Artificial Intelligence in Medicine 84 (2018) 146–158

Contents lists available at ScienceDirect

Artificial Intelligence in Medicine

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

Random ensemble learning for EEG classification

Mohammad-Parsa Hosseini a,e,∗ , Dario Pompili a , Kost Elisevich b,c ,
Hamid Soltanian-Zadeh d,e
a
Department of Electrical and Computer Engineering, Rutgers University, NJ 08854, United States
b
Division of Neurosurgery, College of Human Medicine, Michigan State University, Grand Rapids 49503, MI, United States
c
Dept. of Clinical Neurosciences, Spectrum Health, Grand Rapids, MI 49503, United States
d
CIPCE, School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran
e
Image Analysis Lab, Depts. of Radiology and Research Administration, Henry Ford Health System, MI 48202, United States

a r t i c l e i n f o a b s t r a c t

Article history: Real-time detection of seizure activity in epilepsy patients is critical in averting seizure activity and
Received 15 June 2017 improving patients’ quality of life. Accurate evaluation, presurgical assessment, seizure prevention, and
Received in revised form emergency alerts all depend on the rapid detection of seizure onset. A new method of feature selection
19 December 2017
and classification for rapid and precise seizure detection is discussed wherein informative components of
Accepted 21 December 2017
electroencephalogram (EEG)-derived data are extracted and an automatic method is presented using infi-
nite independent component analysis (I-ICA) to select independent features. The feature space is divided
Keywords:
into subspaces via random selection and multichannel support vector machines (SVMs) are used to clas-
Brain–computer interface
Distributed computing system
sify these subspaces. The result of each classifier is then combined by majority voting to establish the
Electroencephalogram final output. In addition, a random subspace ensemble using a combination of SVM, multilayer perceptron
Ensemble learning (MLP) neural network and an extended k-nearest neighbors (k-NN), called extended nearest neighbor
Epileptic seizure detection (ENN), is developed for the EEG and electrocorticography (ECoG) big data problem. To evaluate the solu-
Computational neuroscience tion, a benchmark ECoG of eight patients with temporal and extratemporal epilepsy was implemented
in a distributed computing framework as a multitier cloud-computing architecture. Using leave-one-out
cross-validation, the accuracy, sensitivity, specificity, and both false positive and false negative ratios of
the proposed method were found to be 0.97, 0.98, 0.96, 0.04, and 0.02, respectively. Application of the
solution to cases under investigation with ECoG has also been effected to demonstrate its utility.
© 2017 Elsevier B.V. All rights reserved.

Motivation: Epilepsy exists for any patient when two or more regarding localization [3]. The advent of responsive neurostimu-
seizures occur in a 24-h period. The prevalence of epilepsy in lation [4], a closed-loop autonomic therapy, requires automatic
developed countries is 4–10 cases per 1000 people. In developing reliable detection of seizure onset and necessitates refinement in
countries, a much higher prevalence of 14–57 cases per 1000 people autonomic computing methods via brain–computer interface (BCI)
[1] is observed. The most efficient real-time strategy that enables before timely intervention can be provided.
prevention of seizures requires a computational system to detect Vision: Autonomic computing provides a self-management
seizure onset and to promptly institute a therapeutic response. capability that allows intervention. A means of averting seizure
Therapy typically includes administration of antiepileptic drugs, activity through timely intervention early in the course of ictal evo-
surgical resection and/or neurostimulation. Accurate seizure detec- lution is the desired outcome. The closed loop system is created in
tion imposes several demands in the current environment. The two steps – the development of a BCI to detect ictal onset followed
patient may not always have sufficient warning of an impending by an appropriate stimulus to abort the seizure [5]. In this study, we
seizure or may lack the presence of mind to activate a response [2]. focus first on detecting ictal onset with a high sensitivity and speci-
Current EEG analysis involves both automated detection method- ficity. A false positive, in which the system detects a seizure when
ology which is subject to artefact-induced error. Visual perusal there is none, leads to overstimulation. A false negative reflects a
by electroencephalographers, at times, leads to differing opinions failure to detect and, therefore, to abort a seizure.
Challenges: To develop an effective automatic seizure detec-
tion system, salient features of the EEG signal must be extracted
∗ Corresponding author. and distinguished as normal or epileptogenic. Scalp EEG record-

https://doi.org/10.1016/j.artmed.2017.12.004
0933-3657/© 2017 Elsevier B.V. All rights reserved.
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 147

• Developing a multi-tier distributed computing seizure detection

implemented in the cloud as;
• Developing a feature-selection technique using I-ICA to elect
independent features and to infer the number of features auto-
matically from the input data;
• Proposing a random subspace ensemble method with support
vector machines (SVMs) as the base classifiers which fits big data
problems by parallel processing;
• Proposing a random subspace ensemble method using a combi-
nation of different methods as the base classifiers which fits big
data problems by parallel processing;

The proposed seizure detection system is evaluated on a

Fig. 1. Cloud-computing framework enabling ubiquitous healthcare, which is com- benchmark clinical dataset and the performance is presented in
posed of: resource provider, data provider, service requester, arbitrator, and cloud comparison to a previous method as a real-life support mechanism
storage.
for epilepsy management.
Outline: The rest of the article is organized as follows.

