Computers in Biology and Medicine 119 (2020) 103671

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: http://www.elsevier.com/locate/compbiomed

Selection of features for patient-independent detection of seizure events

using scalp EEG signals
Shuhan Yang a, Bo Li d, Yinda Zhang a, Meiyu Duan a, Shuai Liu a, Yexian Zhang a, Xin Feng a, b,
Renbo Tan c, Lan Huang a, Fengfeng Zhou a, *
a
BioKnow Health Informatics Lab, College of Computer Science and Technology, Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of
Education, Jilin University, Changchun, 130012, Jilin, China
b
Jilin Institute of Chemical Technology, Jilin, 132000, Jilin, China
c
Cancer Systems Biology Center, The China-Japan Union Hospital, Jilin University, Changchun, 130033, China
d
BioKnow Health Informatics Lab, College of Software, Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University,
Changchun, 130012, Jilin, China

A R T I C L E I N F O A B S T R A C T

Keywords: Epilepsy involves brain abnormalities that may cause sudden seizures or other uncontrollable body activities.
EEG Epilepsy may have substantial impacts on the patient’s quality of life, and its detection heavily relies on tedious
Epileptic seizure detection and time-consuming manual curation by experienced clinicians, based on EEG signals. Most existing EEG-based
Time-domain feature extraction
seizure detection algorithms are patient-dependent and train a detection model for each patient. A new patient
MinMaxHist
Nonlinear features
can only be monitored effectively after several episodes of epileptic seizures. This study investigates the patient-
Feature selection independent detection of seizure events using the open dataset CHB-MIT Scalp EEG. First, a novel feature
XGBoost extraction algorithm called MinMaxHist is proposed to measure the topological patterns of the EEG signals.
Following this, MinMaxHist and several other feature extraction algorithms are applied to parameterize the EEG
signals. Next, a comprehensive series of feature screening and classification optimization experiments are con­
ducted, and finally, an optimized EEG-based seizure detection model is presented that can achieve overall values
for accuracy, sensitivity, specificity, Matthews correlation coefficient, and Kappa of 0.8627, 0.8032, 0.9222,
0.7504 and 0.7254, respectively, with only 30 features. The classification accuracy of the method with Min­
MaxHist features was 0.0464 higher than that without MinMaxHist features. Compared with existing methods,
the proposed algorithm achieved higher accuracy and sensitivity, as shown in the experimental results.

1. Introduction Although EEG has characteristics of low cost, non-invasiveness, and high
signal resolution [13,14], clinical annotation of EEG signals still relies
Chronic lesion epilepsy induces transient brain dysfunction via the heavily on human screening [15]. The visual inspection of EEG signals is
abnormal discharge of brain neurons [1,2]. Epilepsy is recognized as the recommended by guidelines, but is tedious and time-consuming [16].
second most common neurological disorder, ranked just after stroke [3]. Machine learning algorithms have become the research focus of EEG-
More than 20% of epilepsy patients suffer from seizures, which are based epileptic seizure detection problems [17–19], and several studies
usually refractory to medication [4]. Epileptic seizures occur recurrently have investigated classification based on EEG signals [20,21]. Epileptic
and unpredictably, and substantially impact the quality of life of the seizure detection is formulated as a classification problem of ictal and
patient [5,6]. Various modern imaging technologies have been applied interictal EEG signals; various feature extraction algorithms have been
to diagnose epilepsy, including magnetic resonance imaging (MRI) [7, utilized to generate features from a fixed-length window of EEG signals,
8], computed tomography (CT) [9], positron emission tomography including the time domain [22], frequency domain [23], and non-linear
(PET) [10,11] and electroencephalogram (EEG) [12]. characteristics [24,25]. After the feature extraction step, a variety of
EEG technology records electrophysiological signals to describe feature selection [26,27] and classification algorithms [5,28] may be
brain activity, and has been widely used to detect epilepsy [5,12]. utilized to generate seizure detection models. Deep learning algorithms

* Corresponding author.
E-mail addresses: [email protected], [email protected] (F. Zhou).

