You are on page 1of 6

2018 15th International Conference on Electrical Engineering, Computing Science and Automatic Control (CCE)

Mexico City, Mexico September 5-7, 2018

Pediatric Seizure Forecasting using Nonlinear


Features and Gaussian Mixture Hidden Markov
Models on Scalp EEG Signals
Carlos Emiliano Solórzano-Espı́ndola Blanca Tovar-Corona Álvaro Anzueto-Rı́os
Instituto Politécnico Nacional Instituto Politécnico Nacional Instituto Politécnico Nacional
UPIITA - Lab. Biomecánica UPIITA - Academia de Sistemas UPIITA - Lab. Biomecánica
Mexico City, Mexico Mexico City, Mexico Mexico City, Mexico
csolorzanoe1200@alumno.ipn.mx bltovar@ipn.mx aanzuetor@ipn.mx

Abstract—Seizure forecasting systems have been studied in optimal for the case. Some of these AEDs may have unde-
recent years for improving the quality of life for patients sirable side effects, that is one of the reasons that medical
with epilepsy and gain further understanding about seizures. A doctors require to choose the one that minimizes the risk of
common approach for this is the study of electroencephalography
(EEG) recordings, using signal processing techniques and, more developing side effects [7], and approximately 20-30% of the
recently, machine learning algorithms. A four-stage system is patients cannot find suitable AEDs that reduce or stop seizures
developed for patient-specific seizure prediction; consisting of [8]. Therefore the medic must evaluate and present alternatives
pre-processing, dimensionality reduction, feature extraction and to the patient and, when it is required, surgery.
classification between interictal and preictal EEG signals. A A more recent alternative is the use of electrical stimulation
hybrid method using principal component analysis (PCA) and
independent component analysis (ICA) is applied for dimension- (neuromodulation) systems [7], [8]. Those systems use electric
ality reduction. Nonlinear features are selected for the analysis pulses that target specific nerves in order to attenuate seizures.
and characterization of the signals. A Hidden Markov Model The systems can administrate the pulse intermittently (open
(HMM) with Gaussian mixture emissions is trained for each type loop) or activate the stimulus after some feedback reading
of signal and evaluated as a classifier. A sensitivity of 0.95 and signals from the patient (closed loop).
a specificity of 0.86 were achieved.
Index Terms—seizure, forecasting, electroencephalography, II. E PILEPSY FORECASTING
Markov, entropy
Seizure detection systems provide a way to detect the
I. I NTRODUCTION moment in which the seizure is occurring. The data for the
Epilepsy is a common neurological disorder, with an es- detection can come from many sources like electrocardiogra-
timated incidence between 24 to 70 out of 100,000 people phy (ECG), accelerometry, chemical sensors and others [7].
according to different studies [1]–[3]. This proportion also The most common approach is the electroencephalography
depends on risk factors as some groups have been shown to signals (EEG) since previous research has been done by
have a higher incidence and prevalence of seizures. Some of neurologists to identify parameters that differentiate the base
this factors include age, sex, family history, physical trauma, state (interictal) to the state with an ongoing seizure (ictal).
environmental toxins, among others [4], [5]. Some methods have been optimized for real-time classifi-
Epilepsy is a disease of the brain characterized by the cation and are used in the electric stimulation systems. An
predisposition to have epileptic seizures. Seizures are neuro- example of this is the Neuropace, an implantable device able
logical events with abnormal and excessive electric discharges to identify abnormal neurological activity and deliver pulses
originated from a synchronized activity of neurons [6]. The to the vagus nerve.
onset can be either partial or generalized depending on how
A. Prediction systems
the seizure begins and can be further classified according to
the symptoms that they manifest. Although seizure detection has delivered improvement in
the treatment of seizure control and study, there are still
A. Treatment of Epilepsy benefits on the prediction of seizures compared with simple
The treatment for epilepsy has evolved over the years giving detection. Patients with epilepsy feel that the problem of the
different types of solutions aiming to control of seizures. uncertainty of seizures affects their life quality by inducing
Nowadays there is a broad range of anti-epileptic drugs fear to perform some activities along with psychological and
(AEDs) targeting different types of seizures and stages of the social problems [8].
disease, giving options for physicians to study their patient’s Neuromodulation systems have shown a better performance
conditions and prescribe the AED that they consider to be when the stimulus is applied moments before the seizure