ing is subject to signal distortion and directional misinformation

• Section 1: state of the art.
making feature extraction and classification prone to error. Signal
• Section 2: the prevalence of epilepsy.
recording by ECoG directly from the cerebrum via intracranially
• Section 3.1: distributed computing architecture and system
implanted electrodes circumvents much of these difficulties [6]
model.
through higher spatio-temporal resolution of neuronoglial activ-
• Section 3.2: time and frequency feature extraction methods using
ity. Large amounts of real-time data are generated in the process
wavelet, ICA, etc.
leading to a big data problem. For instance, in a five day period, a
• Section 3.3: feature reduction and selection method via I-ICA.
patient may generate 1.6 Gb from several channels of data at high
• Section 3.4: classification of the selected features via random-
recording frequency (i.e., 2000 Hz) [7]. Both safe storage and high
subspace ensemble learning.
computational resources are required for fast real-time processing
• Section 4: validation of our assumptions through simulations on
while fluctuations in amplitude and frequency create challenges
clinical dataset, improvement over existing solutions.
for feature extraction and classification. Many methods exist for
• Section 5: final discussion and future work.
seizure detection but further improvements are required to provide
reliable detection preictally.
Approach: In order to create an accurate seizure-detection sys- 1. State of the art
tem that is useful, real-time signal processing, machine-learning,
and brain-state prediction on large data sets collected from Several analytic techniques have been applied toward assessing
large populations over extended time periods are required [8,9]. electrographic behavior. Lacunarity and Bayesian linear discrimi-
Next-generation BCI-EEG systems will require connection with nant analysis in intracranial EEG has been used for epileptic seizure
high-performance computing (HPC) servers through the Internet detection [12]. A wavelet-based algorithm to examine how dif-
to adapt prediction models to incoming streaming (i.e., in-transit) ferent frequency ranges in ECoG fluctuate from background has
data [10,11]. We propose a cloud computing framework that auto- provided further insight into varying patterns of detection [13]. An
matically detects epileptic seizures by providing the flexibility to algorithm based on artificial neural networks has allowed for clas-
access databases, computational resources, and storage over the sification of EEG signals into healthy, ictal, and interictal patterns
Internet. This expanding area of IT service offers ubiquitous access [14]. Automatic detection of tonic–clonic seizures, based on surface
with the potential to increase agility at lower operational costs. The electromyography (sEMG) has also been developed to better dis-
platform makes decisions based on comparing extracted EEG pat- tinguish actual clinical seizure activity [15]. The power spectrum
terns with cloud data. Our seizure-detection framework, as shown in EEG segments was developed using an adaptive thresholding
in Fig. 1, is composed of: (i) a resource provider, (ii) a data provider, technique based on short time Fourier transform (STFT) for seizure
(iii) a service requester, (iv) an arbitrator, and (v) cloud storage. classification [16]. Pachori et al. [17] presented a new method
Resource and data providers enable computational resources and for classification of ictal and non-ictal EEG using empirical mode
training data, respectively. A service requester organizes requests decomposition (EMD) and the second-order difference plot (SODP).
for the framework and an arbitrator processes the request, deter- Spectral features using tunable-Q factor wavelet transform (TQWT)
mines the set of service providers, and distributes the workload and bagging are developed for seizure detection in the following
tasks among the resource providers. Both original and processed paper [18].
data is saved through cloud storage. High data-rate telecommu- Phase-Slope Index (PSI) of directed influence was applied to
nication technology and cloud-based computational methodology a multichannel ECoG for seizure detection [19] to good effect.
lend themselves well to this real-time big data processing applica- Moreover, analyses in time and frequency domains have been
tion. developed to find energy distributions in the time-frequency plane
Contributions: A wavelet transform is applied to extract brain [20] indicating effective computational application. Amplitude fea-
signals and divide the EEG into different frequency bands. Multiple tures such as relative average, coefficient of variation, and duration
features from different domains are extracted for classification. A have been used for detection and monitoring [21]. Use of SVM to
new technique for feature selection, based on infinite independent distinguish EEG signals (i.e., epileptic vs nonepileptic) based on
component analysis (I-ICA), is developed. We also propose a new input features has also been shown [22].
classification method, based on ensemble learning and random- A State-Space Model with Cauchy observation noise (SSMC) was
ness for parallel processing, decreasing the false detection rate and developed to detect seizure onset in a long-term EEG monitor-
increasing sensitivity. To summarize, our contributions in neuroin- ing system [23]. A simple rule applied to the acceleration norm
formatics include: entropy HnA has been shown to enhance performance compared
148 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

to other classifiers trained on other feature sets [24]. An eight- tic pathway for the management of epilepsy worldwide is much
channel EEG recorder and seizure detector has also been developed needed and is being guided by the International League Against
[25] and an implantable low-power integrated circuit for real-time Epilepsy. We see the current technology as a step in this process
seizure detection proposed [26]. A nonlinear SVM seizure classifi- that must be considered for the future of all patients.
cation with 8-channel EEG data acquisition and storage has been
suggested [27]. A seizure detection system capable of both iden- 3. Proposed work
tifying electrographic seizure onset and triggering focal blockade
of seizure development implemented [28]. Quantification of the A distributed computing seizure detection system from EEG
sharpness of waveforms has allowed greater refinement in seizure and ECoG signals is developed for real-time usage. A concep-
detection [29]. A seizure model for a priori known seizure activity tual architecture of the proposed BCI as a multi-tier distributed
has been implemented by using optimal null filters as a building bioinformatics system is presented in Section 3.1. In Section 3.2,
block for the detection of similar events [30]. There are some other preprocessing and feature extraction is presented while in Section
methods which use artificial intelligence and machine learning 3.3, the number of features are reduced by an extended version
techniques for seizure detection [31–33]. of ICA called I-ICA to increase classification accuracy by removal
Overall, existing methods have been developed for local storage of less effective features and also to decrease computational time.
and processing with little consideration for the big data prob- In Section 3.4, a classification method using a random selection of
lem. Accurate seizure detection requires long-term multichannel feature subsets and ensemble learning is proposed. The proposed
surveillance and existing works fall short in exploiting fully the method is compatible with big and high-dimensional data such as
considerable volume of data available. A cloud-based system allows ECoG.
for the storage and analysis of this level of demand with a timely
responsiveness available in real-time. However, applications of
cloud-based systems in medicine are still in their infancy and only 3.1. Distributed computing system
a few studies are available. A cloud-based system using the MapRe-
duce parallel-programming system has been developed for analysis The electroencephalographic assessment of an individual’s
epileptogenicity begins first with a standard scalp recording over
of cardiac activity that generates large biosignal volumes [34]. Our
group developed the usage of deep learning models in the cloud an extended duration (i.e., 5–7 days) to gather both interictal and
ictal features that arise over time in an epilepsy monitoring unit
for analyzing the prediction of seizure in big EEG dataset as BCI
systems [35,36]. An on-line BCI-EEG system uses multi-tier fog and (EMU) as shown in Fig. 2. This is commonly referred to as a phase
I study. Consistency in channel recording and in patterning of dis-
cloud computing, semantic-linked data search, and adaptive classi-
fication models [37]. A scalable analysis tool for brain imaging has charges, particularly in the preictal period of an epileptic event,
been used by combining a MapReduce framework with machine- serves to establish certainty in the uniqueness of an ictal origin.
learning algorithms in the cloud [38]. Because of data overfitting, Source modeling by a variety of techniques is used to ascertain the
low sensitivity, computational time, and a high error rate of existing likelihood of a specific locale or, at least, nodal sources of activ-
techniques in feature selection and classification, the application ity in an epileptogenic network. The temporal accumulation of this
was not judged appropriate for real-time implementation. In con- activity ultimately approaches a certain threshold at which a deter-
trast, we propose a new artificial intelligence system for EEG/ECoG mination may be made to terminate recording with one of the
feature selection and classification that exploits the advantages of following outcomes:
cloud computing and simultaneously increases detection accuracy
while decreasing error rate. • Interictal and ictal EEG findings are sufficiently consistent and in
agreement with neuroimaging metrics to warrant either resec-
tion of the area of epileptogenicity or the permanent placement
2. Prevalence of epilepsy of electrodes strategically targeting sites within a defined epilep-
togenic network to provide reliable surveillance of ictal activity
The prevalence of epilepsy globally, according to the World and to execute a signal from an implanted microprocessor that
Health Organization, amounts to about 50 million people with may abort the event.
an almost 80% preponderance in low and middle-income coun- • EEG and neuroimaging features do not correlate sufficiently or
tries where upwards of 75% of patients may not receive adequate remain consistent of themselves but establish some certainty of
first-line treatment of the condition. Endemic conditions certainly an epileptogenic focality to warrant further investigation to bet-
play a major role in this maldistribution of cases (i.e., insufficient ter understand whether a distinct and sufficiently constrained
regional health services, birth injury, neurocysticercosis, road acci- epileptogenic network may be defined. In this circumstance,
dents, etc.). In absolute terms, the proportion of people with active ECoG with strategically directed intracranial electrodes targeting
epilepsy in high income countries ranges 4–10/1000 although in putative sites of epileptogenic behavior (i.e., phase II) is contem-
the low and middle income countries, the estimate appears to be plated in the EMU setting.
7–14/1000 where the annual incidence of cases could be upwards • There is sufficient reason to indicate a generalized pattern of
of 100/100,000 [39]. Our paper addresses an application of a tech- epilepsy or a multifocality that precludes further consideration
nology that affords treatment of medically intractable epilepsy for investigation.
conditions where even conventional surgical approaches cannot be
effectively implemented with a suitable result. Its immediate use Over the relatively short duration of a phase I study, a large
will be in relatively high income society where this technology is amount of scalp EEG data is accumulated and, again, necessitates
more available and where first-line treatment has already failed. In considerable archiving and ongoing feature extraction to arrive at
low and middle-income countries, such first-line treatment must a possible solution within a given interval of time. Correlation then
still be brought up to a sufficient level and this is a matter for global with neuroimaging affords an added measure of certainty of suc-
healthcare economics to address. However, the universality of the cess in identifying the epileptogenic network sufficiently well for
described technology mentioned in this paper can ultimately be definitive therapy. A distributed computing system can be used to
realized with the unlimited resources of the global internet and enable a real-time detection of epileptic activity in the EMU and
cloud computing. A proper clinical investigational and therapeu- to incorporate, simultaneously, a considerable volume of quantita-
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 149

