
 Mother wavelet defines the varying finite length window
functions by changing the ‘‘scale’’ and ‘‘translation’’.
Examples of mother wavelet for CWT are Morlet and
Mexican hat.
 biocybernetics and biomedical engineering 40 (2020) 649–690 665

Table 7 (Continued )
Feature Name Description/Advantages Disadvantages/challenges
extraction
method type
DWT  Easier to implement than CWT.  Choosing the correct choice of ‘‘wavelet’’ is a
tedious task.
 Signal is analysed at different scales w.r.t. time.  It suffers from some other limitations such as
sensitive to translation, shift variance, aliasing,
and lack of directionality [121,66].
 Involves the significant amount of information for  The extent of frequency resolution achieved by
analysing and reconstructing the signal, thus useful for DWT is considered to be coarse for analysing the
designing the less intense models in terms of signals in practical scenarios.
computational time and resources.
 A series of low and high pass filters are used to study the
signal at different ‘‘scales’’.
 ‘‘Level wise’’ decomposition is involved in which the
lists of detailed and approximation coefficients with half
the sampling rate are obtained corresponding to the high
and low frequency components respectively.
 Some of the popular wavelet functions are symlet,
daubechies (db1, db2, db4, db6, and db8), and Haar.
 DWT is the most commonly used method than CWT for
EEG based epilepsy diagnosis application.
DD-DWT [66]  Better time-frequency representation method than
traditional DWT.
 Based on the structure of ‘‘dual-wavelets’’. –
 Input signal is passed through DD-DWT to obtain the
low frequency sub-component and two high frequency
sub-components at each level of decomposition.
 The subtle changes in the EEG signals can be revealed
and localized more accurately through this method.
 Outperformed DWT by having properties of anti-aliasing
and shift invariance.
WPD  ‘‘Wavelet Packets’’ are generated, given some wavelet  Lacks of improved directionality and sensitive
[121,122] function with orthogonal property. to location w.r.t. time.
 It is an extension to DWT:  Involves complex data structures.
- In WPD, signal is passed through more number of filters
as compared to DWT.
- In DWT, only the approximation coefficients are
decomposed further at each level.
- In WPD, at each level of decomposition, both the detailed
and approximation coefficients are decomposed into high
and low frequency components, thus offering much richer
signal analysis.
- For n levels of decomposition, WPD produces 2n sets of
wavelet coefficients as compared to (n + 1) sets as in DWT.
TQWT  Efficient in analysing the subtle variations in the  One should have complete understanding to
[123,124] oscillatory signals. finely tune the Q-factor for TQWT.
 Basically involves the 3 parameters – Q (Q-factor), r  Number of levels should be chosen carefully
(redundancy), and j (number of decomposition levels). because as the number of levels become too
large, it becomes difficult to interpret the
resulting coefficients.
 A signal with frequency rate fs, is decomposed into low
and high pass sub-bands.
 Low pass filter and low pass scaling parameter (a) are
responsible for generating low pass sub-band with
sampling frequency afs, similarly, high pass filter and high
pass scaling parameter (b) are responsible for generating
high pass sub-band with sampling frequency, bfs.
 Q describes the extent of signal resonance and is tuned
depending upon the oscillatory nature of the signal.
 High value of Q is suitable for high frequency signals and
low for low frequency signals.
 Can be implemented with FFTs in case of discrete time
signals.
666 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 7 (Continued )
Feature Name Description/Advantages Disadvantages/challenges
extraction
method type

Time domain decomposition EMD [46]

 It is a time- decomposition method that helps to effectively analyse  For EEG signals, where the data is
domain based the non-linear and non-periodic signals. simultaneously captured using the multiple
channels, EMD is not enough to analyse the
cross-channel interdependence [26].
 The signal is decomposed into a set of intrinsic mode
functions (IMFs) that are used as sub-signals.
 Represents the intrinsic modes present inside the
signals w.r.t. time, thus giving better temporal resolution.

Table 8 – Statistical/wavelet features.

Features Reference Feature Extraction method Sub-bands Purpose
(if applied)
Mean [45,46] Raw/ Preprocessed signals No sub-band divisions Sleep stage classification
[48] Raw/ Preprocessed signals No sub-band divisions Depression
[14] FAWT D2, D3, D4, D5 and A5 Alcoholism
[63,10] DWT (Db4, Levels = 5) D3, D4, D5, A5 Seizure/epilepsy
[72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
[68] DWT (symlet, levels = 4) d, u, a, b, g Epilepsy
Median [47] Raw/preprocessed signals No sub-band divisions Emotion recognition
[48] Raw/preprocessed signals No sub-band divisions Depression
[14] FAWT D2, D3, D4, D5 and A5 Alcoholism
[45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
Variance [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[48] Raw/preprocessed signals No sub-band divisions Depression
[95] Band pass filters d, u, a, b, g Sleep stage classification
[72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
Standard deviation [46] FIR bandpass filters d, u, a, b1, b2, k-complex, spindle wave Sleep stage classification
[46] EMD 7 IMFs Sleep stage classification
[47] Raw/preprocessed signals No sub-band divisions Emotion recognition
[48] Raw/preprocessed signals No sub-band divisions Depression
[14] FAWT D2, D3, D4, D5 and A5 Alcoholism
[94] Raw/preprocessed signals No sub-band divisions Drowsiness detection
[25] DWT (db4, levels = 5) d, u, a, b, g Eyes state detection
[61] Multi-basis MODWPT - Seizure/epilepsy
[63,10] DWT (Db4, levels = 5) D3, D4, D5, A5 Epilepsy/seizure
[67] DWT (levels = 6) D3, D4, D5, D6, and A6 Epilepsy and autism
[68] DWT (symlet, levels = 4) d, u, a, b, g Epilepsy
[22] DWT (Db4, levels = 4) d, u, a, b Sleep stage classification
Kurtosis [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[46] FIR bandpass filters d, u, a, b1, b2, k-complex, spindle wave Sleep stage classification
[46] EMD 7 IMFs Sleep stage classification
[47] Raw/preprocessed signals No sub-band divisions Emotion recognition
[48] Raw/preprocessed signals No sub-band divisions Depression
[14] FAWT D2, D3, D4, D5 and A5 Alcoholism
[22] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
 biocybernetics and biomedical engineering 40 (2020) 649–690 667

Table 8 (Continued )
Features Reference Feature Extraction method Sub-bands Purpose
(if applied)
Maximum [45,46] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[94] Raw/preprocessed signals No sub-band divisions Drowsiness detection
[61] Multi-basis MODWPT - Seizure/epilepsy
[22] DWT (Db4, levels = 4) d, u, a, b Sleep stage classification
Minimum [46] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[94] Raw/preprocessed signals No sub-band divisions Drowsiness detection
[61] Multi-basis MODWPT - Seizure/epilepsy
RMS [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[48] Raw/preprocessed signals No sub-band divisions Depression
[23] Bandpass filters u, a, b Depression
[25] DWT (db4, levels = 5) d, u, a, b, g Eyes state detection
Skewness [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[46] FIR bandpass filters d, u, a, b1, b2, k-complex, spindle wave
[14] FAWT D2, D3, D4, D5 and A5 Alcoholism
[22] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
Energy [46] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[45] IIR Butterworth bandpass d, u, a, sigma, b, g
filters: 0.5–45 Hz
[46] EMD 7 IMFs
[46] FIR bandpass filters d, u, a, b1, b2, k-complex, spindle wave
[117] DWT (Db4, levels = 5) d, u, a, b, g Emotion recognition
[78] DWT (Db4, levels = 3) Level 3 Motor imagery
[57] WPD (Db4, levels = 5) A5, D5, D4, D3 Eyes state detection
[61] Multi-basis MODWPT - Seizure/epilepsy
[22] DWT (Db4, levels = 4) d, u, a, b Sleep stage classification
[17] WPD d, u, a, b Dementia
[72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
Average rectified [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
value
Peak-to-peak [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
amplitude [45] IIR Butterworth bandpass d, u, a, sigma, b, g
filters: 0.5–45 Hz
Forward prediction [72] Low pass, high pass, band pass 10 sub-bands – d, u, a, b, low g, full Seizure/epilepsy
error and band stop filters spectrum excluding 57–63 Hz, 63 Hz to
maximum frequency, 100 Hz to
maximum frequency, 200 Hz to possible
complete spectrum, full spectrum
Zero crossings [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
[48] Raw/preprocessed signals No sub-band divisions Depression
[94] DWT (Db2, levels = 5) Scales – 3 (b), 4 (a), 5 (u) Drowsiness detection
Mean square [117] DWT (Db4, levels = 5) d, u, a, b, g Emotion recognition
amplitude
Moving slope [117] DWT (Db4, levels = 5) d, u, a, b, g Emotion recognition
Integrated EEG [94] DWT (Db2, levels = 5) Scales- 3 (b), 4 (a), 5 (u) Drowsiness detection
Hjorth Parameters: [45] Raw/preprocessed signals No sub-band divisions Sleep stage classification
(mobility and [46] FIR bandpass filters d, u, a, b1, b2, k-complex, spindle wave
complexity) [46] EMD 7 IMFs
[47] Raw/preprocessed signals No sub-band divisions Emotion recognition
[22] Raw/preprocessed signals No sub-band divisions Sleep stage classification
Energy ratios [45] IIR Butterworth bandpass d, u, a, sigma, b, g Sleep stage classification
Filters: 0.5-45 Hz
668 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 8 (Continued )
Features Reference Feature Extraction method Sub-bands Purpose
(if applied)
Cumulative max [48] Raw/ Preprocessed signals No sub-band divisions Depression
and min, Smallest
window elements,
moving median,
max to min
difference, Root
sum of squares
peak and other
related
Auto-regressive order = 6 Raw/preprocessed signals No sub-band divisions Identity authentication
coefficients [21]
AAR coefficients order = 6 Raw/preprocessed signals No sub-band divisions Motor imagery
with RLS or 12 [77]
order = 6 Raw/ Preprocessed signals No sub-band divisions
[78]
Wavelet coefficients [80] DWT (Db8) d, u, a, b, g Emotion recognition
[81] DWT (Db8) d, u, a, b, g Emotion recognition
[83] WPD - Motor imagery
[57] WPD A5, D5, D4, D3 Eyes state detection
[64] DWT (db1, db2, db4, db6, Haar) - Seizure/epilepsy
[74] Wavelet filter bank (scales = 6, Upper 5 scales- d, u, a, b, g Seizure/epilepsy
db4)
Energy (delta/alpha) [22] DWT (Db4, levels = 4) d, u, a, b Sleep stage classification
Wavelet entropy [78] DWT (Db4, levels = 3) Level 3 Motor imagery
Maximum and [11] GST and SVD - Seizure/epilepsy
minimal singular
value

Table 9 – Spectral features.