https://doi.org/10.1016/j.compbiomed.2020.103671
Received 2 January 2020; Received in revised form 20 February 2020; Accepted 20 February 2020
Available online 21 February 2020
0010-4825/© 2020 Elsevier Ltd. All rights reserved.
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Table 1
EEG signal samples retrieved from each patient. Each row gives the total number of EEG signal samples for a given patient. Columns marked “NumSeizure” and
“NumNonIctal” give the numbers of seizure and non-seizure regions recorded for this patient, while the columns marked “NumIctal” and “NumInterIctal” give the
numbers of sampling regions retrieved from the EEG signals of this patient. The column marked “SamplePair” shows the number of pairs of one positive (ictal) and one
negative (interictal) region retrieved as modeling data for this patient.
Patient NumSeizure NumIctal NumNonIctal NumInterIctal SamplePair

chb01 7 70 34 204 70
chb02 3 27 33 198 27
chb03 7 62 31 186 62
chb04 4 62 39 234 62
chb05 5 92 34 204 92
chb06 10 22 11 66 22
chb07 3 53 16 96 53
chb08 5 151 15 90 90
chb09 4 44 15 90 44
chb10 7 70 18 108 70
chb11 3 133 32 192 133
chb13 10 68 4 24 24
chb14 8 24 19 114 24
chb15 20 323 25 150 150
chb17 3 48 17 102 48
chb18 6 51 29 174 51
chb19 3 38 26 156 38
chb20 8 45 22 132 45
chb21 4 32 29 174 32
chb22 3 33 28 168 33
chb23 7 67 6 36 36
chb24 16 79 9 54 54
Total 146 1594 492 2952 1260

such as CapsNet and a deep belief network may also be utilized to 2. Materials and methods
characterize the phenotypic patterns within EEG signals [29,30].
Most studies calculate variables from EEG signals using feature 2.1. Problem setting
extraction algorithms, and formulate EEG-based seizure detection as a
machine learning problem using these extracted features [19,31,32]. This study investigated the patient-independent detection problem
Gabor et al. trained a self-organizing map (SOM) neural network to of epileptic seizures using scalp EEG signals, due to the following two
detect seizures in 24 long-term EEG recordings using time-frequency challenges identified in existing studies.
features, and achieved an accuracy of 90%, thus demonstrating the Previous studies have usually used patient-specific modeling [21,24,
possibility of automated seizure detection [33]. Time-frequency features 38–40], i.e., they have trained a separate model for each patient.
are widely used to describe EEG signals due to their non-stationary Although this approach can achieve high detection accuracy, clinicians
characteristics. Various forms of wavelet transformations have been must wait for several epileptic seizure events before a new patient can be
utilized to extract time-frequency features from EEG signals. Saab and monitored. In addition, the rich array of information from other patients
Gotman used wavelet decomposition to extract features from scalp EEG is not fully utilized in optimizing these patient-specific models.
signals and achieved a sensitivity of 76.0% for the detection of seizures Many existing studies have screened intracranial EEG signals for
[34]. Discrete wavelet transformation (DWT) is a powerful tool for patterns representing epileptic seizures [21,38–41]. An intracranial EEG
analyzing time-frequency domain data, and can give a more flexible (iEEG) signal has a high signal/noise ratio, which facilitates the gener­
representation of signals. An extreme learning machine (ELM) was used ation of highly accurate seizure detection models [42]. However, an
to train a seizure detection model in Ref. [35]. In addition to iEEG monitors electrophysiological signals by placing sensing electrodes
time-frequency features, non-linear features can also be extracted from directly on the exposed surface of the brain, making it difficult to deploy
EEG signals. Kannathal et al. investigated the seizure detection problem outside of a clinical environment. Scalp EEG (sEEG) is a non-invasive
by integrating four entropy estimators: Shannon spectral entropy, technology for detecting electrophysiological signals from the surface
Renyi’s entropy, Kolmogorov-Sinai entropy and approximate entropy. of the head, and has a much higher noise level compared with iEEG
Their entropy-based model achieved a seizure classification accuracy of [42–44]. The detection of seizures from sEEG signals therefore remains a
90% [36]. Another study demonstrated that the non-linear features such challenge.
as correlation dimension (CD), largest Lyapunov exponent (LLE), Hurst
exponent (H), and entropy were accurate when used for seizure detec­ 2.2. Description of dataset
tion [37].
This study proposes a novel feature extraction algorithm in the time The CHB-MIT Scalp EEG database from PhysioNet was used in this
domain for the patient-independent seizure detection (pidSeizure) study [45]. This pediatric scalp EEG dataset was generated at Boston
problem, and demonstrates that the integrated modeling of various Children’s Hospital after the withdrawal of anti-seizure medications. A
feature sources can achieve satisfactory detection performance. detailed description of these samples can be found on the PhysioNet
Although much progress has been achieved in the epileptic seizure website.
detection problem, most prior studies have focused on training a sepa­ All pediatric patients were monitored at a sampling rate of 256 Hz
rate detection model for each patient. In this study, we hypothesize that with 16-bit quantization. Since most of the recordings used 23 EEG
a pidSeizure model might better serve epileptic seizure patients. Our channels, this study only selected samples with 23 channels. These
experimental data also suggest that the extracted features can be further channels were placed following the International Federation of Clinical
selected for better classification performance. The final detection model Neurophysiology 10–20 placement system, and the recordings were
utilizes both feature extraction and feature selection to deliver accurate analyzed accordingly, (i.e., FP1–F7, F7-T7, T7-P7, P7–O1, FP1–F3,
seizure detection performance. F3–C3, C3–P3, P3–O1, FZ-CZ, CZ-PZ, FP2–F4, F4–C4, C4–P4, P4–O2,