Fig. 2. Epilepsy monitoring unit (EMU). Electroencephalographers inspect both ictal and interictal activities typically by visual scanning of computerized records and develop
interpretations of the location of disturbances (i.e., sites of epileptogenicity) which are then debated in conference.

Fig. 3. Conceptual architecture of the proposed EEG seizure detection system. It is developed in a multi-tier distributed computing infrastructure and a semantic linked data
superstructure. The proposed architecture enables two operation scenarios: (1) big data analysis using a cloud computing paradigm; (2) interactive and adaptive prediction
using real-time brain state and relevant data sets for training and refining brain state prediction.

tive neuroimaging analytics to provide a reasoned assessment of Algorithm 1. Finally, the third tier is developed as HPC clusters such
the distribution of epileptogenicity. as cloud servers for delivering plenty of computing power, storage
In this paper, we propose to enhance the traditional health care capacity, and communication bandwidth to offload the computing
delivery system with the provision of a computational intelligence burden of the second tier.
as a pervasive computing system for seizure detection. As shown State-of-the-art techniques for communicating between each
in Fig. 3 a seizure detection system is developed as a multi-tier tier involve the Message Queuing Telemetry Transport (QMTT) pro-
distributed computing structure via mobile device cloud (MDC) and tocol for interacting between EEG and MDC and the RESTful Web
cloud computing. In this framework, MDC performs tasks in parallel Service for interacting between MDC and the Cloud [62,63]. Data is
while sharing the workload among multiple nearby mobile devices sent from the first to the second tier using Bluetooth 4.0 protocol
[40]. and IEEE 802.11n low-power Wi-Fi technology. A transfer rate of up
The first tier involves interfaces between human and IT tech- to 24 Mbps is supported. With cloud computing, EEG data transport
nology derived from EEG, ECoG, and smart phones. The second tier latency through the Internet core runs between 200 and 500 ms
consists of the MDC as an ad hoc conglomerate of IT devices for com- [37]. In the proposed distributed computing system, the MDC as the
putational purposes such as notebooks and home-gateways. Each second tier between cloud and brain sensors provides two advan-
MDC server operates dually as a data hub and a signal processor tages. First, it delivers subsecond real-time responses with minimal
[59]. Fig. 4 represents the proposed workflow for signal processing communication overhead. Second, it reduces the amount of traf-
as an extraction of time and frequency features, feature selection, fic between local area networks and the Internet. In point-of-fact,
random subspacing, and ensemble classification. The extracted fea- most of the data is processed and stored in the second tier and only
tures can also be sent to the next tier for further processing and big data, requiring high computational resources or storage, will be
archiving [60,61]. To find the computational complexity of the over- uploaded to the cloud.
all algorithm, the pseudocode of the proposed system is provided in
150 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

Fig. 5. Wavelet tree showing synthesize signal (s), detail coefficients (d), and
approximation coefficients (a).

may interfere with the information content of ECoG (i.e., intrinsic

physiologic signals, environmental noise, and adjacent electrode
interference). Some sources of noise and artifact may be removed
by filtering. A fourth-order Butterworth bandpass filter may be
applied to eliminate unwanted frequencies. Also, a notch filter may
be applied to remove the interference from the power line. In addi-
tion, forward and backward filtering may be applied to cancel phase
distortion.
Wavelet transforms are used to extract epileptic spike activ-
ity and to capture the rhythmicity of epileptogenic behavior. Also,
transient features can be captured by wavelets and localized in
Fig. 4. Workflow of the proposed method consisting of noise and artifact reduction, both time and frequency domains [47,48]. Fig. 5 demonstrates the
wavelet, time/frequency-based feature extraction, infinite ICA, ensemble learning synthesis of a signal (s), approximation coefficients (a), and detail
by random subspace, ensemble SVM, and majority voting.
coefficients (d) with six levels of decomposition. Finally, seizure-
related features are extracted using the outputs of the wavelet
Algorithm 1. Proposed seizure detection system. transform and filtering. Several time- and frequency-related fea-
tures are extracted including fast Fourier transform (FFT), mobility,
complexity, variance, skewness, kurtosis, average spectral power,
line length, absolute value sum, maximal and minimal values,
energy, entropy, correlation coefficients, mean, fractal dimension,
frequency band power, zero crossing, peak amplitude. Some of the
extracted features are defined as follows:
Hurst Exponent: The EEG/ECoG signals include epileptogenic
disturbances consisting of a variety of observed signals and inde-
pendent sources. Independent component analysis (ICA) is used
for decomposing mixed EEG signals into a set of independent com-
ponents. Then, the Hurst exponent is extracted for finding these
components within the epileptogenic disturbance. For this pur-
3.2. Feature extraction pose, the range of 0.25–0.45 mV is considered epileptogenic [49].
Entropy: The degree of uncertainty is measured with entropy of
Certain sources of noise and artifact may interfere with infor- a signal s and is given as,
mation content in the ECoG (i.e., intrinsic physiologic signals,
environmental noise and adjacent electrode interference) [41,42].

Q

H(s) = − si ln(si ), (1)

Real-time analysis of EEG dataset is susceptible to an enormous
i=1
amount of artifacts, including motion artifacts, which will degrade
the classifier’s performance. Recent papers have proposed spe- where i is the index of the samples and Q is the total number of
cific algorithms for identifying and removing such artifacts [43,44]. samples per signal segment. The disorder of the EEG segment is
The level of artifacts may vary by experimental setup, quality of considered a way to find a seizure.
data acquisition system, and speed of motion. The efficacy of the Amplitude: In general, the amplitude of the rhythmic epilepti-
proposed methods has been questioned recently [45]. Therefore, genic component is greater than that in the normal state. Therefore,
removal of motion artifact is still an important open problem for we extract features based on signal amplitude, including relative
developing real-time systems which should be addressed in the average, coefficient of variation, and duration. To have a normal-
future. In this study, we have used a standard dataset which is free ized measurement, each ECoG epoch is divided by the mean value
of motion artifacts. We have applied the standard filtering methods of the ECoG segment.
suggested for the removal of unwanted frequencies and interfer- Energy: The energy of EEG segments is extracted as an indicator
ence from the power lines. of seizure activity.
When there are motion artifacts, the methods proposed in
[43,44] can be applied. Also, moving average and stride time warp-