Features Feature extraction method Sub-bands Purpose
Mean frequency Welch method [45] MEAN frequency from the whole power Sleep stage classification
spectrum for 1 s epoch
Band power Welch method + IIR Butterworth d, u, a, sigma, b, g Sleep stage classification
bandpass filters: 0.5-45 Hz [45]
[47] u, low (a), a, b, g Emotion recognition
FFT [24] d, u, a, b, g Emotion recognition
Welch method + Butterworth bandpass d, u, a, b Depression
[99]
[48] d, u, a, low (b), b, high (b), g Depression
[9] Welch method d, u, a, b Depression
FFT [50] d, u, a, b, high (b), g, high (g) Alcoholism
Elliptic bandpass [78] d, u, a, b, g, Motor imagery
DWT (db4, 5 levels) [63,10] D3, D4, D5, A5 Seizure/epilepsy
DWT (levels = 6) [67] D3, D4, D5, D6, and A6 Epilepsy and autism
Normalized band power FFT [24] d, u, a, b, g, Emotion recognition
SASI, APV, RGP [49] Modified u, a, modified b, g, Depression
Peak power and their FFT (Welch method), Peak values from the whole PSD graph Epilepsy, alcoholism
corresponding peak
frequencies
Yuler-walker, Burg's Method [55]
Relative power FFT [50] d, u, a, b, high (b), g, high (g) Alcoholism
[51] d, u, a, b-1, b-2, g Dementia
Median frequency [51] Median frequency of the whole PSD Dementia
graph
[48] Median frequency of the whole PSD Depression
graph
Individual alpha [51] Only a band Dementia
frequency
 biocybernetics and biomedical engineering 40 (2020) 649–690 669

Table 9 (Continued )
Features Feature extraction method Sub-bands Purpose
Lowest, mean and FFT [52] d, u, a1, a2, b Dementia
highest relative power
and peak frequencies
from all sub bands and
in range (4–13 Hz)
respectively, theta/
alpha ratio
Central frequency, Q1F, Burg's method with order = 20 [94] d, u, a, b, g Drowsiness detection
Q3F, ratio H/L, RH/L,
SSD, IR, MaxF, AC,
kurtosis coefficient
Other PSD parameters [21] From the PSD graph, but sub-bands not Identity authentication
mentioned
Max to min power ratio DWT (db4, levels = 5) [117] d, u, a, b, g Emotion state recognition
Power ratios, product of IIR Butterworth bandpass filters: 0.5– d, u, a, sigma, b, g, Sleep stage classification
powers 45 Hz [45]
Valence, arousal [48] From a and b sub-bands Depression
EEG alpha FFT (Welch method) [9] From a band Depression
interhemispheric
asymmetry
DE, DASM, RASM [24] d, u, a, b, g Emotion recognition
Power percentage, FFT [27] d, u, a, b, g Drowsiness detection
gravity frequency,
frequency variability

Table 10 – Non-linear features.

Category Features Reference Feature extraction Sub-bands Purpose
method (if applied)
Entropy based Spectral entropy [45] Welch method From the whole power Sleep stage classification
features spectrum
[51] Not mentioned From the whole power Dementia
spectrum
Renyi Entropy [45] Welch method From the whole power Sleep stage classification
spectrum
[27] DWT (db4) A5, D5, D4, D3, D2 Drowsiness detection
Kraskov entropy [46] FIR band pass filter d, u, a, b1, b2, k- Sleep stage classification
complex, spindle wave
[46] EMD 7 IMFs Sleep stage classification
Shannon entropy [106] Raw/preprocessed No sub-band divisions Emotion recognition
signals
[67] DWT (levels = 6) D3, D4, D5, D6, and A6 Epilepsy and autism
Centered [54] TQWT Last detailed subband Alcoholism
correntropy
Sample entropy [51] Raw/preprocessed No sub-band divisions Dementia
signals
[27] DWT (db4) A5, D5, D4, D3, D2 Drowsiness detection
[22] Raw/preprocessed No sub-band divisions Sleep stage classification
signals
[22] DWT (db4, levels = 4) d, u, a, b Sleep stage classification
MSE [88] Raw/preprocessed No sub-band divisions Eye state recognition
signals
[89] Sleep stage classification
Fuzzy entropy [51] Dementia
[66] DD-DWT (levels = 5) 15 sub-bands – c1, c2, Epilepsy/seizure
c3, c4, c5, d11, d21, d31,
d41, d51, d12, d22, d32,
d42, and d52
[22] Raw/preprocessed No sub-band divisions Sleep stage classification
signals
Permutation [98] Dementia
entropy
[22] Sleep stage classification
670 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 10 (Continued )
Category Features Reference Feature extraction Sub-bands Purpose
method (if applied)
MPE [88] Eye state recognition
Tsallis entropy [22] Sleep stage classification
Singular value [11] GST and SVD Seizure/epilepsy
entropy 1
Singular value
entropy 2
Approximate [27] DWT (db4) A5, D5, D4, D3, D2 Drowsiness detection
entropy
Dispersion entropy [95] Band-pass filter d, u, a, b, g Sleep stage classification
Fractral KFD [46] FIR band pass filter d, u, a, b1, b2, k- Sleep stage classification
dimension complex, spindle wave
[46] EMD 7 IMFs Sleep stage classification
[106] Raw/preprocessed No sub-band divisions Emotion recognition
signals
PFD [46] FIR band pass filter d, u, a, b1, b2, k- Sleep stage classification
complex, spindle wave
[46] EMD 7 IMFs Sleep stage classification
HFD [46] FIR band pass filter d, u, a, b1, b2, k- Sleep stage classification
complex, spindle wave
[46] EMD 7 IMFs Sleep stage classification
[106] Raw/preprocessed No sub-band divisions Emotion recognition
signals
[99] Depression
[49] Depression
FD [47] Emotion recognition
[22] Sleep stage classification
Fractral analysis DFA [46] FIR band pass filter d, u, a, b1, b2, k- Sleep stage classification
complex, spindle wave
[46] EMD 7 IMFs Sleep stage classification
[99] Raw/preprocessed No sub-band divisions Depression
signals
[49] Depression
Other LZC [106] Emotion recognition
complexity
measures
[49] Depression
[51] Dementia
[22] Sleep stage classification
KC [106] Emotion recognition
CD [99] Depression
Large Lyapunov [99] Depression
exponents
[22] Sleep stage classification
Hurst exponent [22] Sleep stage classification
[66] DD-DWT (levels = 5) 15 sub-bands – c1, c2, Epilepsy/seizure
c3, c4, c5, d11, d21, d31,
d41, d51, d12, d22, d32,
d42, and d52
CTM [51] Raw/preprocessed No sub-band divisions Dementia
signals
AMI [51] Dementia
Area parameters [70] EWT d, u, a, b, g Seizure/epilepsy
from RPS plots of
rhythms using
CTM
LL2N [100] Six length BDL 7 wavelet sub bands Depression
TCOWFB
Bispectrum [16] Raw/preprocessed No sub-band divisions Dementia
features using HOS signals
RQA [27] DWT (db4) A5, D5, D4, D3, D2 Drowsiness detection
LLE [67] DWT (levels = 6) D3, D4, D5, D6, and A6 Epilepsy and autism
and db4
 biocybernetics and biomedical engineering 40 (2020) 649–690 671

Table 11 – Functional connectivity based features.

Features EEG application References
Topological features based on PSI such as Eglobal, Elocal, C, Bnodal, and Bedge Dementia [17]
RWECN Fatigue classification [111]
SL Alcoholism [15]
Depression or stress [12,102]
Phase lag index Dementia [52]
Cross-correlation Epilepsy and autism [67]

Table 12 – Feature selection methods.

Technique EEG application References
mRMR Emotion recognition [47]
Sleep stage classification [45]
Genetic algorithm Emotion recognition [47]
Identity authentication [20]
Depression [99]
Fast correlation-based filter solution Dementia [51]
Sleep stage classification [22]
Sequential forward floating selection Identity authentication [21]
Sleep stage classification [22]
Wrapper subset evaluation Emotion recognition [24]
Rank based Selection using ROC Depression [12,9]
Alcoholism [15]
Student t-test Depression [100]
Alcoholism [50]
Parkinson [16]
Sleep stage classification [95]
Correlation based method Depression [48]
Wrapper method Depression [48]
Kruskal Wallis test Epilepsy/seizure [70]
Sleep stage classification [46]
FDM, Kullback–Leiber, Bhattacharya distance, Motor [77]
Gini Index, DM imagery
LDA Drowsiness detection [94]
ANOVA test Eyes state detection [109]
Epilepsy [66]
CfscSubset Eval Evaluator (weka toolkit) Eyes state detection [57]
Fisher Score Sleep stage classification [22]
Sequential forward selection Sleep stage classification [22]
RF Sleep stage classification [46]

Table 13 – Dimensionality reduction methods.

Dimensionality reduction method Application References
PCA Depression [99,48]
Alcoholism [54,50]
Epilepsy/seizure [61,63,64,68]
Sleep stage classification [45]
ICA Epilepsy/seizure [61,63]
KPCA, ISOMAP, LLE, and LE Epilepsy/seizure [61]
LDA Epilepsy/seizure [63]
Autoencoder Multi-class task recognition [87]
Generalized Gaussian distribution Epilepsy/seizure [74]
Compensation distance evaluation technique Stress [102]
672 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 14 – Classification algorithms.

Classifiers Applications References
SVM [12,48,9,100,101,47,117,14,15,50,77,83]
[25,26,16,17,63,64,66,68,87,27,21,46]
LR [12,9,99,49,101,100,23,15,26]
NB [12,9,14,15,83,16,17,46]
LDA [99,50,17,74,87,21]
KNN [99,48,100,23,24,83,17,16]
GDA [48]
LS-SVM [100,14,54,70,61]
LD [100]
Complex tree [100]
Bagging [100,78,46,21]
RF [24,83,57,11,68,87,22,46,101]
CNN [13,84,62,87,97]
QDA [23]
MLP [80,24,59,10,50]
DNN [81,95]
LMT [50]
Ensemble classifier [77,78,22]
DT [83,16,46,87]
ANN [83,26,94]
Boosting [78,22,87,46]
DBN [59]
Dropout NN [59]
Fuzzy KNN [16]
PNN [16]
Weighted/unweighted N-TSK [17]
ME [10]
Proposed USVM and UTSVM, USVM, TWSVM, UTSVM [64]
MK-LSVM [61]
Optimized SVM (GA-SVM) [66]
GRNN [72]
GBM [68]
RNN [87]
FLDA [111]
ELM [27]
GBDT [22]
LSTM RNN [45]
Hierarchical classification based on probabilistic output SVM and proportion [89]
based clustering
HDCA, optimized HDCA [20]
GSLT-CNN, CV-CNN [21]

MDD, AUD, etc.) or monitoring of other applications (such as third gives the reference of the application study. The color
emotion recognition, sleep stage classification, etc.) with the coding for the Table 14 has been explained in Table 15 where
help of EEG signals. The machine learning classifiers such as each color signifies the type of an EEG application.
supervised, unsupervised, deep learning neural architectures, (b) Table 16 summarizes the details of the best
and ensemble learning models are used for the classification performing algorithms in the corresponding application
purposes in various EEG research studies. Some important
findings have also been concluded in the following sub-
heading that have applied the statistical tests on the extracted Table 15 – Color coding scheme for EEG applications.
features or other parameters for result analysis. The result
Color Application
analysis phase can be explained through classification models
and statistical analysis as below: Depression
Alcoholism
Dementia
(a) Classification models
MI
The survey for classification algorithms for various EEG
Eye state recognition
applications has been done into two parts: Identity authentication
(a) Table 14 gives an idea to the readers about the extent Emotion recognition
the particular classification algorithm has been explored for Epilepsy
various applications. Ist column gives the name of the Multi-class task recognition
classification algorithm, second column signifies the appli- Drowsiness detection
Sleep stage classification
cations which have worked upon these algorithms, and the
Table 16 – Classification algorithms.
Classifiers References Application Class labels Division ACC SEN SPEC F-score Others
SVM [12] Depression 2 (Normal and 10-fold CV 98 99.9 95 0.97 –
MDD)
[48] Depression 2 (High stress and 10-fold CV 80.32 – – – –
low stress)
[9] Depression 2 (Normal and 10-fold CV 98.4 96.66 100 – –
MDD)
[47] Emotion Recognition 2, 3, and 5 classes 8-fold 2 Classes – 73.06 – – – –
each for Valence (A), 73.14 (V), 3

biocybernetics and biomedical engineering 40 (2020) 649–690

(V) and Arousal (A) classes – 60.7 (A),
62.33 (V), 5 classes
– 46.69 (A) 45.32 (V)
[15] Alcoholism 2 (AUD and 10-fold CV 98 99.9 95 0.97 –
normal)
[77] Motor Imagery 2 (Left and right Holdout and 5- 80 (Holdout), 78.57 – – – –
hand imagery times 10-fold CV (10-fold)
movements)
[25] Eye state detection 2 (EO and EC) 2-fold CV 86.08 – – – –
[16] Dementia 2 (Parkinson and 10-fold CV 96.62 100 99.25 0.98 99.38 (Precision)
not)
[63] Epilepsy 2 (Normal and 800 samples for 100 100 100 – –
epileptic) training and 800
for testing
LR [99] Depression 2 (Normal and LOOCV 90 – – – –
depressed)
[49] Depression 2 (Normal and LOOCV 92 – – – –
depressed)
[101] Depression 2 (Normal and 0.8–0.2 72.41 86.21 – 0.7576 67.57 (Precision)
MDD)
[23] Depression 2 (Congruent and LOOCV 81.1 – – – –
incongruent
conditions)
[26] Eye state detection 2 (EO and EC) 5-fold CV 88.2 – – 0.882 Detection of eye
states takes 2 s
Weighted N-TSK [17] Dementia 2 (AD and control) Data for 46 97.30 95.48 98.32 – –
subjects used as
training and 16 as
testing
LS-SVM [100] Depression 2 (Normal and 10-fold CV 99.58 98.66 99.38 – 1 (AUC), 0.991 (MCC)
depressed)
[14] Alcoholism 2 (Alcoholic and 10-fold CV 99.17 99.17 99.17 99.16 99.44 (Precision),
non-alcoholic) 0.9833 (MCC), 0.9933
(AUC)
[54] Alcoholism 2 (Alcoholic and 10-fold CV 97.02 96.53 97.5 – 0.9494 (MCC)
normal)
[70] Epilepsy 2 (Focal and non- 10-fold CV 90 88 92 – –