978-1-5386-7033-0/18/31.00©2018IEEE
begins. The ability to predict seizures may also lead to alarm rhythms. Studies about epileptogenic activity have shown that
systems that provide a way for patients to perform their daily there is still relevant information above the typical range,
activities more secure by taking preventing actions prior to a considering this, a broader range is used for the pre-processing
seizure. Predictions systems based on EEG signals hypothesize stage [15].
the existence of a transition period called the preictal state, a For the pre-processing stage, a band-pass filter is used to
state where the patients present changes in physiological and attenuate components considered to be noise, coming from
psychological activity. other biological signals and movement artifacts. A 4th order
The preictal state is still a research subject for neurologists, band-pass Butterworth filter is selected for this, the frequency
the difference that it manifest with the interictal state appears range is between 0.5 and 120 Hz.
to be patient dependant. The time window that has to be used To eliminate the noise coming from the power line a 2nd
for forecasting has yet to be determined and also changes from order Butterworth notch filter is used with a rejection band of
patient to patient [9]. 57 to 63 Hz.
B. Proposed System C. Dimensionality reduction
The stages considered for the system in this work are Real time computation of this methods is a key point to the
based on the combination of many approaches for this task, task of seizure prediction, this is why dimensionality reduction
taking elements that have been used for seizure detection and techniques are commonly used [16]. Dimensionality reduction
prediction. The stages are designed considering a Gaussian techniques analyze high-dimensional data and transform it to
Mixture Model-Hidden Markov Model (HMM) based classi- a new space where less variables are needed to represent
fier, proposed as a final step. The steps are: most of the information. The methods take into account what
• Signal pre-processing: Filter the noise in the signals. is expected to be preserved in the new representation and
• Dimensionality reduction: Reduce the amount of data to what form of information is needed. They are also useful for
process in further steps. exploration and visualization when the number of features is
• Feature extraction: Measurements from the signals. too high to be represented on a graph.
• Classification: Train a GMM-HMM per each type of 1) PCA: Principal Component Analysis (PCA) is one of the
signal and patient to identify them. most popular techniques in Machine Learning tasks. It aims
to find a linear combination of the original variables using
III. M ETHODOLOGY the covariance matrix to obtain an orthogonal projection that
A. Materials represents as much variance of the data as possible. The new
The data chosen for this study were scalp EEG records, variables in this space are the principal components (PCs). The
they were acquired from the Children’s Hospital Boston- projection of the data is defined by (1), where X is the original
Massachusetts Institute of Technology, dataset hosted in the data with dimension N by p, W is the matrix containing the
Physionet website [10]. The data consist of scalp EEG samples eigenvectors of the covariance matrix with dimension p by p
of one hour each, sampled at 256 Hz with 16-bit resolution. and Z are the PCs.
The recordings were acquired from 23 pediatric patients with
epilepsy as described in [11]. Z = WTX (1)
The recordings were segmented as in the work of Varathara- Each eigenvector is associated with an eigenvalue λi that
jah et. al. [9] where 10-minutes clips are acquired from an represents the amount of variance explained by that compo-
interval previous to the beginning of a seizure, and considering nent, that can be used to select how many eigenvectors from
a forecast horizon for possible ictal activity undetected by the the W matrix should be preserved in the final transformation,
reviewers that made the annotations of the seizures. In this preserving the ones that represent the highest variance. The
case, a 30-minutes interval previous is proposed, acquiring percentage of variance explained by the eigenvalues 1 to m,
three preictal segments per recording with a seizure if possible. considering being are ordered from highest to lowest, can be
The interictal segments were taken from recordings with no computed using equation (2).
ictal activity, taking care that in the following recording there
Pm
was no ictal activity when they correspond to the same session. j=1 λj
The algorithms for each stage are implemented for the var% = Pp (2)
j=1 λj
Python 2.7 programming language using the libraries of data
processing and machine learning from the SciKits [12]–[14]. In this work, the number of remaining channels should
The computer for the task is an HP (R) 14-k106la with an represent at least approximately 80% of the variance.
Intel (R) i5-4200U 2.3 GHz processor and with Windows 10 2) ICA: Another method commonly used to transform the
(R) as the operating system. data is the independent component analysis (ICA). In this
algorithm, the data is supposed to be a linear mixture of
B. Signal pre-processing non-gaussian components from independent sources. The goal
Traditional study of EEG signals proposes a frequency range for the algorithm is to find the transformation for the data
of 0.5 to 30 Hz, considering the delta, theta, alpha and beta that either minimizes the mutual information (calculated using
measures like the Kullback-Leiber divergence and maximum
entropy) or maximizes the independence (calculated using
non-gaussianity approximate measures like kurtosis and ne-
gentropy) [17]. The new representation of the data is expressed
in (3), where X is the original data, A is the unmixing matrix
and Z the independent components.