Q

E(s) = |si |2 . (2)

ing have been used for motion artifact cleaning [46]. However,
i=1
the efficacy of these methods has been questioned [45]. Therefore,
removal of motion artifacts may be considered as an open problem. A higher energy manifestation is translated into a higher probability
In addition to motion artifacts, other sources of noise and artifact of an impending seizure. Skewness: The third central moment of
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 151

the amplitude histogram is defined as skewness. It is the degree of is applied. In this case, the conditional distribution of one parame-
deviation from the symmetry of a Gaussian distribution, as. ter given all others are sampled by Baye’s rule [50]. The result is a
Q (si −s̄)3
piecewise Gaussian distribution which is defined by,
i=1 N 
SK(s) =  3/2 , (3) N(xki ; − ,  2 ) if xki > 0
Q (si −s̄)2
1p(xki |G, x−ki , yi , zi ) = (6)
i=1 Q −1
N(xki ; + ,  2 ) if xki < 0,
where s̄ is the mean value. For EEG signals with a symmetrical
where
distribution, the skewness has a nonzero value and indicates the
◦
presence of monophasic events. gkT eki ± e2 e2 ◦
± = , 2 = , eki = (eki |zki = 0), (7)
Algorithm 2. Feature selection by I-ICA. gkT gk gkT gk

and gk is the kth column of G.

Algorithm 2 shows the proposed technique for feature selec-
tion. The proposed method is a solution for feature selection in
the big data problem and, by selecting more informative features,
it makes classification less complex, a distinct need for real-time
computation.

Algorithm 3. Random subspace ensemble classification.

3.3. Feature selection via infinite ICA

Given a feature space y ∈ RN , feature-selection methods find a

mapping x = f (y) : RN → RM (M < N) such that the transformed fea-
ture vector x ∈ RM preserves the most information of y. To elect
appropriate and more informative feature sets, different feature-
selection methods have been proposed in the literature such as
principal component analysis (PCA) and ICA. These statistical fea-
ture extractors are used to obtain uncorrelated and independent
features, respectively. 3.4. Classification via random subspace ensemble learning
A linear mapping with a lower dimension with maximum vari-
ance is obtained by PCA. Another technique method for dimension Most existing methods have been developed to use only a small
reduction is ICA. It transforms the observed data y with a lin- training inset which does not measure up to the inherent high
ear transformation G into independent components x as y = Gx + e, dimensionality of the problem. This makes the features-to-instance
where e denotes the Gaussian noise. These methods have been used ratio very high. Using a small training dataset with high dimen-
for feature selection with respect to dimension reduction needs. sionality to train a classifier leads to low classification performance
However, the dimensionality of the new-feature vector cannot be because of overfitting [51]. We propose a new classification tech-
inferred and the number of new features need to be determined nique called random subspace ensemble method with SVM as the
in advance. In this paper, we proposed using an extension of ICA, base classifiers to address this problem and reduce the computa-
called infinite ICA (I-ICA) [50], for solving the above problem and tional load, which is needed for real-time and pervasive analysis
inferring the number of independent features directly from the (see Fig. 6).
input data. For masking a hidden source x, a binary vector z is deter- Ensemble learning techniques have been considered as an alter-
mined, where its elements show activity of kth hidden source for native to single classifiers to avoid overfitting caused by high
the ith data point. dimensionality and small size data. These techniques provide a
collective decision and improve overall learning performance by
Y = G[X  Z] + E, (4)
combining weak classifiers and aggregate multiple learning algo-
where Y, X, Z, E denote the concatenation of {yi }Ni=1
, {xi }N
i=1
, {zi }N
i=1
, rithms such as Bagging and AdaBoost [52]. However, in contrast
and {ei }N
i=1
, respectively; and  denotes the element wise multipli- to previous techniques, the proposed Random subspace ensemble
cation. method uses a random sample of features, as the name implies,
Since Z has infinitely many rows, an infinite number of hidden rather than using all features. In this case, all the features are
sources can be obtained. For N data points and K hidden sources, selected randomly and are assigned to a classifier as described in
the distribution of matrix Z is defined by, Section 3.4.1. In the extended version of this classifier, presented
in Section 3.4.1, all features are used for different classifiers.

K 
N

K
p(Z|1 , . . ., K ) = P(zki |k ) = kmk (1 − k )N−mk , (5)
3.4.1. Random ensemble learning
k=1 i=1 k=1
In the proposed method, decision rules in each SVM classifier
where zki indicates activity of kth source for ith sample using prob- are learned with a randomly-selected feature subspace by all of the
N
ability of k and mk = z shows the total number of active
i=1 ki
training samples. The M dimensional output of I-ICA is randomly
sources. To define E, a Gaussian noise has been considered with divided into P subspaces denoted by r1 , . . ., rP . Then, SVMs, as effi-
variance e2 . At the end, an inference is used for inferring X hidden cient classifiers, are applied on each of the subspaces for classifying
sources from Y observed data with the G mixing matrix (which has the input data as epileptic and nonepileptic. Using nonlinear SVMs,
Z active sources). For sampling elements with zki = 1 Gibbs sampling a nonlinear function (r) is used to map input space into a higher-
152 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

Algorithm 4. Random subspace ensemble classification with

combination of classifiers.

Fig. 6. The proposed random subspace ensemble. The feature space is y ∈ RN ; the
feature selection method (I-ICA) finds a mapping x = f(y): RN → RM (M < N). Ran-
dom subspace utilizes random subsets in the feature space and makes P subspaces 3.4.2. Random ensemble learning with combination of classifiers
denoted by r1 , . . ., rP . Next, SVM’s classifiers are implemented on each subspace
To improve the detection rate in the training and classification
to classify the input data as nonepileptic or epileptic. Finally, the output with the
highest number of votes is chosen by majority voting (MV). phase, a combination of additional classification methods is con-
sidered in our method. For this purpose, multi layer perception
(MLP) neural network and an extended k-nearest neighbors (k-NN)
algorithm called extended nearest neighbor (ENN) [54] are used as
dimensional feature space which corresponds to a linear surface in shown in Fig. 7. For MLP classifiers, two-layer feedforward neu-
the feature space. The decision function v is defined by, ral networks with one hidden layer and sigmoids, as activation
function, are used. Different optimization schemes were applied
for the training of the network and backpropagation algorithms
v = sgn(w.(r) − b), (8) used based on good results. A mathematical hard limit function is
used to code the output of the first layer to 1 and −1 representing
where, b is a bias and sgn is the sign of a real number. Then, the clas- ictal and nonictal detection. A backpropagation algorithm is used to
sification problem is defined as distinguishing normal features from find the derivative of the loss with respect to network parameters
epileptogenic features. This is done by identifying an hyperplane for weight optimization. Assuming error function, E, the gradient
2 component for each weight is found by,
w that divides the features of two classes. This corresponds to the
problem of maximizing a quadratic function of defined variables
subject to their linear constraints. Such quadratic programming ∂E 
N−n 
N−n
∂E ∂xij
l 
N−n 
N−n
∂E l−1
= = y (11)
optimization is solved by, ∂ωab ∂xijl ∂ωab ∂xijl (i+a)(j+b)
i=0 i=0 i=0 i=0