673
focal)
 674
Table 16 (Continued )
Classifiers References Application Class labels Division ACC SEN SPEC F-score Others
[61] Epilepsy 2 (interictal and 10-fold CV 99.67 – – – –
ictal)
CNN [13] Depression 2 (Normal and 0.9–0.10/10-fold 95.49 94.99 96 – 1 (AUC), 0.991 (MCC)
depressed)
[84] Motor imagery 2 (Right hand and 0.80–0.20 99.35 98.80 100 0.994 0.9869 (Kappa
right foot) coefficient)
[62] Epilepsy 3 (Normal, pre- 0.9 (0.70–0.30)–0.10 88.67 95 90 – –
ictal and seizure) with 10-fold CV

biocybernetics and biomedical engineering 40 (2020) 649–690

[97] Sleep stage 5 (Wake, N1, N2, 0.5–0.2–0.3 87 – – 0.87 (F1-micro), 0.81 (Kappa)
classification N3 and REM) 0.78 (F1-macro)
MLP [80] Emotion Recognition 2 (Happy and sad) Out of 32 subjects, 58.5 – – – –
data for 30 used as
training and 2 as
testing
LMT [50] Alcoholism 2 or 3(Alcoholics, 10-fold 96 97 93 0.97 0.97 (AUC)
alcohol abusers,
and controls)
KNN [24] Emotion recognition 2 (Boredom and 10-fold 86.73 – – – 0.92 (AUC)
non-boredom)
[83] Motor imagery 2 (Right hand and 10-fold 94.57 – – – –
right foot)
DNN [81] Emotion recognition Low/high each for Out of 32 subjects, 62.50 (V), 64.25 (A) – – – –
Valence (V) and data for 30 used as
Arousal (A) training and 2 as
testing
[95] Sleep stage 4 and 2 from 10-fold sleep vs wake: – – – –
classification stages (W, S1, S2, 85.51, light sleep
S3, S4 and REM) vs deep sleep:
94.03, NREM vs
REM: 95.71
[26] Eye state detection 2 (EO and EC) 5-fold 88.2 – – 0.882 Detection of eyes
state takes 2 s
[94] Drowsiness detection 2 (Alert and 0.70–0.30 85.5 – – – –
drowsy)
RF [57] Eye state detection 2 (EO and EC) 0.66-0.34 99.8 – – – Classification
speed = 639.5 samples/s
[11] Epilepsy 2 and 3 (Normal, 10-fold 99.63 (seizure vs – – – –
interictal, and non-seizure)
ictal)
[46] Sleep stage 5 (N1, N2, N3, REM, 2-fold 89.4 – – – –
classification W)
Dropout NN [59] Eye state detection 2 (EO and EC) Holdout (0.8–0.1– 97.5 (Holdout) – – – Classification
0.1), 10-fold CV, speed = 1.9 s
LOOCV
ME [10] Epilepsy 2 (Epileptic and 1000 samples for 94.5 95 94 – –
normal) training and 600
for testing
Table 16 (Continued )
Classifiers References Application Class labels Division ACC SEN SPEC F-score Others
Proposed UTSVM [64] Epilepsy 2 (Healthy and 0.5–0.5 99 – – – Training
seizure) time = 0.01756 s
MKLSVM [61] Epilepsy 2 (Interictal and 10-fold CV 99.83 – – – –
ictal)
GA-SVM [66] Epilepsy 2 or 3 (healthy, 10-fold CV – 100 100 – 100 (PPV)
inter-ictal. ictal),
(inter-ictal and
ictal), (seizure and

biocybernetics and biomedical engineering 40 (2020) 649–690

non-seizure)
GRNN [72] Epilepsy 2 (pre-ictal and Training and – – – – Prediction
inter-ictal) testing: rate = 91.6%
approximately half
(labelled training
data, unlabelled
testing)
GBM [68] Epilepsy 3 (ictal, 10-fold CV 96.5 – – – 0.9695 (AUC)
intermittent, and
healthy)
LDA [74] Epilepsy 2 (seizure and non- LOOCV – 98 88 – Average detection
seizure) latency of 4 s thus
very fast method
XGBoost [87] Multi-class task 5 tasks (public 532,000 samples 0.794 0.781 – 0.7883 0.9456 (AUC)
recognition dataset) for training and
28,000 for testing
[87] 6 tasks (local 155,520 samples 0.7485 – – – –
dataset) for training set and
17,280 samples for
testing
FLDA [111] Drowsiness detection 2 (Alert and 10-fold – – – – Proposed method
fatigue) shows improved
performance than
traditional
approach
ELM [27] Drowsiness detection 2 (Alert and LOOCV 95.6 96.8 94 – Accuracy = 96.9
drowsy) (EEG + EOG)
Ensemble classifier [22] Sleep stage 6 (awake, NREM1, 0.70–0.30 with 5- 96.67 – – – 0.96 (Kappa
classification NREM2, NREM3, fold CV coefficient)
NREM4, REM)
[77] Motor Imagery 2 (Left and right Holdout and 5- 80 (Holdout), 78.57 – – – –
hand imagery times 10-fold CV (10-fold)
movements)
[78] Motor Imagery 2 (Left and right 0.5–0.5 (holdout 85.71(Adaboost – – – –
hand movement) technique) (Mix boost
ensemble))
LSTM RNN [45] Sleep stage 5, 4 and 2 (N1, N2, 0.80–0.10–0.10 with 86.74 (five classes) – – – –
classification N3, REM, W) 10-fold CV

675
676 biocybernetics and biomedical engineering 40 (2020) 649–690

studies as- name of the algorithm, reference of the

study, type of an application, class labels, division of the
dataset into training, validation (in some cases) and testing
datasets or name of the Cross Validation (CV) technique if
Others

used for computing the performance parameters, and

the rest of the columns give performance of the classifiers
–

–
in the form of various parameters such as accuracy (ACC)
in (%), sensitivity (SEN) in (%), specificity (SPEC) in (%),
f-score, and other parameters (such as precision (%), MCC,
etc.).
(b) Statistical analysis
F-score

There are some application studies that have used

different statistical tests for EEG signal analysis. Table 17
–

describes the name of the statistical test, features or

parameters on which the test is applied, results, and
SPEC

purpose or the application for which the tests are used

–

along with their references.

SEN
–

7. Other research studies

96 (157 subjects)

There are some research studies that have been separately

summarized in Table 18 with their important findings. Some of
94.26

them have worked upon the ERP measures or the evoked

ACC
91.4

potentials [107,112,114] for developing various applications

based on EEG signals, while we came across two studies that
are based upon classifying two neurological disorders simul-
taneously – (epilepsy and autism) [67] and (epilepsy and
alcoholism) [55], another study is using a combination of ECG
2-fold CV

5-fold CV

5-fold CV
Division

and EEG signals to classify the sleep stages [95], then stress
analysis is done using EEG, salivary cortisol level test, and VAT
[102], one study is proposing the circuit to make reliable ERP
measurements using low cost Emotiv Epoc+ headset [108],
5 (W, S1, S2, SWS,

then the visual comfort level for stereoscopic images are

2 (self-face and

2 (self-face and

predicted using a combination of EEG signals and eye tracking

non-self-face)

non-self-face)
Class labels

data [113], images are used instead of signals for classifying the
and REM)

MI tasks [84], and the MDD patients are classified from normal
subjects by using a combination of EEG signals, GSR, and eye
tracking data in another study [101]. For the attribute 'Dataset'
in Table 18, the description for the data collection has already
Identity authentication

Identity authentication

been summarized in Section 5.

8. Research gaps and future directions

classification
Application
Sleep stage

Unlike the other functional neuroimaging techniques, that are

only useful in diagnosing various neurological disorders, EEG
signals are successful in monitoring the other applications too,
References

such as identity authentication, emotion state recognition and

[89]

[20]

[21]

designing the BCI systems. The subtle variations that come in

the long EEG recordings are very difficult to analyse, time
Table 16 (Continued )

consuming, and their analysis depend upon the human

expertise. The signal processing and machine learning
proportion based
output SVM and

Optimized HDCA

algorithms based on computer aided technologies have been

classification

probabilistic

proven as the beneficial tools for the research. These extract

Hierarchical

clustering
Classifiers

based on

GSLT-CNN

the most important ncharacteristics from the EEG signals and

help to design the automated systems that can reduce the
human burden and increase the accuracy of the models. In
healthcare, the brain signals of the patients can be fed to the
 biocybernetics and biomedical engineering 40 (2020) 649–690 677

Table 17 – Statistical analysis.