Z = AX (3)

The goal of the algorithm is to find the elements of the Fig. 1. Sequence of proposed features for a sine wave with three frequencies.
A matrix that give a representation of the data with the new
components as independent as possible.
3) PCA-ICA: The ICA algorithm performs better if some the Fisher’s coefficient of kurtosis is expressed in (6) with µ
pre-processing like whitening and detrending is applied, both as the mean and σ as the standard deviation.
can be achieved by using PCA before ICA, with the additional N  4
advantage of dimensionality reduction. If dimensionality re- 1 X xi − µ
kurotsis(X) = −3 (6)
duction is used, the remaining matrix will be AW T where N i=1 σ
W ∈ Rn×m and A ∈ Rm×m being n the number of original
4) Feature sequence: The feature extraction is applied to
channels and m the number of remaining channels.
the signal by calculating the given formulas to a segment of the
D. Feature extraction signal with a fixed length or duration. Once all the features
are calculated, a vector is formed with each of the features
Applying a classification algorithm over the raw signals representing a dimension. This process is repeated for each
can be both computationally expensive and difficult to find a segment of the signal until the features for the whole signal
pattern to characterize the classes. The extraction of features are obtained. This is represented in Fig. 1 where the features
allows characterizing the behaviour of signals from hidden for a sine wave with three different frequencies are obtained
information that can be obtained through calculations over the for each segment of 20 samples.
data. The features proposed for the system are the following:
1) Entropy: Entropy is commonly described as a measure- E. Classification
ment of the disorder of a system. In the information theory For the classification part of the system the GMM-HMM
field, entropy is a measurement of the uncertainty of the data. model is proposed. The classifier consists of a Hidden Markov
The Shannon entropy is a nonlinear measurement used to Model where the observable states are assumed to be a mixture
quantify the complexity level of a data series [18]. The formula of Gaussians.
used to calculate the Shannon entropy is described at (4) with 1) Hidden Markov Models: HMMs are used for modeling
p(i) being the probability of that value to appear. a sequence of observable states with assumed hidden states
. HMMs used the Markov assumption, represented in (7),
k
X which states that for a sequence X = (x1 , x2 , .., xt−1 , xt )
H=− p(i) log p(i) (4) the probability for the next state depends only on the actual
i
state [21].
2) Teager-Kaiser energy: The Teager-Kaiser energy is a
non-linear method for estimating the energy of a signal con- P (xt |X) = P (xt |xt−1 ) (7)
sidering a second order or higher system, where both the
In these models the sequences of observable states are
amplitude and the frequency influence the result [19]. The
assumed to be caused by some hidden states. These hidden
calculation is carried out using the Teager-Kaiser operator
states have a probability of emission for each of the finite set
defined by equation (5), considering a discrete sequence
of observable states, and also a probability for changing to
X = (x1 , x2 , ..., xn ).
any other hidden state or to stay in the same one [22]. The
parameters to fit in these models for k hidden states and m
Ψ(x[n]) = x2 [n] − x[n − 1] ∗ x[n + 1] (5) observable states are:
k
3) Kurtosis: Statistical moments give information about • An initial probability vector π ∈ R .
k×k
how the data is distributed. The most common moments are • A transition probability matrix A ∈ R .
k×m
the mean (1st moment) and variance (2nd moment) from the • An emission probability matrix B ∈ R .
normal distribution, Higher order moments, skewness and 2) Gaussian Mixture Models: As the data from the features
kurtosis, describe the shape of the distribution; considering is continuous, a clustering algorithm is needed to discretize the
a normal distribution, the typical values for both are zero and data to a finite set of observable states [23]. GMM is a non-
are nondimensional. Kurtosis describes the relative peakdness supervised clustering algorithm that models a set of data as
or flatness of a distribution [20]. The formula used to calculate the combination of Gaussian distributions [24]. The model is
[b]
Fig. 2. Diagram for a HMM with three hidden states and four observable
states. The probabilities ai ∈ A are the transition probabilities among hidden
states, the bi ∈ B probabilities are the probabilities that the observable state
appears given the hidden state.