2
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Fig. 1. Illustration of the MinMaxHist algorithm. (a) Schematic line chart of an EEG signal, where Ti (i ¼ 1, 2, …, n) represents the absolute value of the difference
between a pair of neighboring peak and valley values. (b) Min-Max-Histogram, where the maximum and minimum values in the array T represent the left and right
limits of this histogram. The number of features returned by this algorithm is WinNum.

FP2–F8, F8-T8, T8-P8, P8–O2, P7-T7, T7-FT9, FT9-FT10, FT10-T8, and set the parameter WinNum ¼ 10, to describe the number of histogram
T8-P8, as in Ref. [46]). The last channel, T8-P8, was removed, since groups, in order to ensure that the new data distribution would not be
there were two of these channels, meaning that each EEG signal used in too scattered or too centralized. Next, we calculated the bin size as
this study contained 22 channels. The interictal regions of the record for (Max_T–Min_T)/WinNum to represent the width of each interval in the
patient chb12 were very short (average value 415 s), as were the ictal histogram. As shown in Fig. 1(b), the lateral axis was evenly divided into
regions of the record for patient chb16 (average value 8.6 s), and these 10 bins, with ranges [Min_T, Min_T þ bin), [Min_T þ bin, Min_Tþ2�bin),
two records were also excluded from further analysis. …, [Min_Tþ9�bin, Max_T]. The values for the vertical axis were the data
frequency of array T in each interval. The detailed algorithm was
2.3. Preprocessing of scalp EEG signals formulated as in Fig. 1. Hence, 10 MinMaxHist features were extracted
from a given EEG signal, and were denoted as MinMaxHist1, Min­
The EEG recordings for these patients were of various lengths and MaxHist2, …, MinMaxHist10.
showed different epileptic seizure times, so a preprocessing step was
conducted to ensure that each sample contained the same number of
2.5. Other time-domain features
features. First, a zero-phase 4th-order Butterworth filter was used to
process the 22 channels of each sample. The band filter was applied
Several time-domain features were summarized from the raw EEG
between 0.5 and 32 Hz, since the ictal EEG signals were observed to have
signals. The minimum, maximum, average, standard deviation and
significant abnormalities within this frequency range [34,47]. Following
variance of a given window of EEG signal X, where X¼<X1, X2, …, XN>,
this, 6-s [38] non-overlapping sliding windows were applied to the
were extracted as the features minX, maxX, meanX, stdX and varX.
epileptic seizure regions, which were taken as positive samples. There PN
jXtþ1 Xt j
were 1594 positive samples in total, retrieved from the 22 patients, as Another measurement totalVariance was defined as ðmaxX t¼1minXÞ�ðN 1Þ.
shown in Table 1. The interictal regions were usually much longer than The skewness of a data distribution X was defined as the expected
the ictal ones, and to ensure a balance between positive and negative �� �3 �
samples, six non-overlapping windows with 6-s lengths were randomly value of E X stdX
meanX
[49]. This feature was abbreviated as Skewness.
retrieved as negative samples from the interictal regions, excluding the Kurtosis is a shape descriptor of a probability distribution (abbreviated
initial 10-min windows. Interictal regions shorter than 10 min were not �� �4 �
used for retrieving negative samples. Each sample contained 22 chan­ as Kurs) and was defined as the expected value of E X stdX meanX
[46].
nels, giving 6 s � 256 Hz � 22 channels ¼ 33,792 voltage values.
The magnitude of a time-series signal X can be measured using the
root mean square value [50], i.e.,
2.4. Feature extraction: Min-Max histogram (MinMaxHist)
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u N
uX .
We observed that ictal EEG signals contained more fluctuations than RMSðXÞ ¼ t X 2t N (1)
interictal regions, and therefore proposed our novel feature extraction t¼1