Method Feature Result Purpose
t-test Alpha band power for both the Significant difference of p < 0.05 for alpha band power is found Depression
hemispheres for depressed and between two groups (depressed and normal) in: [99]
normal
 Five electrodes (C3, P3, O1, F7, T3) of left hemisphere.
 One electrode (O2) of right hemisphere.
 Higher values for depressed than normal.
Statistical test Power, Alpha interhemispheric  Less power in alpha and theta bands in all the regions (central, Depression
for depressed and normal occipital, frontal, parietal and temporal) for depressed than normal. [9]
 For MDD, it is more in right region than left for frontal region and
for rest of regions it is vice-versa. For normal, both the cases are
reverse.
 Significant difference of p < 0.01 for both parameters is found
between two groups (depressed and normal).
Mann–Whitney SASI, APV, RGP, HFD, DFA, LZC  Higher values for parameters for depressed than normal. Depression
test for normal and depressed [49]
 Significant difference ( p < 0.05) found for APV, RGP and HFD
between two groups (depressed and normal).
ANOVA, ROC KFD, HFD, LZC for MDD and  Higher values of complexity measures are found for MDD than Emotional
normal normal – significant difference of p < 0.01, p < 0.05 and p = 0.05 are state
found for features of KFD, HFD and LZC respectively in frontal and recognition
parietal locations. [106]
 When subjected to noise, MDD show more complexity than normal.
 For MDD, more complexity seen when subjected to noise than
music.
 For normal, significant difference found in frontal ( p < 0.05) and
parietal ( p = 0.05) regions for the parameters when subjected to
music than restline state.
 For MDD, significant difference found in frontal region ( p < 0.05) for
KFD and LZC when subjected to music than restline state.
t-test, chi- Power, SDS score and other  No significant difference between age, gender and marital status. Depression
square parameters for normal and [101]
depressed
 Significant difference found for SDS score ( p < 0.01) between two
groups.
 Significant difference of ( p < 0.05) found in EEG power of theta,
alpha, beta and gamma bands, lower values are observed for MDD.
Wilcoxon signed Stress is induced under Significant differences ( p < 0.01) found in: Depression
ranked test, external stimuli for normal [23]
Friedman test, subjects. Reaction times were
Bonferroni- observed for congruent and
corrected incongruent condition for
subject and also performance
of classifiers is also observed
under the two conditions
 Reaction time of subject for congruent and incongruent condition.
 LR and QDA work superior to 3-NN for fused features.
Significant differences ( p < 0.05) found in:
 Classifier for four electrodes (F3, F4, F7 and O2).
t-test AP and RP for both EO and EC  For AUD and controls, significant difference of p < 0.01 for AP and
RP for theta band in both EO and EC.
 For alcoholics and alcoholic abusers, significant difference of Alcoholism
p < 0.01 for AP and RP for delta band in both EO and EC. [50]
 Overall, RP shows more significant results than AP.
One-way MSE  Eight values for MSE were chosen according to their p-values
ANOVA ( p < 0.001 and p < 0.5) for each classification level of hierarchy.
 MSE curves for two channels (Fpz-Cz and Pz-Oz) are same for Sleep stage
healthy people and people with mild difficulty in sleeping. classification
[89]
Bonferroni test Cortisol level  Its value increases immediately after the stress test and after the Stress [102]
recovery of stress with a significant difference of p < 0.001 as
compared to the pre-stress condition
678 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 17 (Continued )
Method Feature Result Purpose
ANOVA test, Peak powers, their  Very low p-values for the features for burg's method. Alcoholism
ROC corresponding frequencies and and epilepsy
ratio of peak power and [55]
frequency derived from PSD for
normal, epileptic and alcoholic
using Welch, Yulker walker
and Burg's method
 For ROC, burg's method gives the most distincting results.
t-test EPN  Significant difference of p < 0.001 is observed values of EPN for Emotion
occipital electrodes O1 and O2 under the emotion effects state
recognition
[108]
Mann-Whitney MPE and MSE for EO and EC  Higher significant results found for MPE ( p < 0.05) than MSE. Higher Eye state
test, ROC significant difference for MPE found at electrodes: recognition
[88]
F5 0.00001
Fp1 0.000028
F2 0.00001
F4 0.00013
Fp2 0.00017
P4 0.0163
P3 0.0179
O1 0.023
 For ROC, maximum value of AUC at F5 (AUC = 0.90750) and we get p
value for F5 = 0.00001
Wilcoxon rank- PSI  Significant difference of p < 0.01 between AD and control for PSI Dementia
sum test value. [17]
 Lesser PSI value for Alzheimer patients depicts the weaker
functional connections than healthy subjects.
 Significant difference of p < 0.01 between AD and control for Eglobal,
Elocal and C, the values of these parameters are lower for AD patients
than normal subjects.
 Significant difference of p < 0.01 between AD and control for Bnodal
and Bedge, the values of these parameters are higher for AD patients
than normal subjects.
t-test Wavelet entropy and RWECN For states alert and fatigue, significant difference of p < 0.01 for Driver
wavelet entropy and RWECN. drowsiness
[111]

models that give automated decisions about their disorder, perform different tasks and so on. The various research gaps
thus reducing the time of the patients as well as the doctors in and the future directions have also been suggested by the
hospitals. Even the prediction models can be developed using authors in their studies based on different applications of EEG
EEG signals that keep on monitoring the mental state of the signals. These have been summarized in Table 19. The most
person and depending upon the change in behaviour, a common ones are:
warning system can be generated that help the patients to
reduce the risk of the disease. Suppose in one case, if the (a) The unavailability of larger datasets for most of the
seizure attacks of a person can be predicted in advance, a lot of applications that restricts the validation of the models
accidents or injuries can be prevented and in another case, the for practical use and restricts their feasibility for clinical
progression of the disease can also be measured like MCI can use.
lead to the Alzeihmer if not taken seriously or normal stress (b) The need of the hour is to develop the mobile, portable and
can lead to MDD if not diagnosed in advance. In other cheap models that are computationally less intensive and
applications too, the different functional states of the brain requires less storage space.
can be analysed by the automated systems based on EEG (c) The model should be simple and consists of least number
signals and help to design the applications such as authenti- of EEG channels.
cating the identity of the person, designing the warning (d) The real time data is full of artifacts, so very efficient
systems to alert the drivers about their drowsy state and so on. filtering techniques need to be proposed that help to
The BCI systems based on EEG signals can help the disabled reduce the noise and increase the performance of the
people a lot to perform their daily tasks such as controlling models.
their wheel chairs, closing or opening the doors of the lifts with (e) A variety of feature extraction techniques and machine
their motor imaginary based brain signals, controlling the learning classifiers can be explored for different datasets.
television screen in their houses or work environments to
 biocybernetics and biomedical engineering 40 (2020) 649–690 679

Table 18 – Other research studies.

Purpose Dataset Methodology Findings
Emotion recognition Own dataset Single trial ERP and latency  The ERP attributes have been directly obtained from the
[107] features + SVM filtered EEG signals at different latencies.
(polynomial)
 4 class emotion classification has been performed with good
accuracies under the subject independent as well as subject
dependent scenario.
 The proposed methodology in this study motivates to
develop the real time applications based on emotion
classification with reduced processing time.

brain computer Own dataset SSVEP responses  An asynchronous SSVEP based BCI system for spelling
interface [112] interface for high frequency RVS has been developed using a
single EEG channel.
 LASSO algorithm has been adopted as the frequency
detection method to evaluate the SSVEP responses in the
range 6-60 Hz.
 The frequency set of 35-40 Hz is best suited for BCI with best
accuracy of 99.2%, ITR of 67.1 bit/min and maximal user
comfort level of 80%.

Visual, auditory, and Own dataset VEP, AEP, and SEP  The best positions on the scalp have been investigated for
somatosensory measured under different the measurement of the evoked potentials so that the
stimulations [114] stimulating conditions. wearables with few electrodes can be used for real time EEG
applications.
 The activities occurring due to the visual, auditory, and
somatosensory stimulations are detected in the Lobus
Occipitalis and Lobus temporalis.
 The combination of VEP, AEP, and SEP can be detected at Oz,
O1, O2, TP9, and TP10.
 It is also investigated that segmentation frequency should
match the stimulation frequency with an accuracy of at least
99.92% for VEP detection and 99.95% for AEP and SEP
detection.

Epilepsy and autism Own + public DWT, Shannon entropy  Based on two class scenario: Epileptic vs normal for single
[67] dataset and KNN channel and autistic vs normal for single and multi-channel,
as well as three class scenario: epileptic vs normal vs autistic
for single and multi-channel.
 Thus it is able to classify the data for two neurological
disorders (autism and epilepsy) at the same time.
 It gives an overall accuracy of 94.6% for the three class
classification scenario.

Alcoholism and Public dataset PSD using Welch method  Analysis has been done to distinguish the normal, epileptic
epilepsy [55] and AR model and alcoholic subjects using ROC method.
 Burg's method gives the most distinguishing features.

Sleep stage Public dataset ECG (IMFs) + EEG (sub-  The proposed method is able to get the highest classification
classification [95] bands): (RQA, dispersion accuracies for different combinations of classes using the
entropy and variance) features from both the RR-time series (ECG) as well as EEG
+ Deep neural architecture data.
 Uses only the single channel data recordings of both EEG
and ECG data.
 Simulation time for evaluating the features from RR time
series and EEG signals take 0.03 and 4.89 s respectively,
classification using both the features takes 0.275 s.

Stress [102] Own dataset SL + CDET + SVM (RBF)  Salivary cortisol levels and VAT are also measured along
with EEG to check the stress level changes during the three
different scenarios.
 The effect of stress vanishes after 20 min of the test.
 Accuracies for EO and EC for alpha waves are 74.32% and
91.21% respectively, i.e., alpha waves are more sensitive to
anxiety in EC state.
 Accuracies for EO and EC for beta waves are 92.31% and
93.62% respectively.
 Good classification accuracies for alpha and beta
bandwidths are observed for monitoring the stress levels.
680 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 18 (Continued )
Purpose Dataset Methodology Findings
Emotion state Own dataset ERP and stimuli display  The marking circuit for determining the stimuli display time
recognition [108] timestamps, EPN stamps has been proposed for attaining the reliable ERP
measurements in case of low cost Emotic Epoc+ headset.
 Differences can be observed for EPN components under the
effect of different facial expressions using the proposed
modifications.
 The proposed scheme can be a very good solution to carry
out the research using the low cost EEG devices.

Visual comfort level Own dataset Multi-taper method + Eye  This aims at finding the set of significant features for
of images [113] tracking features and EEG predicting the visual comfort level for stereoscopic images.
activity from frontal lobe
 The best results are found for 2-s and 5-s pre-DPI windows
for the selected feature set.
 Increased activity was observed for beta, theta-alpha ratio,
alpha-high beta ratio in case of visually uncomfortable
stereoscopic perception.

Motor imagery [84] Public dataset Signals transformed to  The EEG signals are not directly used instead the
images using STFT and transformed images are fed to the deep CNN to classify the
CWT + AlexNet motor imagery tasks.
 CWT performs better than the STFT transform.
 Overall accuracy of 99.35% is achieved through the proposed
approach.

Depression [101] Own dataset EEG, GSR, Eye tracking  Using only EEG signals gives an f1 score of 75.76% for
data + LR classification of MDD and normal.
 The combination of EEG, eye tracking and GSR data achieves
an improved performance with f1 score of 80.70%.

Table 19 – Research gaps and future directions.

Purpose Research gaps and future directions
Depression  Hospitalized MDD patients can be explored for more realistic results [101,12,9].
 The depression study is still under-diagnosed among the older subjects, so there is a scope to carry further study on
the depressed older patients [101].
 Smaller dataset is the limitation of the most of the studies [12,100,99,49,9,48].
 For more detailed analysis, EEG can be combined with other modalities such as fMRI or fNIRS.
 More heterogeneous dataset can be collected covering the different severity stages for depression and people of varied
age groups and genders [100,13,49].
 The most effected regions of the brain due to depression need to be found and also EEG data can be collected under
varied conditions instead of rest state only [99].
 The various modalities such as EEG, GSR, and eye-tracking can be integrated to get high performance for the classifier
models.
 Significant hemispheric asymmetry can be observed in MDD patients by including more number of EEG channels [101].
 Apart from low and high, more number of classes of stress can be studied further by using voting algorithms, etc. [48].
 To the best of our knowledge, there are no public datasets available for depression based on EEG signals, so it is a
challenging task to collect the own heterogeneous dataset of MDD patients for different age groups using the different
EEG headsets.
 Most of the research studies are carried out under the laboratory conditions, they need to be tested in mobile and non-
laboratory environment for real time deployment of research models in daily life.
 For more detailed analysis, EEG data can be combined with the other parameters such as heart rate, blood pressure,
breathing patterns, and body posture.
 The levels of various biomarkers such as cortisol, alpha amylase enzyme and catecholamine should also be observed
for more validated results.
 To the best of our knowledge, no datasets are available for depression and stress, so efforts can be made to collect the
own dataset covering the various levels of the problem in varied age groups such as workers at construction sites,
children in school or the youth taking the drugs due to the various life stresses.

Alcoholism  The work can be carried further on the bigger datasets with proper selection of participant's biological characteristics
such as age group, gender, etc. [14,15,54,50].
 Systems consuming lesser storage space need to be designed [54].
 biocybernetics and biomedical engineering 40 (2020) 649–690 681

Table 19 (Continued )
Purpose Research gaps and future directions
 The automated machine learning methods can be combined with various traditional alcohol screening methods such
as questionnaire based to achieve higher efficiency of the system and make its deployment feasible in clinical practice
[50].