Fig. 3. Histogram distribution for both states using signals of 20 minutes of


shown in (8) where φi represents the probability of belonging recording for each state.
to the cluster i, and each cluster has a normal distribution
N (µi , Σi ) with a vector of means µi and a covariance matrix
Σi . IV. R ESULTS
The recordings are first filtered using the proposed filters.
k
X With the resulting data, the ICA algorithm was applied for
p(x) = φi N (µi , Σi ) (8) dimensionality reduction. The FastICA algorithm was used
i for this purpose, which is an efficient iterative algorithm
A parameter to measure if the number of clusters in the to approximate non-guassianity maximization. The algorithm
resulting fitting is considered appropriate is the Bayesian In- first performs PCA for whitening and centering.
formation Criterion (BIC), that takes into account the number The channels were reduced to 6 as it was found that for 23
of parameters and the likelihood of the model to evaluate the channels in the recordings, the first 6, after PCA, preserved
model [25]. The BIC is expressed in (9) with L̂ being the at least ≈ 80% of the variance. The resulting signals are
maximized likelihood, n the sample size and k the number of then unmixed using FastICA,which may converge to a local
parameters of the model. minimum result as an iterative algorithm, this is why the
algorithm was run several times and the resulting matrix of
BIC = −2 log (L̂) + k log (n) (9) the combined best result was saved per each patient.
Once the data dimension was reduced and the resulting
The models above are combined in a more complex model channels are as independent as possible, the feature extraction
called GMM-HMM. This models have been used for se- was applied. A running window of the equivalent to one
quences of continous data where, for each each observation of second of data (256 samples) calculates the three features
the sequence, first the GMM allows to identify the probability proposed and are saved as the feature vector for all the
of belonging to a cluster that will then be the observable segments in the signal. The histograms for the calculated
state for the HMM, the HMM then decodes the probability features for both states are shown in Fig 3.
of belonging to a hidden state given the observation and the Several GMMs were fitted using samples from preictal and
previous state (A matrix) or in the case of being the first interictal states using a range from 5 to 20 clusters to find the
of the sequence, the probability of the sequence to start in model that best described the distribution. For the clustering of
that state (π matrix). The overall likelihood of a sequence the given features, 15 clusters were found as optimal according
X = (x1 , x2 , ..., xT ) for a given model is calculated by (10). to the BIC criterion. The results of the clustering are shown
in Fig. 4.
T
Y The HMM models were trained using a total of 169 clips
p(x1 , x2 , ..., xT ) = π(z1 )B(y1 |z1 ) A(zt |zt − 1)B(yt |zt ) of preictal state and 393 clips of interictal state, equivalent
t=2 to a total of 93.6 hours of data. The log-likelihood was
(10) calculated for each sequence of data for each state, and the
Where yt denotes the observable state and zt the hidden model that presents the greater was saved as the model for the
state at time t. The observable state yt can be obtained as corresponding state.
the most likely cluster for a GMM with θ parameters as For each patient, a pair of models, one for each state, were
represented in (11). saved and evaluated, to generate a confusion matrix. Table I
shows the results for all the patients, obtaining a sensitivity
yt = argmax(P (Y |xt , θ)) (11) of 0.95 and a specificity of 0.86. In these type of systems a
Fig. 5. Sequence of the extracted Teager-Kaiser Energy for a preictal state.
A seizure is marked between the dashed lines.