algorithm, the Min-Max Histogram (MinMaxHist). Spikes and sharp

The largest peak value of X was defined as Peak(X) ¼ max(|maxX|, |
waveforms were repeatedly observed in the EEG signals at seizure onset
minX|), since the voltages of the EEG signal X may be positive or nega­
[48]. The purpose of the proposed MinMaxHist algorithm was to
tive values. The peak-to-average power ratio (PARS) was also defined as
quantitatively describe the waveform characteristics of these spikes and
PARS(X) ¼ [Peak(X)/RMS(X)]2 [51].
sharp waves by describing the scales of change in the EEG signals in the
feature extraction step and to distinguish seizure events from interictal
events based on these types of features. 2.6. Nonlinear features of EEG signals
The aim of the MinMaxHist algorithm is to create a histogram
describing a given EEG signal, as shown in Fig. 1. First, we constructed a EEG signals have also been described by nonlinear features in many
one-dimensional array T¼<T1, T2, …, Tn>, where Ti was the absolute existing studies, since they are inherently non-stationary, low-energy
value of the difference between a pair of neighboring peak and valley and transient [52]. Nonlinear descriptors can robustly capture the hid­
values, as shown in Fig. 1(a). Then, Max_T and Min_T were defined as the den patterns within such noisy signals [53,54].
maximum, and minimum values in array T, representing the right- and Entropy is a representative descriptor of the complexity of a time-
left-hand limits of the lateral axis in the histogram, respectively. We then series variable, and increases with an increase in this complexity [55].

3
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

The nonlinear dynamic analysis wavelet entropy (abbreviated as

WaveletEntropy) was used to described the EEG signals in this study,
and the Daubechies wavelet (db4) was used to decompose the EEG
signals into six frequency sub-bands [56–58].
Richman et al. proposed a similar measurement to describe the data
regularity of the EEG signals, abbreviated as SampleEntropy [59]. The
definition and parameter choices for SampleEntropy may be found in
Refs. [60,61].
Permutation entropy has demonstrated a very good anti-noise
capability and a low computational complexity [62,63], and was
therefore used in this study (abbreviated as PermutEntropy) to describe
the complexity of the EEG signals.
The Hurst Exponent (abbreviated as HurstExp) was proposed in 1951
to quantize the long-term correlations in a time-series variable [64]. The
HurstExp feature is usually calculated by the R/S method [65,66], and a
value in the range [0.5, 1] demonstrates the existence of long-range
dependence.
Non-static and non-stationary time-series variables can also be
described by fractal patterns, e.g., using detrended fluctuation analysis
(DFA) [67] and the Petrosian fractal dimension (PFD) [68]. DFA can
robustly describe the fractal characteristics and self-similarity in a
non-stationary time-series variable [67], while the fractal dimension
describes the self-similarity within a time-series variable and demon­
strates a better temporal resolution than Fourier analysis [68–70]. The
Petrosian version of fractal dimension (PFD) was utilized in this study, as
it has shorter time requirements [69,70].
The Hjorth parameters were proposed by Bo Hjorth in 1970 to
describe EEG signals [71], and we utilized the Hjorth mobility and
Hjorth complexity to detect seizures. The Hjorth mobility is defined as
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
HjorthMobilityðXðtÞÞ ¼ varðdXðtÞ = dtÞ = var ðXðtÞÞ (2)

where var(X(t)) is the variance of the variable X(t), while the Hjorth
complexity is defined as
HjorthComplexityðXðtÞÞ ¼ HjorthMobilityðdXðtÞ = dtÞ = HjorthMobilityðXðtÞÞ
(3)

2.7. Feature selection algorithms

Various feature selection algorithms were utilized to find a feature

subset that outperformed the list of 638 original features extracted from
the EEG signals. Filters are the simplest form of feature selection and
usually have the lowest time complexity [72]. The student’s t-test (Ttest)
is a widely-used statistical test for evaluating the difference in a feature
between two groups of samples [5,73,74]. The statistical p-value of a
Ttest evaluates the null hypothesis in which the distributions of a given
feature in two groups of features are identical to each other. Only fea­
tures with p � 0.05 are selected for further analysis. The rank sum test
(Utest) evaluates whether a feature was collected from two populations
with the same distribution, and does not assume that this feature follows
a normal distribution [75,76]. The same p-value threshold of 0.05 is
applied to select features using a Utest. In this study, we ranked the
Fig. 2. Seizure detection performance of 10 classifiers using raw EEG data
features in descending order based on their accuracy, calculated using a (Raw) and extracted features (FE). Values were calculated using 5-fold cross-
5-fold cross-validation strategy and two classifiers: a support vector validation for (a) accuracy, (b) MCC and (c) Kappa for each classifier on the two
machine (SVM) with a linear kernel [77] and the extreme gradient datasets (Raw and FE). (d) Difference in detection performance between the FE
boosting algorithm (XGB) [78]. The top 50% of the features were kept and raw data. Five performance measurements (Acc, Sn, Sp, MCC and Kappa)
for further analysis. These two filter feature selection strategies were were evaluated.
denoted as Acc-SVM and Acc-XGB, respectively.
Recursive feature elimination (RFE) is a popular feature selection be embedded in RFE as long as it returns the feature coefficients after the
framework [79] that iteratively trains a classification model over a classification model is trained. The most popular embedded classifier is
feature set using the embedded classifier and removes the feature with an SVM [80].
the lowest feature coefficient returned by the trained model. The feature Another feature elimination framework BackFS has also been uti­
subset with the best 5-fold cross-validation accuracy is returned for lized to remove features in an exhaustive way [5,81]. BackFS does not
further analysis. If multiple feature subsets achieve the same best ac­ assume that the coefficients from the trained model are positively
curacy, the one with the fewest features is returned. Any classifier may