Epilepsy/seizure  Most of the research studies have focused on the dataset provided by Bonn University, Germany. It has been made
available publicly for more the last 15 years and is small and less comprehensible. So, it is required to work upon more
larger and comprehensible datasets.
 The intracranial EEG recordings for epilepsy need to be replaced with more effective and simpler solutions based on
scalp EEG.
 New methods can be proposed in order to automate the selection of optimized kernel parameters for LS-SVM used in
the study [70].
 The performance can be further improved by working on larger datasets, extracting much better features, using more
pre-processing methods and using more machine learning classifiers [70,61,62,11,66].
 A work can be done to design new muscle artifact removal techniques in order to filter the epileptic dataset used in the
study [61].
 Apart from epilepsy, the proposed methodology can be used for the diagnosis of various neurological disorders based
on EEG signals such as Alzheimer, coronary artery disease [61,11].
 Bagging algorithms can be used to increase the performance of the proposed work [62].
 Efforts should be made to develop the research models that use least number of electrodes in order to make the
wearable, portable and mobile applications to be deployed in real time applications.
 Work can be done to improve the computational time of the study and to solve the multi-class classification problems
based on EEG signals [64].
 Research on seizure studies can be extended to work upon more different number of time-frequency transforms [64].
 Optimized-SVM can be used for the diagnosis of other classification problems due to its splendid ability of
classification [66].
 Different severity stages of the disease can be studied for deeper analysis or for understanding the progression of the
disease [67].
 Alzeihmer can also be included in the given classification problem along with epilepsy and autism [67].
 Optimized GSO can be used for the automatic selection of optimized parameters for different classifiers [68] or even
some other optimization algorithms can also be used for the selection procedure.
 Prediction models based on early warnings before the onset of epileptic seizures can be designed that can help to avoid
severe accidents or injuries.
 The long-term EEG recordings can be used for thorough evaluation of the proposed work [74].
 The false positives that occur due to the noise and random peaks in the epileptic signals can be handled by adding the
regularization parameters at the training stage [74].
 It is necessary to track the dynamic offsets arising due to the seizures because they lead to the degradation in
classification performance [74].
 The source location for EEG signals is required in order to characterize the spatio-temporal wave patterns [74].

Identity  The bigger datasets containing data for multiple days, tasks, and sessions are needed to build more validated models
authentication for biometric identification [21].
 The authentication models needs to be designed for personal needs keeping in mind the requirements such as high
performance and low cost.
 Portable EEG devices such as Emotiv Epoc can be used for the research purposes so that the models can be deployed in
practical scenarios [20].
 The performance of the system can be improved further by choosing a set of more random people who are not a part of
the training phase, thereby giving more realistic results [20].

Alzeihmer  Large and heterogeneous dataset for MCI needs to be collected and analysed for MCI and its various sub-types. Also the
study about the progression of MCI to Alzheimer can be worked upon [51].
 More advanced classification algorithms such as spiking neural networks and SVMs can be included in the work as an
extension [51].
 The combination of fuzzy systems and functional networks can be used as an effective tool for the automatic diagnosis
of various neurological diseases by making use of various machine learning algorithms [17].
 The potential biomarkers for AD can be explored with the help of N-TSK system [17].
 The early warning systems of AD with the help of machine learning algorithms can be designed.
 The concept of network theory can be integrated with the unsupervised or semi-supervised learning methods to
improve the identification performance.

Sleep stage  The accuracy for the NREM1 stage can be improved by using different feature extraction and analysis techniques [46].
classification
 Analysis can be done on the patients with sleep related disorders.
 Large and heterogeneous databases can be used for more validated results.
 Sleep stage classification can be improved by using multiple channels of EEG and more physiological signals such as
EMG and EOG can be combined with the EEG signals for improved performance [97,46].
 The combination of RNN and CNN can be used where the raw signals are fed to the CNN for the feature extraction and
RNN does the classification task [45].
682 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 19 (Continued )
Purpose Research gaps and future directions
 CNN and RNN can be used for the spatial and the temporal extraction of features respectively [45].
 Classification performance for the 'wake' stage can be improved and also the layer by layer errors that arise in the
hierarchical structure needs to be handled [89].
 More non-linear features such as HOS, entropies and teager energy as well as the deep learning architectures can be
used for high performance results [95].
 Study for sleep stage classification can be done on the patients with disorders such as bruxism, epilepsy, and insomnia
using various non-linear features [95].

Motor imagery  An adaptive scheme for selection of k-value (parameter for selecting the feature sub-set in the FDM) is required to be
developed with the aim to increase the discernibility between the decision classes [77].
 Various time-frequency methods can be used to convert the signals into the images [84].
 The study can be extended on exploring the various CNN models such as VGGNet, ResNet and GoogleNet [84].
 The different types of classifiers can be ensemble in a mixture mode to get more improved performance results [78].
 Bigger datasets can be explored for the research work for more validated results.
 A lot of scope is there to carry research studies on neuroprosthesis for designing the BCI systems for disable people.

Eye state  Many of the research studies are based upon the UCI dataset for eye state recognition which is a very small dataset
recognition containing the data samples for a single user. So, the research can be extended on bigger datasets for more validated
results.
 Temporal RBMs for time-series data models can be used for more improved learning [59].
 Work can done to increase the classification speed of the proposed models by following the algorithms having the less
intense and fast computations.
 The ensemble of deep learning architectures may be proven as the effective models for the classification problems
based on time-series data [59].
 The dataset for eye state recognition can be used for controlling various applications such as BCI, driver drowsiness,
and home automation.
 The fuzzy cognitive maps generated using LSTM and GA can be merged with IoT to provide person-centric monitoring
in the healthcare applications [58].
Emotion state  Work can be done on more than two emotional states so that the model can be deployed to control the real-time
recognition applications such as brain controlled wheel-chairs, video-games, etc. [80,47].
 More number of classifiers can be explored as an extended research work [80].
 Bigger and balanced datasets can be used for the research.
 Overfitting issues can be solved for improved performance results [24].
 GSR data can be combined with EEG to gain higher accuracies [24].
 More model design parameters such as the different frequency bands, number of subjects, etc., can be varied for the
testing the performance of the study [47].
 More number of complexity measures such as RQA, HOS, and sample entropy can be used to analyse the MDD patients
under different emotional states [106].
 The proposed methodology can be applied to the other neurological disorders such as schizophrenia, epilepsy, and
bipolar depression with the aim to compare the findings for all the studies [106].
 The factors such as level of severity, gender and the medication status of the subjects should also be considered in the
research work as a part of more detailed analysis [106].
 Deep neural architectures such as RNN can be used for the classification [81].
 Advanced filtering techniques and averaging methods can be designed in order to pre-process the data taken using the
Emotiv Epoc+ headset in the research studies [108].

ERP related  The concept of evoked potentials can be leveraged to find the least number of electrodes for various EEG applications
[114].
 The various machine learning techniques can be combined with the ERP features for improving the classification
performance [108].

Deep learning and ensemble architectures can help to neuronal activities. This fusion gives a more clear picture of
improve the accuracy of the models. the neuronal structure and functions and can help in finding
(f) Non-linear and functional connectivity features can be the more accurate biomarkers for diagnosing various neuro-
studied for various applications to understand the com- logical and neuropsychiatric disorders. For an instance, EEG
plexity of the EEG signals. and fMRI as the single modalities may not give high spatial and
temporal resolutions respectively. But their fusion can achieve
both. Now-a-days, the fusion of data from multiple modalities
is considered as a new research challenge because different
9. Multi-modal fusion of brain signals
modalities may represent the data in the form of uncommon
patterns and with different orders and it is not easy to directly
Integrating the brain signals from different neuroimaging fuse them. In the present work, we have explored some of the
modalities can give better understanding and analysis of research studies which have contributed in proposing the
 biocybernetics and biomedical engineering 40 (2020) 649–690 683

Table 20 – Multi-modal fusion of brain signals.

Application Data Methodology Future scope/limitations
Schizophrenia  EEG + fMRI + sMRI  Structure-revealing CMTF (coupled matrix  Done on limited number of
[127] and tensor factorization) method (advanced patients.
CMTF), as a fusion method, has been used
to exploit the potential biomarkers for
diagnosis of schizophrenia.
 HC: 21, Schizophrenia  Study can be carried forward to
patients: 11 during an auditory different neurological disorders.
oddball task (AOD).
 Pre-processing the fused data is a
new area of interest.
 It aims at finding the unique features for
identifying the patients using multi-modal
data with shared and unshared patterns.
 It is able to provide excellent temporal and
spatial resolution.

Brain–computer  EEG + NIRS (near-infrared  An asynchronous BCI system has been  The model needs to be validated in
interface [128] spectroscopy) developed to accurately detect an idle class. real-time asynchronous BCI-based
paradigm.
 HC: 14 (right-handed  Two subject-dependent classification
participants) under cue-based algorithms have been proposed.
paradigm
 Two subject-dependent classification
algorithms have been proposed.
 The hybrid EEG-NIRS model is not only
successful in improving the classification
performance but also improves the delays
of the overall model that are caused due to
the slow hemodynamic response of NIRS.

Alzeihmer [129] MRI + PET + CSF  Zero-masking method has been used to  Four-class classification
fuse the data from multiple modalities. performance is required to be
improved for deployment of the
model in clinical scenarios.
 200 normal instances, 400  A deep learning architecture combining  More number of training samples
MCI instances, 200 AD patients the multi-modal data for computer-aided with smaller variance are required
diagnosis of AD has been presented. to check the validity of the model.
 A performance gain has been achieved for
binary and multi-class classification
scenario.

Alzeihmer [130]  Voxel based morphometry  The present fusion methods have  In future, focus can be laid on
(VBM) + fluorodeoxyglucose following limitations: – unable to preserve solving the non-convex
positron emission tomography the structure information while fusing the optimization problem in multi-
(FDG) + F-18 florbetapir PET data from different modalities, – the modal fusion models.
scans amyloid imaging (AV45) contributions of the samples from different
data is considered as equal in the fused
data.
 To address the above problems, a latent
correlation embedded multi-modal fusion
(LLM2F) algorithm has been proposed.
 This method is able to provide compact
representation of the multi-modal data by
exploring the latent correlations among the
different modalities and fusing the data into
a common feature level.
 To dynamically evaluate the contribution
of each sample in the fused model, a self-
optimized learning method has been
adopted.

Epilepsy [131]  EEG + fMRI  At the time of data fusion from multiple  In future, efforts will be made to
modalities, a joint BSS (blind source better understand why jointICA and
separation) problem arises. This has been CMTF are able to provide stable EEG
mathematically formulated in the present and fMRI voxel signatures
work. respectively.
684 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 20 (Continued )
Application Data Methodology Future scope/limitations
 Epilepsy patients: 5 (right  To solve the above issue, different  Efforts will be made to explore
temporal) and 5 (left temporal) approaches such as joint independent more advanced models such as
component analysis (jointICA) and coupled advanced matrix-tensor
matrix-tensor factorization (CMTF) have decomposition (ACMTF).
been presented as solutions.
 These have been used to fuse EEG and
fMRI data for analysing the interictal
activities with high spatio-temporal
resolution.