Fig. 4. Scatter matrix of features with the clusters found by the GMM
algorithm, each color representing a cluster. The diagonal of the matrix
represents an estimate of the kernels densities.

higher value of sensitivity than the specificity is preferred as


patients prefer false alarms than not detected seizures [8].

TABLE I
C ONFUSION M ATRIX

Predicted State Fig. 6. Sequence of features in the 10 minutes prior to a seizure. The seizure
Interictal Preictal Total is marked by the dashed lines.
True Interictal 345 48 393
True Preictal 7 162 169
Total 352 210 562
and the entropy measures the disorder in the system, however,
this can also happen in the interictal state.
TP A key difference for the preictal state is that the kurtosis
Sensitivity = ≈ 0.95 value shows a mirroring behaviour with increments at the
TP + FN
same time that the entropy decreases. This can be seen in
TN Fig. 6 and 7 with 10-minutes segments of the sequence of
Specif icity = ≈ 0.86 features (entropy, Teager-Kaiser energy and kurtosis from top
TN + FP
to bottom) containing a seizure. Both were acquired from the
patient chb01. It is important to mention that the reviewers
V. D ISCUSSION AND C ONCLUSIONS
marked a seizure between the seconds 2996 and 3036 as
Although the resulting matrix for ICA reduces the amount of highlighted by the dashed lines in Fig. 6 and a seizure is
information, the calculated features in the signals after being between 1467 and 1494 in Fig. 7.
processed with PCA and ICA, display simmilar patterns as It was also observed that the Teager-Kaiser energy displays
in the non-processed counterpart. The sequence of features is an increment as shown in Fig. 6 and 7 at the onset of the
preserved with the scale being the main difference. This is seizure and sometimes a few seconds in advance. This allows
illustrated in Fig. 5 for the same segment of a recording with to exclude the ictal state from the ones considered for the
a seizure and the calculated Teager-Kaiser feature before and study and it was found useful for the identification of the
after the processing. beginning of a seizure, thus obtaining the average seizure
Something to be noted is that in the preictal states some horizon to exclude from the preictal state. This also may lead
authors present the tendency of the entropy to decrease [18], to a technique for efficient seizure detection in a further work.
[26], [27]. This is observed also in this work at many intervals The proposed system is able to detect the preictal signals
before seizures. This is to be expected, as a seizure by with an accuracy of 0.9, comparable to other systems [28]. The
definition is caused by synchronous activity of the neurons system described in this work is able to characterize each state
[12] W. McKinney, “Data structures for statistical computing in python,” in
Proceedings of the 9th Python in Science Conference, S. van der Walt
and J. Millman, Eds., 2010, pp. 51 – 56.
[13] E. Jones, T. Oliphant, P. Peterson et al., “SciPy: Open source scientific
tools for Python,” 2001–, [Online; accessed 11/2017]. [Online].
Available: http://www.scipy.org/
[14] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
esnay, “Scikit-learn: Machine learning in Python,” Journal of Machine
Learning Research, vol. 12, pp. 2825–2830, 2011.
[15] C. Alvarado-Rojas, M. Valderrama, A. Fouad-Ahmed, H. Feldwisch-