In the ENN, the nearest neighbors of the test sample to the

w(˛) = 0.5w2 − ˛[o(w.(r) − b) − 1], (9)
classes and each class’ nearest neighbors to the test samples are
considered thus improving classification accuracy compared to the
where, ␣ is defined as a set of Lagrange multipliers. Since, most of k-NN method. In this classifier, the unknown sample x is iteratively
the time, EEG training sets are not separable using a linear function, assigned to each possible class j, j = 1, 2 and the class membership
a Gaussian Radial Basis Function (GRBF) kernel is used to minimize is predicted by,
the classification error [53]. This kernel function is defined as, ⎧ ⎫
⎨ nj + ki − kT 
N
n
j⎬
i i i
v = argmaxj ∈ 1,2 − (12)
k(ri , rj ) = T (ri ).(rj ) = exp{ri − rj 22 /2 2 }, (10) ⎩ (ni + 1)k ni k ⎭
i=j i=
/ j

where ri and rj are two feature vectors, and ri − rj 22 is the squared where k is the number of the nearest neighbors (defined by user),
Euclidean distance between ri and rj . Majority voting (MV) is ki is the number of the nearest neighbors of the test sample x from
j
applied to find the output with the highest number of votes and class i, ni is the number of training data for class i, ni indicates
as the final output of the overall system. A cross-validation tech- the variation of the k nearest neighbors for class i when the test
nique is used for increasing generality since the strength of a sample x is supposed to be class j, and Ti indicates the general-
correlation does not necessarily predict the outcome of a new ized classwise statistic of original class i. The pseudocode of the
observation. Leave-one-out as an exhaustive cross-validation is combination of weak classifiers in the proposed random subspace
applied for directly estimating the predictive accuracy of a par- ensemble learning is shown in Algorithm 4.
ticular statistical model. Therefore, the model is fitted to subsets
of data and the accuracy of the model is obtained with the held- 4. Performance evaluation
out sample. Algorithm 3 presents the pseudocode of the proposed
classification method. Dataset: To evaluate the proposed methods, we have used the
Following is the algorithm of Ensemble classification with com- clinical ECoG dataset of eight epileptic patients, including 104
bination of classifiers interictal (normal) segments and 104 ictal (seizure) segments. The
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 153

Fig. 7. Random subspace ensemble with different base classifiers using I-ICA for feature selection. A combination of classifiers are used on each subspace to classify the input
data as nonepileptic or epileptic. Using majority voting (MV), the output with the highest number of votes is chosen.

Fig. 8. Three-dimensional rendering of wide array left hemispheric surface electrode coverage undertaken for phase II (ECoG) recording of a patient. A closed loop device
(RNS; Neuropace) was placed subsequently with two electrode arrays targeting newly established sites of epileptogenicity.

database was developed by the University of Pennsylvania and the locations were used for recording of the ECoG database. Ictal and
Mayo Clinic, and also sponsored by the American Epilepsy Soci- interictal segments of a sampling rate from 500 to 5000 Hz have
ety [55]. Depth electrodes implanted along the anterior–posterior been used for training and testing. The ictal data segments were
axis of the hippocampus and in subdural electrode grids in various organized into 1 s EEG clip sections. The average electrographic
154 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

Fig. 9. Synthesized signal for 1-s seizure data segment: S, six levels of approximation coefficient: a1–a6 (left), six levels of detail coefficients: d1–d6 (right). The x-axis
represents samples and the y-axis shows the signal in ␮V. Scale 4–6 of detail coefficients representing seizure activity.

duration for each subject is covered in the interictal data segments for ECoG data. The detail coefficients were investigated to find
with a limit of no less than 1 h prior to or after a seizure. In addition, the frequency range of 3–25 Hz, the range in which ictal activity
anonymized epilepsy cases which had undergone RNS implanta- on ECoG occurred commonly. First, the sampling frequency of the
tion (Neuropace) following Phase II-derived definition of site(s) data (500 Hz) was considered and the maximum frequency of data
of epileptogenicity were used (see Fig. 8) to further develop our obtained at 250 Hz by the Nyqvist criteria. Finally, the frequency
methods. range of 3–30 Hz was covered in scales of 4, 5, and 6 by coefficient
Testbed: We have developed a proof-of-concept prototype of representation in each scale.
the computational EEG-based seizure detection model in the Cen- Fig. 9 shows the six levels of the approximation coefficient
ter of Cloud and Autonomic Computing (CAC), Rutgers University. A (a1–a6) and the corresponding detail coefficients (d1–d6) in which
clinical dataset of epilepsy is used in the first tier. In the second tier, ictal activity was covered in the d4–d6 scales. Some EEG patterns
an HP laptop with Intel i5 processor, 8 GB RAM, and a 4.4 Ah bat- are morphologically epileptiform but are not associated with an
tery is used. A supercluster of computers hosted by Amazon Elastic epileptic event or other neurological abnormality. These patterns
Compute Cloud (EC2) is used for the third tier. do not have any clinical significance and are termed “epileptiform
Transformation of the data format to different hierarchical data normal variants”. These variants are one of the major reasons for
models such as Multiscale Electrophysiology File (MEF), Standard false seizure detection in automatic systems and the recognition of
JavaScript Object Notation (JSON), JSON, Neurophysiology Data these patterns is important to avoid misinterpretation. Small sharp
Translation Format (NDF), and BiosignalML for representation and spikes, phantom waves, wicket spikes, and paroxysmal rhythmic
encoding were analyzed in order to organize and coordinate large discharges are some examples [56]. After analyzing these patterns,
quantities of numerical ECoG data [7]. The first step was to atten- we remove the specific features that resemble those patterns (see
uate various sources of noise and artifact via filtration, where a Fig. 10) using frequency filtering followed by block scaling of ampli-
fourth-order Butterworth bandpass filter (0.5–150 Hz) was used to tude and slope [57]. Table 1 provides the results of a standard
cut frequencies. A notch filter set at 60 Hz was applied for removal ANOVA analysis [1], while Figs. 11 and 12 show some metrics by
of unwanted frequencies of the oscillator. Then phase distortion via quartile for the extracted features in nonepileptic and epileptic sub-
backward- and forward-filtering was canceled. Based on input sig- jects. The large F-statistic and small value of p in Table 1 correspond
nal and application, the necessary wavelet and decomposition level to a large difference in the centerlines of the box plots between
were chosen. We used Daubechies 4 (Db4) to determine details Figs. 11 and 12.
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 155

Fig. 10. EEG segments of 32 channels showing epileptiform normal variants in a

few channels (top) and its stack EEG segments (bottom) in highlighted areas. The
horizontal line is time [s] and the vertical line is amplitude [␮V].

Table 1
Source of variability, sum of squares (SS) for each source, degrees of freedom (df) for
a source, mean square (MS) value as the ratio SS/df, F-statistic or ratio of the mean
squares, and p-value obtained from the cumulative distribution function (CDF) of F
are presented as the standard ANOVA. Fig. 12. Graphical depiction for nine of the extracted features for ictal segments.
Maximum of amplitude which is multiplied by scale factor 100, variance, energy
Group Source SS df MS F-statistic p-Value
which is multiplied by scale factor 0.01, skewness, power, sum of absolute value
Normal Columns 2.89e+09 8 3.62e+08 11.45 1.85e−14 which is multiplied by scale factor 100, FFT, and mean value which is multiplied by
Error 1.11e+10 351 3.16e+07 scale factor 0.01 are shown, respectively. Outliers are plotted by plus signs.
Total 1.39e+10 359

Seizure Columns 1.96e+10 8 2.45e+09 21.92 4.64e−27

Error 3.92e+10 351 1.11e+08
Total 5.88e+10 359

Fig. 13. Linear classification in one the subset of features (P1, P2). The red signs are
extracted features from ictal segments and the green signs are extracted features
from nonictal segments. The blue sign is the outcome of a test segment which is
correctly classified as nonictal. (For interpretation of the references to color in this
legend, the reader is referred to the web version of the article.)