solutions for fusing the data from multiple modalities. These signals with the help of artifact handling methods, then for
have been tabulated in Table 20. feature extraction studies, maximum studies have worked
upon the statistical features and in future, there is a greater
scope to work upon the functional connectivity based features,
10. Conclusion
then post-processing is done with the aim to reduce the
computational burden, for that, maximum studies are based
The functional neuroimaging capabilities such as excellent upon selecting the most significant features with the help of
temporal resolution, non-invasiveness, inexpensiveness, and various feature selection methods, and lastly, for the result
safe nature makes the study of EEG to be very crucial for analysis phase, maximum number of studies are focusing on
understanding the dynamically changing complex processes the classification algorithms for developing the automated
of the brain. The varied frequency rhythms are associated with recognition systems for various applications.
different functional states of the brain. Any minute changes in It is very time consuming and tedious task to manually
the frequencies of these rhythms can be well captured by the analyse the complex and non-stationary EEG signals and the
EEG signals. These signals are analysed with the help of analysis results vary a lot depending upon the expertise
computer-aided technologies with greater accuracies and experience of the visualizers. So, now-a-days, a lot of research
speed. is going on analysis the EEG signals using various computer-
The present study explores a number of data acquisition aided technologies that could automate the analysis task
methods for wide variety of applications based on EEG signals. thereby giving fast and highly accurate results. From the
A comparative analysis of the signal processing methods has survey, it can be concluded that these computer-based
been made that involves a number of pre-processing, feature systems make use of various signal processing and machine
extraction, and postprocessing techniques. Then, the result learning schemes to automatically conclude the happening of
analysis stage is discussed, mainly focusing on the classifica- some neuronal activity using EEG signals. Computer-based
tion methods based upon various machine learning models. programming languages or softwares such as MATLAB,
From the studies, it can be concluded that every stage has its python, R, WEKA, and so on are used to implement these
own crucial role in processing the raw EEG signals. Each of the signal processing methods and machine learning techniques.
stages- preprocessing, feature extraction, post-processing, It can be concluded that EEG based computer-aided systems
and result analysis play a very significant role in processing have shown their potential successfully in various research
the raw time-domain EEG signals for developing the comput- applications, covering the diagnosis of different neurological
er-aided automated decision models. Pre-processing the raw disorders such as epilepsy/seizure, alcohol related disorders,
signals at the first stage diminishes the unwanted frequency depression, and dementia to the monitoring of other applica-
components and noise from the signals, thereby, enhancing tions including emotion recognition, identity authentication,
the quality of the signals. At next stage, various feature sleep stage classification, eye state detection, motor imagery
extraction methods are adopted to represent the high- and drowsiness monitoring. Then, the future scope of the
dimensional EEG data in the form of most discriminating various studies has been summarized in order to inspire the
features, without this step, the performance of the decision readers to take the study of EEG signals based on computer-
model might get degraded. Then, if the data is still very high aided technologies to more higher level of research. Finally,
dimensional or suffers from a problem of overfitting, feature some of the research studies have been explored that focuses
selection and reduction algorithms play their significant role, on the fusion of brain signals from multiple modalities.
thereby reducing the burden on the computational resources In future, efforts will be made to explore more advanced
and cost of the model. The signals are still of no use, unless applications on EEG signals and study a wide variety of signal
they are not processed through the classification models (such processing and classification methods used in their analysis.
as traditional algorithms or deep learning architectures) or the There are various applications that are least or not covered in
statistical tests for some decision making or deducing some the present study that are working upon EEG based computer-
findings through them. From the survey, it can be concluded aided methods in their analysis. Such as study of EEG signals
that maximum number of studies are preprocessing the for patients with neurological disorders like Huntington's
 biocybernetics and biomedical engineering 40 (2020) 649–690 685

disease, Schizophernia, Autism, Strokes, Rett syndrome, work, it has been planned to extend the work on EEG based
attention deficit hyperactivity disorder (ADHD), and sleep depression diagnosis by creating the huge dataset of the
related disorders. Efforts will be made to cover the extent of patients and working towards their analysis using various
research going on signal processing and classification meth- signal processing and classification methods.
ods used in these applications. It has been observed that
maximum number of studies on EEG based diagnosis of
Authors' contribution
epilepsy using computer-aided methods have worked upon a
very small dataset containing the EEG signals for only 5
patients. So, in future, efforts will be made to collect the data Ashima Khosla: conceptualization, writing – original draft
for more number of epileptic patients from a renowned preparation, writing – reviewing and editing, visualization.
hospital to create huge and heterogeneous (data of patients Padmavati Khandnor: writing – reviewing and editing, super-
with varying age, gender, and so on) dataset. Advanced signal vision. Trilok Chand: writing – reviewing and editing,
processing methods and classification methods will be used to supervision.
process these signals with the aim to get more validated
results and higher classification accuracies. Next, to the best of
Conflict of interest
our knowledge, it has been studied that no EEG dataset is
publicly available for Major Depressive Disorder (MDD) that
has led to the limited research on this field. As a part of future None declared.

Appendix A. List of acroynms and abbreviations

The list of abbreviations and acroynms has been summarized in Table 21.

Table 21 – List of acroynms and abbreviations.

Acronynm/ Full form Acronynm/ Full form
abbreviation abbreviation
AAR Adaptive auto-regressive KFD Katz fractral dimension
AC Asymmetry coefficient KNN K-nearest neighbor
AD Alzheimer's disease KPCA Kernel principal component analysis
ADhall+ AD with hallucinations LD Linear discriminant
ADhall- AD without hallucinations LDA Linear discriminant analysis
AEP Auditory evoked potentials LE Laplacian eigenmaps
AMI Auto-mutual information LMS Least mean square
ANN Artificial neural network LLE Locally linear embedding
ANOVA One-way analysis of variance LL2N Logarithm of L2 norm
AP Absolute power LMT Logistic model tree
APV Alpha band power variability LOOCV leave one out cross validation
AUC Area under curve LR Logistic regression
AUD Alcohol use disorder LSTM Long short – term memory
BCI Brain computer interface LS-SVM Least square support vector machine
BDL Bandwidth-duration localized LASSO Least absolute shrinkage and selection
operator
BOLD Blood oxygenation level depredent LVHA Low valence high arousal
signal
Bedge Edge betweeness LVLA Low valence low arousal
Bnodal Node betweeness LZC Lempel-ziv complexity
C Clustering coefficient M Male
CD Correlation dimension MaxF Maximum frequency
CDET Compensation distance evaluation MCC Matthews correlation coefficient
technique
CWT Continuous wavelet transform MCI Mild cognitive impairment
CNN Convolutional neural network MDD Major depressive disorder
CTM Central tendency measure ME Mixture of experts
CV Cross validation MEG Magnetoencephalogram
DASM Differential asymmetry MKLSVM Multiple kernel learning SVM
DBN Deep belief network MLP Multi-layer perceptron
DD-DWT Double-density discrete wavelet MLPNN Multi-layer perceptron neural networks
transform
DE Differential entropy MMSE Mini Mental State Examination
DEAP Database for Emotion Analysis using MODWPT Maximal overlap discrete wavelet
Physiological Signals package transform
686 biocybernetics and biomedical engineering 40 (2020) 649–690

Table 21 (Continued )
Acronynm/ Full form Acronynm/ Full form
abbreviation abbreviation
DLB Dementia with Lewy bodies mRMR Minimum redundancy maximum
relevance
DLBhall+ Dementia with Lewy bodies with MSE Multiscale sample entropy
hallucinations
DFA Defrended fluctuation analysis MPE Multiscale permutation entropy
DM Discernibility matrix NB Naïve Bayesian
DNN Deep neural network NN Neural network
DPI Depth perception indication NREM Non-rapid eye movement
DT Decision tree N-TSK Network based Takagi–Surgeno–Kang
DWT Discrete wavelet transform PCA Principal component analysis
EC Eyes closed PD Parkinson disease
ECG Electrocardiogram PET Positron emission tomography
EEG Electroencephalogram PFD Petrosian fractral dimension
EI Electrode impedance PNN Probabilistic neural network
EMD Empirical mode decomposition PPV Positive predictive value
EMG Electromyogram PSD Power spectral density
ELM_sig Extreme learning machine with PSI Phase synchronization index
sigmoid activation function
ELM_RBF Extreme learning machine with RBF QDA Quadratic discriminant analysis
kernel
EO Eyes open Q1F First quartile frequency
EOG Electrooculogram Q3F Third quartile frequency
EPN Early posterior negativity RASM Rational asymmetry
EPSPs Exhibitory post synaptic graded RBF Radial basis function
potentials
ERP Event related potentials RBM Restricted Boltzmann machine
EWT Empirical wavelet transform REM Rapid eye movement
Eglobal Global efficiency RF Random forest
Elocal Local efficiency RGP Relative gamma power
F Female RLS Recursive least square
FAWT Flexible analytical wavelet transform RNN Recurrent neural network
FD Fractral dimension ROC Receiver operating curve
FDM Fuzzy discernibility matrix RP Relative power
FFT Fast Fourier transform RPS Reconstructed phase space
FIR Finite impulse response RQA Recurrence quantification analysis
FLDA Fisher linear discriminant analysis RVS Repetitive visual stimulus
fMRI Functional magnetic resonance RWECN Relative wavelet entropy in complex
imaging networks
fNIRS Functional near infrared spectroscopy SASI Spectral asymmetry index
GBDT Gradient boosting decision tree SDS Self-rating depression score
GBM Gradient boosting machine SEP Somatosensory evoked potentials
GDA Gaussian discriminant analysis SHHS Sleep Heart Health Study
GGD Generalized Gaussian distribution SHR Split-half-reliability
GRNN Generalized regression neural network SSVEP Steady-state visual evoked potentials
GSLT-CNN CNN with global spatial and local SQUIDs Superconducting quantum interference
temporal filter devices
GSO Grid search optimizer SSD Spectral standard deviation
GSR Galvanic skin response STFT Short time Fourier transform
GST Generalized Stockwell transform SVD Single value based decomposition
HC Healthy control SVM Support vector machine
HDCA Hierarchical discriminant component SWS Slow wave sleep
analysis
HE Hurst exponent TCOWFB Three-channel orthogonal wavelet filter
bank
HFD Higuchi fractral dimension TMS Transcranial magnetic stimulation
HOS Higher order spectra TSST Trier social stress test
HVHA High valence high arousal TWSVM Twin SVM
HVLA High valence low arousal TQWT Tunable-Q wavelet transform
IAPS International Affective Picture System UCI University of California Irvine
ICA Independent component analysis USVM Universum SVM
IIR Infinite impulse response UTSVM Universum twin SVM
IMFs Intrinsic mode functions VAT Visual analog test
IPSPs Inhibitory post synaptic graded VEP Visually evoked potentials
potentials
IR Interquartile range WT Wavelet transform
 biocybernetics and biomedical engineering 40 (2020) 649–690 687

Table 21 (Continued )
Acronynm/ Full form Acronynm/ Full form
abbreviation abbreviation
ISOMAP Isometric feature mapping WPD Wavelet packet decomposition
KC Kolmogorov complexity XGBoost Extreme gradient boosting