Drentrup, M. Ihle, C. A. Teixeira, F. F. Sales, A. Schulze-Bonhage,
C. Adam, A. Dourado, S. Charpier, V. Navarro, and M. L. V. Quyen,
“Slow modulations of high-frequency activity (40-140 hz) discriminate
preictal changes in human focal epilepsy,” in Scientific reports, 2014.
Fig. 7. Sequence of features in the 10 minutes prior to a seizure. The seizure [16] F. La Foresta, N. Mammone, and F. Morabito, “Pca-ica for automatic
is marked by the dashed lines. identification of critical events in continuous coma-eeg monitoring,”
Biomedical Signal Processing and Control, vol. 4, no. 3, pp. 229–235,
7 2009.
of the signals for the analyzed recordings in order to predict [17] A. Hyvrinen and E. Oja, “Independent component analysis:
algorithms and applications,” Neural Networks, vol. 13,
the most probable state. The usage of dimensionality reduction no. 4, pp. 411 – 430, 2000. [Online]. Available:
algorithms allows to extract features able to characterize each http://www.sciencedirect.com/science/article/pii/S0893608000000265
state with less calculations, something required for real-time [18] A. S. Zandi, G. A. Dumont, M. Javidan, and R. Tafreshi, “An entropy-
based approach to predict seizures in temporal lobe epilepsy using scalp
applications. eeg,” in 2009 Annual International Conference of the IEEE Engineering
in Medicine and Biology Society, Sept 2009, pp. 228–231.
R EFERENCES [19] E. Kvedalen, “Signal processing using the teager energy operator and
[1] J. Christensen, M. Vestergaard, M. G. Pedersen, C. B. Pedersen, other nonlinear operators cand. scient thesis,” Ph.D. dissertation, Uni-
J. Olsen, and P. Sidenius, “Incidence and prevalence of epilepsy in versity of Oslo, 2003.
denmark,” Epilepsy Research, vol. 76, no. 1, pp. 60 – 65, 2007. [20] R. Kalpana and M. Chitra, “Classifiction of different brain states
[Online]. Available: https://doi.org/10.1016/j.eplepsyres.2007.06.012 from eeg signal using statistical parameters,” International Journal of
[2] W. A. Hauser, J. F. Annegers, and L. T. Kurland, “Incidence of epilepsy Advanced Computing, Engineering and Application, vol. 1, no. 3, 2012.
and unprovoked seizures in rochester, minnesota: 19351984,” Epilepsia, [21] S. Wong, A. B. Gardner, A. M. Krieger, and B. Litt, “A
vol. 34, no. 3, pp. 453–458, 1993. stochastic framework for evaluating seizure prediction algorithms
[3] K. M. Fiest, K. M. Sauro, S. Wiebe, S. B. Patten, C.-S. Kwon, using hidden markov models,” Journal of Neurophysiology,
J. Dykeman, T. Pringsheim, D. L. Lorenzetti, and N. Jetté, “Prevalence vol. 97, no. 3, pp. 2525–2532, 2007. [Online]. Available:
and incidence of epilepsy,” Neurology, vol. 88, no. 3, pp. 296–303, http://jn.physiology.org/content/97/3/2525.full.pdf
2017. [Online]. Available: http://n.neurology.org/content/88/3/296 [22] M. H. Abdullah, J. M. Abdullah, and M. Z. Abdullah, “Seizure detection
[4] M. J. van Putten and J. Hofmeijer, “Generalized periodic discharges: by means of hidden markov model and stationary wavelet transform
Pathophysiology and clinical considerations,” Epilepsy & Behavior, of electroencephalograph signals,” in Proceedings of 2012 IEEE-EMBS
vol. 49, no. Supplement C, pp. 228 – 233, 2015, sI:Status Epilepticus. International Conference on Biomedical and Health Informatics, Jan
[Online]. Available: https://doi.org/10.1016/j.yebeh.2015.04.007 2012, pp. 62–65.