Fig. 11. Graphical depiction for nine of the extracted features for nonictal segments.
Maximum of amplitude which is multiplied by scale factor 100, variance, energy
Table 2
which is multiplied by scaled factor 0.01, skewness, power, sum of absolute value
Accuracy, sensitivity, specificity, FPR, and FNR [64,65] for (1) proposed method with
which is multiplied by scale factor 100, FFT, and mean value which is multiplied by
combination classifiers, (2) proposed method with nonlinear SVM, (3) proposed
scale factor 0.01 are shown, respectively. Outliers are plotted by plus signs.
method with linear SVM, (4) nonlinear SVM, (5) ENN, (6) linear SVM, and (7) MLP.

Methods Accuracy Sensitivity Specificity FPR FNR

We use 64 features as the input parameters of SVMs in the pro-
Ran Sub com of Class 0.97 0.98 0.96 0.04 0.02
posed ensemble method. We find five subsets of features by random
Ran Sub NL SVM 0.95 0.96 0.94 0.06 0.04
selection in each subset. Then, features in each subset are used to Ran Sub L SVM 0.91 0.92 0.90 0.10 0.08
train an SVM (Figs. 13 and 14). The random combination of multiple Non-linear SVM 0.85 0.85 0.87 0.13 0.15
features provides a more accurate classification compared to using ENN 0.85 0.84 0.83 0.17 0.15
all the features. The results of the proposed method with nonlinear Linear SVM 0.84 0.83 0.85 0.15 0.17
MLP neural network 0.82 0.81 0.83 0.17 0.19
and linear base classifiers are compared to ENN, single SVM, and
MLP neural network in Table 2.
156 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

correlated with EEG data to render an initial assumption of the site

of epileptogenicity and these may be reported with varying degrees
of certainty. Based upon this preliminary assessment, definitive
therapy may be decided in the form of resective surgery or entirely
discounted on the basis of multifocality suggesting greater than
two sites of independent epileptogenicity. When uncertainty exists
regarding the location of a particular focality or a need exists to
establish the elequence of cerebral function in the vicinity of a puta-
tive site, than intracranial electrographic investigation is required
in the form of extraoperative ECoG. This requires the intracra-
nial placement of surface and/or depth electrode arrays in specific
locations of the brain to better understand the distribution of the
epileptogenic network (phase II) and a further admission to the
EMU. The results will often declare the approach to be taken ther-
apeutically [58].
In the circumstance where permanent implantation of electrode
arrays is necessary for the longterm surveillance of a defined epilep-
togenic network and for stimulus delivery upon ictal onset, a closed
loop system is implanted (i.e., RNS; Neuropace) to provide ongoing
Fig. 14. Results of nonlinear classification for one subset of features (P1, P2), similar management of the epilepsy. A period of data accrual occurs at the
to Fig. 13. outset following implantation in which preictal patterns are recog-
nized and metrics are applied to serve as means by which decisions
are rendered to initiate stimulus delivery in order to abort an
impending seizure. As implantable electrocorticographic surveil-
lance becomes increasingly complex through greater monitoring
capacity, a better appreciation of the behavior of an epileptogenic
network arises as does a need for a continuous iterative approach to
define this pathophysiologic entity sufficiently to intelligently miti-
gate its actions. The network may be distributed broadly or confined
to a more immediate location in the central nervous system. Its
nodal sites are likely to display variable patterns of activity depen-
dent upon their connectivity and other circumstances of which we
are less aware. A means of autonomously monitoring this activity to
establish the criteria for preictal pattern recognition and creating a
suitable set of stimulus delivery parameters that would effectively
eliminate ictal progression through an iterative approach would
provide a more timely course of initiation. In the longterm, con-
tinued data accrual would monitor any deviation of electrographic
activity from that previously established to provide warning of a
Fig. 15. Round trip time (RTT) between Amazon EC2 servers (third tier) and a local mutating epileptogenicity expressing itself in a different location.
machine (first tier) for EEG clips. At any time, the lowest RTT is observed for the A means of therapeutic interaction with an area of epilepto-
server located in Virginia (the closest to our location). genicity, that does not entail removal of a portion of the brain,
first requires adequate detection of ictal onset. The use of com-
To illustrate the delay associated with different tiers, e.g., at the puters to help physicians in the acquisition, management, storage,
first layer and at the cloud layer, we conducted experiments on the and reporting of brain (i.e., EEG) signals is well established. To this
round trip time (RTT) measured on a 64B EEG segments as shown end, there are computer-aided detection applications that use a
in Fig. 15. BCI. In order for a BCI system to work effectively, computational
algorithms must reliably identify periods of increased probabil-
5. Discussion and future work ity of an ictal occurrence in order to abort its development. Such
ictal periods may be of variable duration and may not afford suit-
To better understand the task at hand, it is first useful to review able latency to provide current methodologies with sufficient time
the current investigational aspects involved in elucidating the for signal deployment to achieve control in all circumstances. The
patient’s epilepsy. In those patients declared to have an epilepto- development of an automated method that delivers on such short
genicity that can be further investigated to establish its location notice would optimize seizure control and bring about an improved
in the brain, a number of standard neuroimaging, functional and quality of life.
electroencephalographic studies are undertaken. These include The fast growing volume of multimodal implanted electrodes
magnetic resonance imaging (MRI), single photon emission com- as ECoG for recording electrophysiological signal data is playing a
puted tomography (SPECT), positron emission tomography (PET), crucial role in health care and clinical research among many dis-
inpatient scalp EEG and video monitoring (phase I), sodium amo- ease domains, such as epilepsy. There is an urgent need to develop
barbital study and a neuropsychological profile. In select cases, a new methods using advanced technologies such as cloud and
variety of further MR postprocessing applications and magnetoen- mobile computing in order to assist in the processing of EEG data
cephalography (MEG) are applied [66]. and develop pervasive computing applications such as real-time
Several quantitative neuroimaging metrics have been applied seizure detection. Our solution – which addresses the challenges
to provide greater precision and reproducibitlity in defining puta- associated with near-real-time processing of EEG big data for diag-
tive sites of epileptogenicity particularly as it applies to the most nosis and treatment of refractory epilepsy, may change the quality
common area of involvement, the mesial temporal lobe. These are of life of patients. The proposed system provides pervasive big
 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158 157