references

preliminary study. 2015 37th Annual International

[1] La Vaque T. The history of EEG Hans Berger: Conference of the IEEE Engineering in Medicine and
psychophysiologist. A historical vignette. J Neurother Biology Society (EMBC). IEEE; 2015. p. 1468–71.
1999;3(2):1–9. [19] Müller-Putz GR, Ofner P, Schwarz A, Pereira J, Pinegger A,
[2] Bronzino JD. Biomedical engineering handbook, vol. 2. CRC Dias CL, et al. Towards non-invasive EEG-based arm/hand-
Press; 1999. control in users with spinal cord injury. 2017 5th
[3] Kandel ER, Schwartz JH, Jessell TM, D. of Biochemistry, International Winter Conference on Brain–Computer
Jessell MB, Siegelbaum S, et al. Principles of neural science, Interface (BCI). IEEE; 2017. p. 63–5.
vol. 4. New York: McGraw-Hill; 2000. [20] Zeng Y, Wu Q, Yang K, Tong L, Yan B, Shu J, et al. EEG-
[4] Beres AM. Time is of the essence: a review of based identity authentication framework using face rapid
electroencephalography (EEG) and event-related brain serial visual presentation with optimized channels.
potentials (ERPS) in language research. Appl Psychophysiol Sensors 2019;19(1):6.
Biofeedback 2017;42(4):247–55. [21] Chen J, Mao Z, Yao W, Huang Y. EEG-based biometric
[5] Novik O, Smirnov F, Volgin M. Structures of the brain. identification with convolutional neural network.
Electromagnetic Geophysical Fields. Springer; 2019. p. 69– Multimedia Tools Appl 2019;1–21.
89. [22] Wang Q, Zhao D, Wang Y, Hou X. Ensemble learning
[6] Das S, Tripathy D, Raheja JL. An insight to the human brain algorithm based on multi-parameters for sleep staging.
and EEG. Real-Time BCI System Design to Control Arduino Med Biol Eng Comput 2019;1–15.
Based Speed Controllable Robot Using EEG. Springer; 2019. [23] Blanco JA, Vanleer AC, Calibo TK, Firebaugh SL. Single-trial
p. 13–24. cognitive stress classification using portable wireless
[7] Teplan M, et al. Fundamentals of EEG measurement. Meas electroencephalography. Sensors 2019;19(3):499.
Sci Rev 2002;2(2):1–11. [24] Seo J, Laine TH, Sohn K-A. Machine learning approaches
[8] Da Silv FL. Electroencephalography: basic principles, for boredom classification using EEG. J Amb Intell Human
clinical applications, and related fields. Lippencott Comput 2019;1–16.
Wlliams & Wilkins; 2005. [25] Kaur B, Singh D, Roy PP. Eyes open and eyes close activity
[9] Mumtaz W, Xia L, Ali SSA, Yasin MAM, Hussain M, Malik recognition using EEG signals. International Conference on
AS. Electroencephalogram (EEG)-based computer-aided Cognitive Computing and Information Processing.
technique to diagnose major depressive disorder (MDD). Springer; 2017. p. 3–9.
Biomed Signal Process Control 2017;31:108–15. [26] Saghafi A, Tsokos CP, Goudarzi M, Farhidzadeh H. Random
[10] Subasi A. Eeg signal classification using wavelet feature eye state change detection in real-time using EEG signals.
extraction and a mixture of expert model. Expert Syst Appl Expert Syst Appl 2017;72:42–8.
2007;32(4):1084–93. [27] Chen L-l, Zhao Y, Zhang J, Zou J-z. Automatic detection of
[11] Zhang T, Chen W, Li M. Generalized Stockwell transform alertness/drowsiness from physiological signals using
and SVD-based epileptic seizure detection in EEG using wavelet-based nonlinear features and machine learning.
random forest. Biocybern Biomed Eng 2018;38(3):519–34. Expert Syst Appl 2015;42(21):7344–55.
[12] Mumtaz W, Ali SSA, Yasin MAM, Malik AS. A machine [28] Brašic´ JR, Mohamed M. Human brain imaging of autism
learning framework involving EEG-based functional spectrum disorders. Imaging of the human brain in health
connectivity to diagnose major depressive disorder (MDD). and disease. Elsevier; 2014. p. 373–406.
Med Biol Eng Comput 2018;56(2):233–46. [29] Ewers M, Sperling RA, Klunk WE, Weiner MW, Hampel H.
[13] Acharya UR, Oh SL, Hagiwara Y, Tan JH, Adeli H, Subha DP. Neuroimaging markers for the prediction and early
Automated EEG-based screening of depression using deep diagnosis of Alzheimer's disease dementia. Trends
convolutional neural network. Comput Methods Programs Neurosci 2011;34(8):430–42.
Biomed 2018;161:103–13. [30] Crosson B, Ford A, McGregor KM, Meinzer M, Cheshkov S,
[14] Anuragi A, Sisodia DS. Alcohol use disorder detection Li X, et al. Functional imaging and related techniques: an
using EEG signal features and flexible analytical wavelet introduction for rehabilitation researchers. J Rehabil Res
transform. Biomed Signal Process Control 2019;52:384–93. Dev 2010;47(2). vii.
[15] Mumtaz W, Kamel N, Saad MNbM, Ali SSA, Malik AS. An [31] Wong DF, Gründer G, Brašic´ JR. Brain imaging research:
EEG-based functional connectivity measure for automatic does the science serve clinical practice? Int Rev Psychiatry
detection of alcohol use disorder. Artif Intell Med 2007;19(5):541–58.
2018;84:79–89. [32] Malhi G, Lagopoulos J. Making sense of neuroimaging in
[16] Yuvaraj R, Acharya UR, Hagiwara Y. A novel Parkinson's psychiatry. Acta Psychiatr Scand 2008;117(2):100–17.
disease diagnosis index using higher-order spectra [33] Mier W, Mier D. Advantages in functional imaging of the
features in EEG signals. Neural Comput Appl 2018;30 brain. Front Human Neurosci 2015;9:249.
(4):1225–35. [34] Walsh V, Cowey A. Transcranial magnetic stimulation and
[17] Yu H, Lei X, Song Z, Liu C, Wang J. Supervised network- cognitive neuroscience. Nat Rev Neurosci 2000;1(1):73.
based fuzzy learning of EEG signals for Alzheimer's [35] He B, Yang L, Wilke C, Yuan H. Electrophysiological
disease identification. IEEE Trans Fuzzy Syst 2019. imaging of brain activity and connectivity – challenges
[18] Ofner P, Müller-Putz GR. Movement target decoding from and opportunities. IEEE Trans Biomed Eng 2011;58
EEG and the corresponding discriminative sources: a (7):1918–31.
688 biocybernetics and biomedical engineering 40 (2020) 649–690

[36] Babiloni C, Pizzella V, Del Gratta C, Ferretti A, Romani GL. [57] Zhou Z, Li P, Liu J, Dong W. A novel real-time EEG based
Fundamentals of electroencefalography, eye state recognition system. International Conference on
magnetoencefalography, and functional magnetic Communications and Networking in China. Springer; 2018.
resonance imaging. Int Rev Neurobiol 2009;86:67–80. p. 175–83.
[37] Braisby N. Cognitive psychology: a methods companion. [58] Duneja A, Puyalnithi T, Vankadara MV, Chilamkurti N.
Oxford University Press; 2005. Analysis of inter-concept dependencies in disease
[38] Rossini PM, Dal Forno G. Integrated technology for diagnostic cognitive maps using recurrent neural network
evaluation of brain function and neural plasticity. Phys and genetic algorithms in time series clinical data for
Med Rehabil Clin 2004;15(1):263–306. targeted treatment. J Amb Intell Human Comput 2018;1–9.
[39] Michel CM, Murray MM. Towards the utilization of EEG as [59] Reddy TK, Behera L. Online eye state recognition from EEG
a brain imaging tool. NeuroImage 2012;61(2):371–85. data using deep architectures. 2016 IEEE International
[40] Sörnmo L, Laguna P. Bioelectrical signal processing in Conference on Systems, Man, and Cybernetics (SMC). IEEE;
cardiac and neurological applications, vol. 8. Academic 2016. p. 000712–7.
Press; 2005. [60] Andrzejak RG, Lehnertz K, Mormann F, Rieke C, David P,
[41] Adrian ED, Matthews BH. The berger rhythm: potential Elger CE. Indications of nonlinear deterministic and finite-
changes from the occipital lobes in man. Brain 1934;57 dimensional structures in time series of brain electrical
(4):355–85. activity: dependence on recording region and brain state.
[42] Walter WG, Dovey VJ. Electroencephalography in cases of Phys Rev E 2001;64(6):061907.
sub-cortical tumour. J Neurol Neurosurg Psychiatry 1944;7 [61] Zhang T, Chen W, Li M. Classification of inter-ictal and
(3–4):57. Jul. ictal EEGs using multi-basis MODWPT, dimensionality
[43] Campisi P, La Rocca D. Brain waves for automatic reduction algorithms and LS-SVM: a comparative study.
biometric-based user recognition. IEEE Trans Inform Biomed Signal Process Control 2019;47:240–51.
Forensics Secur 2014;9(5):782–800. [62] Acharya UR, Oh SL, Hagiwara Y, Tan JH, Adeli H. Deep
[44] Kumar JS, Bhuvaneswari P. Analysis of convolutional neural network for the automated detection
electroencephalography (EEG) signals and its and diagnosis of seizure using EEG signals. Comput Biol
categorization – a study. Proc Eng 2012;38:2525–36. Med 2018;100:270–8.
[45] Michielli N, Acharya UR, Molinari F. Cascaded lstm [63] Subasi A, Gursoy MI. EEG signal classification using PCA,
recurrent neural network for automated sleep stage ICA, LDA and support vector machines. Expert Syst Appl
classification using single-channel EEG signals. Comput 2010;37(12):8659–66.
Biol Med 2019;106:71–81. [64] Richhariya B, Tanveer M. Eeg signal classification using
[46] Jiang D, Lu Y-n, Yu M, Yuanyuan W. Robust sleep stage universum support vector machine. Expert Syst Appl
classification with single-channel EEG signals using 2018;106:169–82.
multimodal decomposition and hmm-based refinement. [65] Faust O, Acharya UR, Adeli H, Adeli A. Wavelet-based EEG
Expert Syst Appl 2019;121:188–203. processing for computer-aided seizure detection and
[47] Atkinson J, Campos D. Improving BCI-based emotion epilepsy diagnosis. Seizure 2015;26:56–64.
recognition by combining EEG feature selection and kernel [66] Li M, Chen W, Zhang T. Automatic epilepsy detection
classifiers. Expert Syst Appl 2016;47:35–41. using wavelet-based nonlinear analysis and optimized
[48] Jebelli H, Hwang S, Lee S. EEG-based workers' stress SVM. Biocybern Biomed Eng 2016;36(4):708–18.
recognition at construction sites. Autom Constr [67] Ibrahim S, Djemal R, Alsuwailem A.
2018;93:315–24. Electroencephalography (EEG) signal processing for
[49] Bachmann M, Päeske L, Kalev K, Aarma K, Lehtmets A, epilepsy and autism spectrum disorder diagnosis.
Ööpik P, et al. Methods for classifying depression in single Biocybern Biomed Eng 2018;38(1):16–26.
channel EEG using linear and nonlinear signal analysis. [68] Wang X, Gong G, Li N. Automated recognition of epileptic
Comput Methods Programs Biomed 2018;155:11–7. EEG states using a combination of symlet wavelet
[50] Mumtaz W, Vuong PL, Xia L, Malik AS, Rashid RBA. processing, gradient boosting machine, and grid search
Automatic diagnosis of alcohol use disorder using EEG optimizer. Sensors 2019;19(2):219.
features. Knowl-Based Syst 2016;105:48–59. [69] Andrzejak RG, Schindler K, Rummel C. Nonrandomness,
[51] Ruiz-Gómez S, Gómez C, Poza J, Gutiérrez-Tobal G, Tola- nonlinear dependence, and nonstationarity of
Arribas M, Cano M, et al. Automated multiclass electroencephalographic recordings from epilepsy
classification of spontaneous EEG activity in Alzheimer's patients. Phys Rev E 2012;86(4):046206.
disease and mild cognitive impairment. Entropy 2018;20 [70] Bhattacharyya A, Sharma M, Pachori RB, Sircar P, Acharya
(1):35. UR. A novel approach for automated detection of focal EEG
[52] Dauwan M, Linszen MM, Lemstra AW, Scheltens P, Stam signals using empirical wavelet transform. Neural Comput
CJ, Sommer IE. EEG-based neurophysiological indicators of Appl 2018;29(8):47–57.
hallucinations in Alzheimer's disease: comparison with [71] Brinkmann BH, Wagenaar J, Abbot D, Adkins P, Bosshard
dementia with Lewy bodies. Neurobiol Aging 2018;67:75– SC, Chen M, et al. Crowdsourcing reproducible seizure
83. forecasting in human and canine epilepsy. Brain 2016;139
[53] Zhang XL, Begleiter H, Porjesz B, Wang W, Litke A. Event (6):1713–22.
related potentials during object recognition tasks. Brain [72] Sudalaimani C, Sivakumaran N, Elizabeth TT, Rominus VS.
Res Bull 1995;38(6):531–8. Automated detection of the preseizure state in EEG signal
[54] Patidar S, Pachori RB, Upadhyay A, Acharya UR. An using neural networks. Biocybern Biomed Eng 2019;39
integrated alcoholic index using tunable-q wavelet (1):160–75.
transform based features extracted from EEG signals for [73] Shoeb AH. Application of machine learning to epileptic
diagnosis of alcoholism. Appl Soft Comput 2017;50:71–8. seizure onset detection and treatment. Massachusetts
[55] Faust O, Acharya R, Allen AR, Lin C. Analysis of EEG signals Institute of Technology; 2009 [Ph.D. thesis].
during epileptic and alcoholic states using AR modeling [74] Quintero-Rincón A, Pereyra M, D'Giano C, Risk M,
techniques. IRBM 2008;29(1):44–52. Batatia H. Fast statistical model-based classification of
[56] Rösler O, Suendermann D. A first step towards eye state epileptic EEG signals. Biocybern Biomed Eng 2018;38
prediction using EEG. Proc of the AIHLS; 2013. (4):877–89.
 biocybernetics and biomedical engineering 40 (2020) 649–690 689