4
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Table 2 many supervised learning algorithms. Another popular boosting algo­

Summary of the features extracted from the EEG signals. Two families of rithm is the gradient-boosted decision tree (GBDT), which can be used to
features were used in this study, from the time domain and the nonlinear solve both classification and regression problems [90–92]. Chen and
domain. The columns marked “Abbreviation” and “Description” give abbrevia­ Guestrin recently proposed a scalable tree boosting system (XGB), which
tions and basic descriptions of these extracted features, while the column has a very fast running speed and is a scalable end-to-end boosting
marked “Number” gives the number of features extracted for each row.
system [78].
Family Abbreviation Description Number

Time MinMaxHist1- Waveform 10 2.9. Performance measurements

domain MinMaxHist10
maxX, minX, meanX, stdX Maximum, minimum, mean, 5
and varX standard deviation and In this study, we used a 5-fold cross-validation strategy to evaluate
variance of a given variable X how well each model performed for a binary classification problem,
totalVariation, Skewness, Total variation, skewness, 5 using the measures of accuracy, sensitivity, and specificity. Let P and N
Kurs, RMS, and peak kurtosis, root mean square,
be the numbers of positive and negative samples, respectively. The
and peak
PARS peak-to-average power ratio 1 performance measurement accuracy (Acc) is defined as the percentage
of correctly detected samples, i.e., Acc¼(TP þ TN)/(P þ N), where TP
Non- WaveletEntropy, Sample Wavelet, sample and 3
linear Entropy and permutation entropy
(true positive) and TN (true negative) are the numbers of correctly
domain PermutEntropy measurements detected positive and negative samples, respectively.
HurstExp Hurst exponent 1 Sensitivity (Sn) and specificity (Sp) are defined as the ratios of
DFA Detrended fluctuation 1 correctly detected positive and negative samples, respectively. Hence,
analysis
sensitivity Sn ¼ TP/P, and specificity Sp ¼ TN/N. The numbers of
PFD Petrosian fractal dimension 1
HjorthMobility and Hjorth parameters: mobility 2 incorrectly detected positive and negative samples were defined as FP
HjorthComplexity and complexity (false positive) and FN (false negative), respectively.
The Matthews correlation coefficient (MCC) and Kappa are also used
to evaluate the performance of our approach. The range of MCC is be­
correlated with the importance of each feature. In each round of the tween zero and one, with one representing the best classification result.
iteration, BackFS removes a feature if this achieves the smallest decrease MCC is defined as
in the accuracy of the embedded classifier. Any classifier can be
embedded, and BackFS is used as a wrapper framework. MCC ¼ TP � TN FP
.pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
� FN ðTP þ FPÞ � ðTP þ FNÞ � ðTN þ FPÞ � ðTN þ FNÞ (4)
2.8. Classification algorithms
The Kappa metric (also called Cohen’s kappa) is a coefficient for the
Ten popular classifiers were utilized in this study. SVM is a fast and qualitative (categorical) samples [93,94] that takes values in the range
effective classifier in many cases of binary classification problems, and [0, 1]. Kappa is defined as
calculates a hyperplane to maximize the margins between two groups of Kappa ¼ ðAcc pe Þ=ð1 pe Þ (5)
samples within the hyperspace, using a given kernel function [77]. In

pe ¼ ½ðTP � FPÞ þ ðTP � FNÞ þ ðTN � FPÞ þ ðTN � FNÞ� =ðTP þ TN þ FP þ FNÞ2 (6)