data collection and analysis useful for real-time support of epilepsy [27] Altaf MAB, Yoo J. A 1.83 ␮J/classification, 8-channel, patient-specific epileptic
patients. The best strategy to prevent epileptic seizures is to detect seizure classification soc using a non-linear support vector machine. IEEE
Trans Biomed Circ Syst 2016;10:49–60.
them accurately at onset and to attempt promptly the appropriate [28] Mirzaei M, Salam MT, Nguyen DK, Sawan M. A fully-asynchronous low-power
therapy (e.g., by administering anticonvulsant drugs or applying implantable seizure detector for self-triggering treatment. IEEE Trans Biomed
brain neurostimulation), which depends on noninvasive and inva- Circ Syst 2013;7:563–72.
[29] Yadav R, Shah A, Loeb JA, Swamy M, Agarwal R. Morphology-based automatic
sive monitoring and investigation. seizure detector for intracerebral EEG recordings. IEEE Trans Biomed Eng
2012;59:1871–81.
[30] Yadav R, Swamy M, Agarwal R. Model-based seizure detection for intracranial
References EEG recordings. IEEE Trans Biomed Eng 2012;59:1419–28.
[31] Buteneers P, Verstraeten D, van Mierlo P, Wyckhuys T, Stroobandt D, Raedt R,
[1] Hosseini M-P, Nazem-Zadeh M-R, Pompili D, Jafari-Khouzani K, Elisevich K, Hallez H, Schrauwen B. Automatic detection of epileptic seizures on the
Soltanian-Zadeh H. Comparative performance evaluation of automated intra-cranial electroencephalogram of rats using reservoir computing. Artif
segmentation methods of hippocampus from magnetic resonance images of Intell Med 2011;53:215–23.
temporal lobe epilepsy patients. Med Phys 2016;43:538–53. [32] Luca S, Karsmakers P, Cuppens K, Croonenborghs T, Van de Vel A, Ceulemans
[2] Blum DE, Eskola J, Bortz JJ, Fisher RS. Patient awareness of seizures. Neurology B, Lagae L, Van Huffel S, Vanrumste B. Detecting rare events using extreme
1996;47:260–4. value statistics applied to epileptic convulsions in children. Artif Intell Med
[3] Correa AG, Orosco L, Diez P, Laciar E. Automatic detection of epileptic seizures 2014;60:89–96.
in long-term EEG records. Comput Biol Med 2014. [33] Kassahun Y, Perrone R, De Momi E, Berghöfer E, Tassi L, Canevini MP,
[4] Carrette S, Boon P, Sprengers M, Raedt R, Vonck K. Responsive Spreafico R, Ferrigno G, Kirchner F. Automatic classification of epilepsy types
neurostimulation in epilepsy. Expert Rev Neurother 2015;15:1445–54. using ontology-based and genetics-based machine learning. Artif Intell Med
[5] Hosseini M-P, Hajisami A, Pompili D. Real-time epileptic seizure detection 2014;61:79–88.
from EEG signals via random subspace ensemble learning. In: 2016 IEEE [34] Sahoo SS, Jayapandian C, Garg G, Kaffashi F, Chung S, Bozorgi A, Chen C,
international conference on autonomic computing (ICAC). IEEE; 2016. p. Loparo K, Lhatoo SD, Zhang G. Heart beats in the cloud: distributed analysis of
209–18. electrophysiological ‘big data’ using cloud computing for epilepsy clinical
[6] Kuzum D, Takano H, Shim E, Reed JC, Juul H, Richardson AG, de Vries J, Bink H, research. J Am Med Inf Assoc 2014;21.
Dichter MA, Lucas TH, Coulter DA, Cubukcu E, Litt B. Transparent and flexible [35] Hosseini M-P, Pompili D, Elisevich K, Soltanian-Zadeh H. Optimized deep
low noise graphene electrodes for simultaneous electrophysiology and learning for EEG big data and seizure prediction BCI via internet of things.
neuroimaging. Nat Commun 2014;5. IEEE Trans Big Data 2017;3(4):392–404.
[7] Jayapandian C, Chen C-H, Dabir A, Lhatoo S, Zhang G-Q, Sahoo SS. Domain [36] Hosseini M-P, Tran TX, Pompili D, Elisevich K, Soltanian-Zadeh H. Deep
ontology as conceptual model for big data management: application in learning with edge computing for localization of epileptogenicity using
biomedical informatics. In: International conference on conceptual modeling. multimodal rs-fMRI and EEG big data. In: 2017 IEEE international conference
Springer; 2014. p. 144–57. on autonomic computing (ICAC). IEEE; 2017. p. 83–92.
[8] Pagán J, Risco-Martín JL, Moya JM, Ayala JL. Modeling methodology for the [37] Zao JK, Gan T, You C, Chung C, Wang Y, Mndez SJ, Mullen T, Yu C, Kothe C,
accurate and prompt prediction of symptomatic events in chronic diseases. J Hsiao C, et al. Pervasive brain monitoring and data sharing based on multi-tier
Biomed Inform 2016;62:136–47. distributed computing and linked data technology. Front Hum Neurosci
[9] Ombao H, Lindquist M, Thompson W, Aston J. Handbook of neuroimaging 2014;8.
data analysis; 2016. [38] Da B, Tudoran R, Costan A, Varoquaux G, Brasche G, Conrod P, Lemaitre H,
[10] Cui L, Sahoo SS, Lhatoo SD, Garg G, Rai P, Bozorgi A, Zhang G-Q. Complex Paus T, Rietschel M, Frouin V, et al. Machine learning patterns for
epilepsy phenotype extraction from narrative clinical discharge summaries. J neuroimaging-genetic studies in the cloud. Front Neuroinform 2014;8.
Biomed Inform 2014;51:272–9. [39] Bell GS, Neligan A, Sander JW. An unknown quantity – the worldwide
[11] Combi C, Pozzani G, Pozzi G. Telemedicine for developing countries: a survey prevalence of epilepsy. Epilepsia 2014;55:958–62.
and some design issues. Appl Clin Inform 2016;7:1025–50. [40] Kiskani MK, Sadjadpour H. Secure and private cloud storage systems with
[12] Zhou W, Liu Y, Yuan Q, Li X. Epileptic seizure detection using lacunarity and random linear fountain codes. arXiv preprint 2017.
Bayesian linear discriminant analysis in intracranial EEG. IEEE Trans Biomed [41] Hosseini M-P, Nazem-Zadeh MR, Pompili D, Soltanian-Zadeh H. Statistical
Eng 2013;60:3375–81. validation of automatic methods for hippocampus segmentation in MR
[13] Khan YU, Gotman J. Wavelet based automatic seizure detection in images of epileptic patients. IEEE international conference of the engineering
intracerebral electroencephalogram. Clin Neurophysiol 2003;114:898–908. in medicine and biology society (EMBC) 2014:4707–10.
[14] Ghosh-Dastidar S, Adeli H, Dadmehr N. Mixed-band wavelet-chaos-neural [42] Kiranyaz S, Ince T, Zabihi M, Ince D. Automated patient-specific classification
network methodology for epilepsy and epileptic seizure detection. IEEE Trans of long-term electroencephalography. J Biomed Inform 2014;49:16–31.
Biomed Eng 2007;54:1545–51. [43] Urigüen JA, Garcia-Zapirain B. EEG artifact removal – state-of-the-art and
[15] Conradsen I, Beniczky S, Hoppe K, Wolf P, Sorensen HB. Automated algorithm guidelines. J Neural Eng 2015;12:031001.
for generalized tonic–clonic epileptic seizure onset detection based on SEMG [44] Nathan K, Contreras-Vidal JL. Negligible motion artifacts in scalp
zero-crossing rate. IEEE Trans Biomed Eng 2012;59:579–85. electroencephalography (EEG) during treadmill walking. Front Hum Neurosci
[16] Hopfeng R, Kasper BS, Graf W, Gollwitzer S, Kreiselmeyer G, Stefan H, Hamer 2016;9:708.
H. Automatic seizure detection in long-term scalp EEG using an adaptive [45] Castermans T, Duvinage M, Cheron G, Dutoit T. About the cortical origin of the
thresholding technique: a validation study for clinical routine. Clin low-delta and high-gamma rhythms observed in EEG signals during treadmill
Neurophysiol 2014;125:1346–52. walking. Neurosci Lett 2014;561:166–70.
[17] Pachori RB, Patidar S. Epileptic seizure classification in EEG signals using [46] Gwin JT, Gramann K, Makeig S, Ferris DP. Removal of movement artifact from
second-order difference plot of intrinsic mode functions. Comput Methods high-density EEG recorded during walking and running. J Neurophysiol
Programs Biomed 2014;113:494–502. 2010;103:3526–34.
[18] Hassan AR, Siuly S, Zhang Y. Epileptic seizure detection in EEG signals using [47] Najmi A, Webber W, Lesser H, Lesser R. Characterization of subdural
tunable-Q factor wavelet transform and bootstrap aggregating. Comput stimulation-induced after discharge activity using the continuous wavelet
Methods Programs Biomed 2016;137:247–59. transform. IEEE Trans Biomed Eng 2015.
[19] Rana P, Lipor J, Lee H, Van Drongelen W, Kohrman MH, Van Veen B. Seizure [48] Zandi AS, Javidan M, Dumont GA, Tafreshi R. Automated real-time epileptic
detection using the phase-slope index and multichannel ECOG. IEEE Trans seizure detection in scalp EEG recordings using an algorithm based on
Biomed Eng 2012;59:1125–34. wavelet packet transform. IEEE Trans Biomed Eng 2010;57:1639–51.
[20] Orosco L, Correa AG, Leber EL. Epileptic seizures detection based on empirical [49] Arunkumar N, Balaji VS, Ramesh S, Natarajan S, Likhita VR, Sundari S.
mode decomposition of EEG signals. Manag Epilepsy Res Results Treat 2011:1. Automatic detection of epileptic seizures using independent component
[21] Gotman J. Automatic seizure detection: improvements and evaluation. analysis algorithm. Proc. of IEEE international conference of advances in
Electroencephalogr Clin Neurophysiol 1990;76:317–24. engineering, science and management (ICAESM) 2012:542–4.
[22] Gabor AJ. Seizure detection using a self-organizing neural network: validation [50] Knowles D, Ghahramani Z. Infinite sparse factor analysis and infinite
and comparison with other detection strategies. Electroencephalogr Clin independent components analysis. In: Independent component analysis and
Neurophysiol 1998;107:27–32. signal separation. Springer Berlin Heidelberg; 2007. p. 381–8.
[23] Wang Y, Qi Y, Zhu J, Zhang J, Wang Y, Pan G, Zheng X, Wu Z. A cauchy-based [51] Leiva-Murillo A, Arts-Rodrguez JM. Maximization of mutual information for
state-space model for seizure detection in EEG monitoring systems. IEEE supervised linear feature extraction. IEEE Trans Neural Netw
Intell Syst 2015;30:6–12. 2007;18:1433–41.
[24] Becq G, Kahane P, Minotti L, Bonnet S, Guillemaud R. Classification of epileptic [52] Dietterich TG. Ensemble methods in machine learning. In: Multiple classifier
motor manifestations and detection of tonic–clonic seizures with acceleration systems. Springer; 2000. p. 1–15.
norm entropy. IEEE Trans Biomed Eng 2013;60:2080–8. [53] Hosseini MP, Nazem-Zadeh MR, Mahmoudi F, Ying H, Soltanian-Zadeh H.
[25] Do Valle BG, Cash SS, Sodini CG, Low-power. 8-channel EEG recorder and Support vector machine with nonlinear-kernel optimization for lateralization
seizure detector ASIC for a subdermal implantable system. IEEE Trans Biomed of epileptogenic hippocampus in MR images. IEEE international conference of
Circ Syst 2016;10:1058–67. the engineering in Medicine and Biology Society (EMBC) 2014:1047–50.
[26] Salam MT, Sawan M, Nguyen DK. A novel low-power-implantable epileptic [54] Tang B, He H, Enn. Extended nearest neighbor method for pattern recognition
seizure-onset detector. IEEE Trans Biomed Circ Syst 2011;5:568–78. [research frontier]. IEEE Comput Intell Mag 2015;10:52–60.
158 M.-P. Hosseini et al. / Artificial Intelligence in Medicine 84 (2018) 146–158