[75] Alhaddad MJ, Kamel MI, Malibary HM, Alsaggaf EA, Thabit subjects', the rechtschaffen & kales (1968) standard.
K, Dahlwi F, et al. Diagnosis autism by fisher linear Psychiatry Clin Neurosci 2001;55(3):305–10.
discriminant analysis flda via EEG. Int J Bio-Sci Bio- [91] Mourtazaev M, Kemp B, Zwinderman A, Kamphuisen H.
Technol 2012;4(2):45–54. Age and gender affect different characteristics of slow
[76] BCI-competition-II, dataset III. Department of Medical waves in the sleep EEG. Sleep 1995;18(7):557–64.
Informatics, Institute for Biomedical Engineering, [92] Kemp B, Zwinderman AH, Tuk B, Kamphuisen HA, Oberye
University of Technology Graz; 2004, JJ. Analysis of a sleep-dependent neuronal feedback loop:
http://www.bbci.de/competition/ii/. the slow-wave microcontinuity of the EEG. IEEE Trans
[77] Chatterjee R, Maitra T, Islam SH, Hassan MM, Alamri A, Biomed Eng 2000;47(9):1185–94.
Fortino G. A novel machine learning based feature [93] Ichimaru Y, Moody G. Development of the
selection for motor imagery EEG signal classification in polysomnographic database on CD-ROM. Psychiatry Clin
internet of medical things environment. Fut Gen Comput Neurosci 1999;53(2):175–7.
Syst 2019;98:419–34. [94] Correa AG, Orosco L, Laciar E. Automatic detection of
[78] Chatterjee R, Datta A, Sanyal DK. Ensemble learning drowsiness in EEG records based on multimodal analysis.
approach to motor imagery EEG signal classification. Med Eng Phys 2014;36(2):244–9.
Machine Learning in Bio-Signal Analysis and Diagnostic [95] Tripathy R, Acharya UR. Use of features from RR-time
Imaging. Elsevier; 2019. p. 183–208. series and EEG signals for automated classification of sleep
[79] Koelstra S, Muhl C, Soleymani M, Lee J-S, Yazdani A, stages in deep neural network framework. Biocybern
Ebrahimi T, et al. Deap: a database for emotion analysis; Biomed Eng 2018;38(4):890–902.
using physiological signals. IEEE Trans Affect Comput [96] Quan SF, Howard BV, Iber C, Kiley JP, Nieto FJ, O'Connor
2011;3(1):18–31. GT, et al. The sleep heart health study: design, rationale,
[80] Pandey P, Seeja K. Emotional state recognition with EEG and methods. Sleep 1997;20(12):1077–85.
signals using subject independent approach. Data Science [97] Sors A, Bonnet S, Mirek S, Vercueil L, Payen J-F. A
and Big Data Analytics. Springer; 2019. p. 117–24. convolutional neural network for sleep stage scoring from
[81] Pandey P, Seeja K. Subject-independent emotion detection raw single-channel EEG. Biomed Signal Process Control
from EEG signals using deep neural network. International 2018;42:107–14.
Conference on Innovative Computing and [98] Tylová L, Kukal J, Hubata-Vacek V, Vyšata O. Unbiased
Communications. Springer; 2019. p. 41–6. estimation of permutation entropy in EEG analysis for
[82] Blankertz B, Muller K-R, Krusienski DJ, Schalk G, Wolpaw alzheimer's disease classification. Biomed Signal Process
JR, Schlogl A, et al. The BCI competition III: validating Control 2018;39:424–30.
alternative approaches to actual BCI problems. IEEE Trans [99] Hosseinifard B, Moradi MH, Rostami R. Classifying
Neural Syst Rehabil Eng 2006;14(2):153–9. depression patients and normal subjects using machine
[83] Behri M, Subasi A, Qaisar SM. Comparison of machine learning techniques and nonlinear features from EEG
learning methods for two class motor imagery tasks using signal. Comput Methods Programs Biomed 2013;109
EEG in brain–computer interface. 2018 Advances in (3):339–45.
Science and Engineering Technology International [100] Sharma M, Achuth P, Deb D, Puthankattil SD, Acharya UR.
Conferences (ASET). IEEE; 2018. p. 1–5. An automated diagnosis of depression using three-
[84] Chaudhary S, Taran S, Bajaj V, Sengur A. Convolutional channel bandwidth-duration localized wavelet filter bank
neural network based approach towards motor imagery with EEG signals. Cogn Syst Res 2018;52:508–20.
tasks EEG signals classification. IEEE Sens J 2019;19 [101] Ding X, Yue X, Zheng R, Bi C, Li D, Yao G. Classifying major
(12):4494–500. depression patients and healthy controls using EEG, eye
[85] Schalk G, McFarland DJ, Hinterberger T, Birbaumer N, tracking and galvanic skin response data. J Affect Disord
Wolpaw JR. Bci2000: a general-purpose brain-computer 2019;251:156–61.
interface (BCI) system. IEEE Trans Biomed Eng 2004;51 [102] Lotfan S, Shahyad S, Khosrowabadi R, Mohammadi A,
(6):1034–43. Hatef B. Support vector machine classification of brain
[86] Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov states exposed to social stress test using EEG-based brain
PC, Mark RG, et al. Physiobank, physiotoolkit, and network measures. Biocybern Biomed Eng 2019;39(1):199–
physionet: components of a new research resource for 213.
complex physiologic signals. Circulation 2000;101(23): [103] Pereira J, Ofner P, Müller-Putz GR. Goal-directed or
e215–20. aimless? EEG differences during the preparation of a
[87] Zhang X, Yao L, Zhang D, Wang X, Sheng QZ, Gu T. Multi- reach-and-touch task. 2015 37th Annual International
person brain activity recognition via comprehensive EEG Conference of the IEEE Engineering in Medicine and
signal analysis. Proceedings of the 14th EAI International Biology Society (EMBC). IEEE; 2015. p. 1488–91.
Conference on Mobile and Ubiquitous Systems: [104] Ofner P, Müller-Putz GR. Using a noninvasive decoding
Computing, Networking and Services. ACM; 2017. p. 28–37. method to classify rhythmic movement imaginations of
[88] Hussain L, Aziz W, Saeed S, Shah SA, Nadeem MSA, Awan the arm in two planes. IEEE Trans Biomed Eng 2014;62
IA, et al. Complexity analysis of EEG motor movement (3):972–81.
with eye open and close subjects using multiscale [105] Ofner P, Schwarz A, Pereira J, Müller-Putz GR. Movements
permutation entropy (MPE) technique. Biomed Res 2017;28 of the same upper limb can be classified from low-
(16). frequency time-domain EEG signals. Proceedings of the
[89] Tian P, Hu J, Qi J, Ye X, Che D, Ding Y, et al. A hierarchical Sixth International Brain-Computer Interface Meeting: BCI
classification method for automatic sleep scoring using Past, Present, and Future (June 2016); 2016.
multiscale entropy features and proportion information of [106] Akar SA, Kara S, Agambayev S, Bilgiç V. Nonlinear analysis
sleep architecture. Biocybern Biomed Eng 2017;37(2):263–71. of EEGs of patients with major depression during different
[90] S. C. C. O. T. J. S. O. S. R. S. (JSSR), Hori T, Sugita Y, Koga E, emotional states. Comput Biol Med 2015;67:49–60.
Shirakawa S, Inoue K, et al. Proposed supplements and [107] Singh MI, Singh M. Development of a real time emotion
amendments to 'a manual of standardized terminology, classifier based on evoked EEG. Biocybern Biomed Eng
techniques and scoring system for sleep stages of human 2017;37(3):498–509.
690 biocybernetics and biomedical engineering 40 (2020) 649–690

[108] Kotowski K, Stapor K, Leski J, Kotas M. Validation of Speech enhancement techniques for digital hearing aids.
emotiv EPOC+ for extracting ERP correlates of emotional Springer; 2019. p. 63–74.
face processing. Biocybern Biomed Eng 2018;38(4):773–81. [120] Polikar R. The wavelet tutorial – part I, 2nd ed.
[109] Öner M, Hu G. Analyzing one-channel EEG signals for [121] Sovic´ A, Seršic´ D. Signal decomposition methods for
detection of close and open eyes activities. 2013 Second reducing drawbacks of the DWT. Eng Rev 2012;32(2):70–7.
IIAI International Conference on Advanced Applied [122] Cohen I, Raz S, Malah D. Orthonormal shift-invariant
Informatics. IEEE; 2013. p. 318–23. wavelet packet decomposition and representation. Signal
[110] Karamacoska D, Barry RJ, Steiner GZ. Using principal Process 1997;57(3):251–70.
components analysis to examine resting state EEG in [123] Selesnick IW. Wavelet transform with tunable q-factor.
relation to task performance. Psychophysiology 2019;56(5): IEEE Trans Signal Process 2011;59(8):3560–75.
e13327. [124] Huang W, Sun H, Wang W. Resonance-based sparse signal
[111] Gao Z, Li S, Cai Q, Dang W, Yang Y, Mu C, et al. Relative decomposition and its application in mechanical fault
wavelet entropy complex network for improving EEG- diagnosis: a review. Sensors 2017;17(6):1279.
based fatigue driving classification. IEEE Trans Instrum [125] Paraschiv-Ionescu A, Aminian K. Nonlinear analysis of
Meas 2018. physiological time series. Advanced biosignal processing.
[112] Ajami S, Mahnam A, Abootalebi V. Development of a Springer; 2009. p. 307–33.
practical high frequency brain-computer interface based [126] Faust O, Bairy MG. Nonlinear analysis of physiological
on steady-state visual evoked potentials using a single signals: a review. J Mech Med Biol 2012;12(04):1240015.
channel of EEG. Biocybern Biomed Eng 2018;38(1):106–14. [127] Acar E, Schenker C, Levin-Schwartz Y, Calhoun VD, Adali
[113] Abromavicius V, Serackis A. Eye and EEG activity markers T. Unraveling diagnostic biomarkers of schizophrenia
for visual comfort level of images. Biocybern Biomed Eng through structure-revealing fusion of multi-modal
2018;38(4):810–8. neuroimaging data. Front Neurosci 2019;13:416.
[114] Stehlin SA, Nguyen XP, Niemz MH. Eeg with a reduced [128] Lee M-H, Fazli S, Mehnert J, Lee S-W. Subject-dependent
number of electrodes: where to detect and how to improve classification for robust idle state detection using multi-
visually, auditory and somatosensory evoked potentials. modal neuroimaging and data-fusion techniques in BCI.
Biocybern Biomed Eng 2018;38(3):700–7. Pattern Recogn 2015;48(8):2725–37.
[115] Grundlehner B, Mihajlovic´ V. Ambulatory EEG monitoring; [129] Liu S, Liu S, Cai W, Che H, Pujol S, Kikinis R, et al.
2019. Multimodal neuroimaging feature learning for multiclass
[116] Jiang X, Bian G-B, Tian Z. Removal of artifacts from EEG diagnosis of alzheimer's disease. IEEE Trans Biomed Eng
signals: a review. Sensors 2019;19(5):987. 2014;62(4):1132–40.
[117] Ghare PS, Paithane A. Human emotion recognition using [130] Zhu Q, Xu X, Yuan N, Zhang Z, Guan D, Huang S-J, et al.
non linear and non stationary EEG signal. 2016 Latent correlation embedded discriminative multi-modal
International Conference on Automatic Control and data fusion. Signal Process 2020;107466.
Dynamic Optimization Techniques (ICACDOT). IEEE; 2016. [131] Hunyadi B, Van Paesschen W, De Vos M, Van Huffel S.
p. 1013–6. Fusion of electroencephalography and functional
[118] Al-Fahoum AS, Al-Fraihat AA. Methods of EEG signal magnetic resonance imaging to explore epileptic network
features extraction using linear analysis in frequency and activity. 2016 24th European Signal Processing Conference
time-frequency domains. ISRN Neurosc 2014. (EUSIPCO). IEEE; 2016. p. 240–4.
[119] Borisagar KR, Thanki RM, Sedani BS. Fourier transform,
short-time fourier transform, and wavelet transform.