this study, we utilized two popular kernel functions for SVM, the
Gaussian and linear kernels. These two models are abbreviated as
rbfSVM and lSVM, respectively.
Two simple classifiers were evaluated, the k-nearest neighbors
(KNN) and naive Bayes (NBayes) approaches. KNN utilizes the essence 3. Results and discussion
of a classification problem (i.e., the intra-class similarity) and assigns a
class label to a sample based on majority voting by the sample’s k 3.1. Extraction of features from raw EEG data
nearest neighbors [82], where the default value of k is 5. NBayes as­
sumes inter-feature independence and calculates the likelihood of a Equal numbers of positive and negative samples were collected from
sample being in each class, based on Bayes’ theorem [83]. This study each patient for further analysis, as shown in Table 1. Hence, the number
used a Gaussian kernel NBayes approach. A logistic regression (LR) bi­ of sample pairs collected from each patient was the minimum of the
nary classifier was evaluated in this study; both L1-and L2-regulariza­ numbers of positive and negative samples for this patient. A total of
tions were used as loss functions to avoid overfitting, and are denoted as 1260 pairs of samples were collected from all patients.
LR-L1 and LR-L2, respectively [84]. Two families of features were extracted from each channel of a given
Tree-based learning algorithms form another popular family of EEG sample, i.e., the time domain and the nonlinear domain, as shown
classifiers. The classification and regression tree (CART) is a widely used in Table 2. We calculated 21 time-domain features from each 6 s EEG
classifier with low model complexity and fast running speed [85]. A sample. The MinMaxHist feature extraction algorithm was used to
CART-based decision tree classifier was evaluated, referred to here as generate 10 features (MinMaxHist1 to MinMaxHist10) from a given EEG
the DTree algorithm. sample. Five basic statistics metrics were also calculated: the maximum
Boosting is an integration strategy that can create a strong classifier (maxX), minimum (minX), mean (meanX), standard deviation (stdX)
by summarizing the results of multiple weak classifiers [86]. Adaptive and variance (varX) for this sample, giving a one-dimensional vector
boosting (AdaBoost) was proposed by Freund and Schapire [87] in 1997 with 6 s � 256 Hz ¼ 1536 data points. Five further time-domain metrics
to generate a boosting classifier in an adaptive manner [88,89]. The were also defined to describe the EEG signals: the total variance
AdaBoost algorithm is a meta-algorithm and can be integrated with

5
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Fig. 3. Heatmap of the performance of 10 classifiers on five feature sets. The FE feature set was the original list of all 638 features extracted from an EEG sample.
The other four feature sets were selected using four filters: Ttest, Utest, Acc-SVM and Acc-XGB. The (a) accuracy, (b) MCC and (c) Kappa were calculated using each of
the 10 classifiers with a 5-fold cross-validation strategy. The column marked AvgC gives the average value for each classifier, while the column marked AvgFS gives
the average value for each feature set.

(totalVariance), skewness (Skewness), kurtosis (Kurs), root mean square in order to show the differences between the two sets of results. The
(RMS) and peak (Peak). The peak-to-average power ratio (PARS) was experimental results showed that except for rbfSVM, the classifiers
also calculated as a feature. achieved better seizure detection performance, with average improved
We also calculated eight nonlinear features to describe the EEG values using the extracted features of 0.0918, 0.1819 and 0.1859 for
signals, as shown in Table 2. Three types of entropy were calculated: the Acc, MCC and Kappa, respectively, compared to using the raw EEG data.
wavelet entropy (WaveletEntropy), sample entropy (SampleEntropy), The largest improvement in Acc (0.1900) was achieved by the KNN
and permutation entropy (PermutEntropy). Three other widely used classifier. The rbfSVM classifier performed in the same way as a random
nonlinear features were also calculated: the Hurst exponent (HurstExp), classifier, with a value of Acc of around 0.5000 using both raw EEG data
detrended fluctuation analysis (DFA), and Petrosian fractal dimension and the extracted features, meaning that the small improvement in ac­
(PFD). The Hjorth parameters of mobility (HjorthMobility) and curacy for rbfSVM between the dataset FE and raw EEG may be simply
complexity (HjorthComplexity) were also used due to their discrimina­ due to chance.
tive capabilities. It was therefore essential to extract features from the EEG signals in
the time and nonlinear domains in order to achieve better seizure
3.2. Detection of seizures using raw EEG data and extracted features detection performance.
The DFA and Hurst exponent features describe the general patterns
Overall, the extracted features achieved better seizure detection in seizure EEG signals [95,96], and have been widely used to detect
performance than the raw EEG data, as shown in Fig. 2. The best values epileptic seizures based on EEG data [96,97]. In this study, we evaluated
of Acc, MCC, and Kappa for seizure detection were achieved by the the classification performance of these two feature types alone and
classifiers on the extracted features (dataset FE), as shown in Fig. 2(a–c). compared it with all feature types. The XGB classifier achieved the best
The GBDT and XGB boosting classifiers were the best performing. GBDT classification performance (Acc ¼ 0.7901, MCC ¼ 0.6057, Kappa ¼
achieved the best values of Acc (0.8210) and MCC (0.6761), while the 0.5802) using these two feature types, i.e., DFA and HurstExp. When all
second-best classifier achieved values of 0.8190 for Acc and 0.6723 for types of features were used, the XGB classifier again achieved the best
MCC. The third best classifier, AdaBoost, achieved values of around 0.80 classification model, with Acc ¼ 0.8190, MCC ¼ 0.6723 and Kappa ¼
for Acc and 0.6138 for MCC. The other classifiers did not achieve values 0.6381. Both the experimental data and the literature suggest that these
of Acc larger than 0.80 or MCC larger than 0.6. These results demon­ two general features (DFA and HurstExp) can contribute to the problem
strate that the boosting classifiers gave good performance for this of epileptic seizure detection using EEG signals.
problem.
We also demonstrated that this step achieved better detection per­ 3.3. Finding seizure-associated features using filters
formance by utilizing the features extracted from the EEG signals. Fig. 2
(d) displays the results of dataset FE minus the results of raw EEG data, Due to the inherently noisy nature of the data, it is always