[55] Stead M, Bower M, Brinkmann BH, Lee K, Marsh W, Meyer FB, Litt B, Van G, Conference on Signal and Information Processing (GlobalSIP). IEEE; 2016. p.
Worrell G. Microseizures and the spatiotemporal scales of human partial 1151–5.
epilepsy; 2010. [61] Tran TX, Hosseini MP, Pompili D. Mobile edge computing: recent efforts and
[56] Pedley TA. EEG patterns that mimic epileptiform discharges but have no five key research directions. IEEE COMSOC MMTC Commun -Front 2017.
association with seizures. In: Current clinical neurophysiology: update on [62] Mousaei M, Smida B. Optimizing Pilot Overhead for Ultra-Reliable
EEG and evoked potentials. New York: Elsevier/North Holland; 1980. Short-Packet Transmission. arXiv preprint 2017.
[57] Barkmeier DT, Shah AK, Flanagan D, Atkinson MD, Agarwal R, Fuerst DR, [63] Hosseini MP. Proposing a new artificial intelligent system for automatic
Jafari-Khouzani K, Loeb JA. High inter-reviewer variability of spike detection detection of epileptic seizures. J Neurol Disord 2015;3.4.
on intracranial EEG addressed by an automated multi-channel algorithm. Clin [64] Hosseini MP, Soltanian-Zadeh H, Akhlaghpoor Sh. Detection and severity
Neurophysiol 2012;123:1088–95. scoring of chronic obstructive pulmonary disease using volumetric analysis of
[58] Wasade VS, Elisevich K, Tahir R, Smith B, Schultz L, Schwalb J, Spanaki-Varelas lung CT images. Iran J Radiol 2012;9(1):22–7.
M. Long-term seizure and psychosocial outcomes after resective surgery for [65] Hosseini MP, Soltanian-Zadeh H, Akhlaghpoor Sh. Computer-aided diagnosis
intractable epilepsy. Epilepsy Behav 2015;43:122–7. system for the evaluation of chronic obstructive pulmonary disease on CT
[59] Shirazi M, Sani A, Vosoughi A. Sensor Selection and Power Allocation via images. Tehran Univ Med J TUMS Publ 2011;68(12):718–25.
Maximizing Bayesian Fisher Information for Distributed Vector Estimation. [66] Hosseini MP, Nazem-Zadeh MR, Pompili D, Jafari-Khouzani K, Elisevich K,
arXiv preprint 2017. Soltanian-Zadeh H. Automatic and manual segmentation of hippocampus in
[60] Hosseini MP, Soltanian-Zadeh H, Elisevich K, Pompili D. Cloud-based deep epileptic patients MRI. arXiv preprint 2016.
learning of big eeg data for epileptic seizure prediction. In: 2016 IEEE Global