[5] J. Godoy., Electroencefalografı́a en las epilepsias. Epilepsias: Di- [23] S. Esmaeili, B. N. Araabi, H. Soltanian-Zadeh, and L. Schwabe,
agnóstico y Tratamiento. Chile: Mediterraneo, 2004. “Variational bayesian learning for gaussian mixture hmm in seizure
[6] R. S. Fisher, J. H. Cross, J. A. French, N. Higurashi, E. Hirsch, prediction based on long term eeg of epileptic rats,” in 2014 21th Iranian
F. E. Jansen, L. Lagae, S. L. Mosh, J. Peltola, E. Roulet Perez, Conference on Biomedical Engineering (ICBME), Nov 2014, pp. 138–
I. E. Scheffer, and S. M. Zuberi, “Operational classification of 143.
seizure types by the international league against epilepsy: Position [24] S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach,
paper of the ilae commission for classification and terminology,” 3rd ed. Upper Saddle River, NJ, USA: Prentice Hall Press, 2009.
Epilepsia, vol. 58, no. 4, pp. 522–530, 2017. [Online]. Available: [25] D. Posada, T. R. Buckley, and J. Thorne, “Model selection and model
http://dx.doi.org/10.1111/epi.13670 averaging in phylogenetics: Advantages of akaike information criterion
[7] K. Gadhoumi, J.-M. Lina, F. Mormann, and J. Gotman, “Seizure and bayesian approaches over likelihood ratio tests,” Systematic
prediction for therapeutic devices: A review,” Journal of Neuroscience Biology, vol. 53, no. 5, pp. 793–808, 2004. [Online]. Available: +
Methods, vol. 260, no. Supplement C, pp. 270 – 282, 2016, methods http://dx.doi.org/10.1080/10635150490522304
and Models in Epilepsy Research. [26] V. Srinivasan, C. Eswaran, and N. Sriraam, “Approximate entropy-
[8] S. Ramgopal, S. Thome-Souza, M. Jackson, N. E. Kadish, I. S. Fernndez, based epileptic eeg detection using artificial neural networks,” IEEE
J. Klehm, W. Bosl, C. Reinsberger, S. Schachter, and T. Loddenkemper, Transactions on Information Technology in Biomedicine, vol. 11, no. 3,
“Seizure detection, seizure prediction, and closedloop warning systems pp. 288–295, May 2007.
in epilepsy,” Epilepsy and Behavior, vol. 37, pp. 291 – 307, 2014. [27] A. A. Bruzzo, B. Gesierich, M. Santi, C. A. Tassinari, N. Birbaumer,
[9] Y. Varatharajah, R. K. Iyer, B. M. Berry, G. A. Worrell, and and G. Rubboli, “Permutation entropy to detect vigilance changes and
B. H. Brinkmann, “Seizure forecasting and the preictal state in preictal states from scalp eeg in epileptic patients. a preliminary study,”
canine epilepsy,” International Journal of Neural Systems, vol. 27, Neurological Sciences, vol. 29, no. 1, pp. 3–9, Feb 2008. [Online].
no. 01, p. 1650046, 2017, pMID: 27464854. [Online]. Available: Available: https://doi.org/10.1007/s10072-008-0851-3
https://www.worldscientific.com/doi/abs/10.1142/S0129065716500465 [28] T. N. Alotaiby, S. A. Alshebeili, T. Alshawi, I. Ahmad, and F. E. Abd
[10] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, El-Samie, “Eeg seizure detection and prediction algorithms: a survey,”
P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. EURASIP Journal on Advances in Signal Processing, vol. 2014, no. 1,
Peng, and H. E. Stanley, “Physiobank, physiotoolkit, and physionet,” p. 183, Dec 2014.
Circulation, vol. 101, no. 23, pp. e215–e220, 2000. [Online]. Available:
http://circ.ahajournals.org/content/101/23/e215
[11] A. H. Shoeb, “Application of machine learning to epileptic seizure onset
detection and treatment,” Ph.D. dissertation, MIT Division of Health
Sciences and Technology, Massachusetts, USA, 7 2009.

You might also like