6
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Fig. 4. Comparison of the detection perfor­

mance of filters and filter-embedded RFE algo­
rithms. The FE feature set contained a list of all
638 features extracted from the EEG signals. The
other four feature sets were first selected using four
filters: Ttest, Utest, Acc-SVM and Acc-XGB. The
column marked Filter gives the seizure detection
performance for the original and filter-selected
features, while the column marked FilterRFE
gives the detection performance for features that
have been further screened using a filter-embedded
RFE strategy. A filter was embedded for the FE-
based RFE process, and the detection performance
was evaluated based on (a) Acc, (b) Sn and (c) Sp,
calculated using 5-fold cross-validation and the
best classifier (XGB) identified in the previous
section. The numbers of features are shown in (d).

challenging to accurately identify phenotype-associated patterns in EEG were independent of each other and the features were ranked based on
signals [43,44,52]. Some EEG-extracted features may be removed in the discriminative power of each feature individually [57,72,99]. The
order to refine EEG-based detection models for various biomedical is­ features were scaled to a [0, 1] normal distribution.
sues, e.g., treatment responses for depressive disorders [98]. Fig. 3 demonstrates that the use of these four simple filters can
First, we evaluated four filter algorithms for finding seizure- improve the seizure detection model for nine of the 10 classifiers. The
associated features. Each sample contained 22 channels, and 29 fea­ largest improvement of 0.1320 in the value of accuracy was achieved by
tures were calculated for each of these channels, meaning that the the lSVM model with the Ttest-selected features. No improvement was
original FE feature set for each sample in this study had 22 � 29 ¼ 638 achieved by the four filter algorithms for the NBayes classifier. The
features. The Ttest and Utest filters were used to select 320 and 419 GBDT boosting classifier needed all 638 features to achieve its best
features, respectively, with p � 0.05, while Acc-SVM and Acc-XGB were classification accuracy. It was also interesting to observe that the three
used to select 319 features. All four filters assumed that the features top-performing models were trained with the ensemble classifiers, i.e.,

7
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Table 3
List of the 30 features selected by BackFS. The columns marked “No.” and
“Channel” give the IDs and channels for the features, while the feature names are
given in the column marked “FeatureName”.
No. Channel FeatureName No. Channel FeatureName

1 P7_O1 MinMaxHist8 16 C3_P3 MinMaxHist2

2 F4_C4 MinMaxHist2 17 T7_FT9 MinMaxHist7
3 T8_P8 MinMaxHist4 18 P7_O1 MinMaxHist2
4 P7_O1 MinMaxHist3 19 P7_T7 MinMaxHist3
5 FT10_T8 MinMaxHist2 20 T7_FT9 MinMaxHist2
6 FT9_FT10 MinMaxHist4 21 T8_P8 RMS
7 FP1_F7 MinMaxHist5 22 T7_FT9 HjorthMobility
8 F7_T7 MinMaxHist2 23 T7_P7 HjorthMobility
9 T7_P7 MinMaxHist3 24 FT9_FT10 MinMaxHist5
10 FP2_F4 MinMaxHist1 25 T7_P7 meanX
11 FP2_F8 MinMaxHist2 26 T7_FT9 PFD
12 C4_P4 MinMaxHist1 27 P3_O1 DFA
13 P8_O2 MinMaxHist2 28 FP2_F4 PermutEntropy
14 F3_C3 MinMaxHist2 29 FT9_FT10 totalVariation
15 P8_O2 MinMaxHist7 30 FT9_FT10 RMS

model training.
FE-RFE achieved the best seizure detection accuracy (Acc ¼ 0.8512),
as shown in Fig. 4(a). The RFE strategy improved the results for all four
filter algorithms, with a minimum improvement in Acc of 0.0198. It was
also interesting to observe that the detection performance Acc/Sn/Sp
was improved by the RFE strategy in all cases, as shown in Fig. 4(b) and
(c). The RFE strategy also reduced the number of features by at least
91%, as shown in Fig. 4(d), thus demonstrating its power to improve
features extracted from an EEG for better classification performance.

3.5. Selection of features using classifier-embedded RFEs

Six classifiers were evaluated in terms of their performance as the

embedded classifier in the RFE framework, as shown in Fig. 5. The three
classifiers lSVM, LR-L1 and LR-L2 performed the worst, with an average
accuracy of only about 0.7690, while the other three classifiers achieved
values of at least 0.8300. XGB and GBDT achieved the best detection
accuracy of 0.8512 on the FE features. A larger difference between Sn
and Sp was seen for GBDT (Sn ¼ 0.7992 and Sp ¼ 0.9032), while XGB
achieved values of 0.8056 and 0.8968 for Sn and Sp, respectively. XGB
was identified as the best embedded classifier, since it used only 11
features to achieve the same seizure detection accuracy as GBDT, which
used 19.
Fig. 5. Evaluation of the RFE strategy using various embedded classifiers.
The detection performance was evaluated based on (a) Acc, (b) Sn, (c) Sp and
(d) FNum (feature number). The six embedded classifiers were lSVM, LR-L1, 3.6. Removal of redundant features using BackFS
LR-L2, AdaBoost, GBDT and XGB.
The union of the 11 detected features and MinMaxHist features was
XGB (Acc ¼ 0.8210, MCC ¼ 0.6779 and Kappa ¼ 0.6413), GBDT (Acc ¼ refined using the BackFS strategy, as shown in Table 3. Thirty features
0.8210, MCC ¼ 0.6761 and Kappa ¼ 0.6429) and AdaBoost (Acc ¼ were chosen to achieve the overall best values of Acc, MCC and Kappa of
0.8004, MCC ¼ 0.6332 and Kappa ¼ 0.6008). Although both XGB and 0.8627, 0.7504 and 0.7254, respectively, using the XGB classifier as
GBDT achieved the best value of 0.8210 for Acc, XGB (320 features) shown in Fig. 6. The values of Acc, MCC and Kappa for MinMaxHist
achieved a better MCC of 0.6779 and used only about half of the features features were 0.0464, 0.084 and 0.0929 higher, respectively, than for
compared with GBDT (638 features). The principle of Occam’s Razor classification without the MinMaxHist features. The final results for Acc,
suggests that a simple model with a similar detection performance MCC and Kappa showed that the proposed method achieved a more
should be preferred over a complex one [100,101]. Hence, the experi­ accurate classification. From Fig. 6, it can be seen that the best results
mental data suggest that a feature selection step can be utilized to were obtained when XGB was used as the classifier for the input feature
further improve the seizure detection model based on the features matrix. The second best classifier, GBDT, was also an ensemble classi­
extracted from an EEG. fier. This observation suggests that ensemble classifiers may be good
choices for modeling EEG-based classification problems, as shown in
Fig. 6. The NBayes classifier achieved imbalanced values of Sn ¼ 0.1540
3.4. Finding a better feature list using filter-embedded RFEs and Sp ¼ 0.9381, suggesting that the features extracted from the EEG did
not facilitate the calculation of the Bayesian probabilities.
This section evaluates how the features can be refined by filter- The final list of 30 features consisted of 21 MinMaxHist features and
embedded RFE strategies, as shown in Fig. 4. The best classifier identi­ nine other features, as shown in Table 3. The nine other features were
fied in the previous section, XGB, was then utilized to evaluate a given shared with the best subset of 11 features selected in the previous sec­
feature subset. The RFE strategy eliminates one feature per iteration of tion. The RMS for channel T7_P7 and PermutEntropy for channel P7_O1

8
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Fig. 6. Seizure detection performance of eight classifiers on the 30 features selected by BackFS. The horizontal axis lists the classifiers, while the vertical axis
shows the detection performance measurements.

were not selected. This demonstrated that the MinMaxHist features specific methods, it has a wider range of applications.
extracted from the EEG signals exhibited the seizure-specific EEG pat­
terns, and had potential applications in the EEG-based detection of other 3.9. Comparison with existing studies
diseases.
Various machine learning algorithms for automatic seizure detection
3.7. Performance of patient-specific seizure detection have been proposed by other researchers. We carried out a compre­
hensive comparison between the model proposed in this study and
Many existing EEG-based seizure detection studies have focused on existing models, as shown in Table 4.
patient-specific models [38], in which a separate model is trained for The open dataset CHB-MIT has been widely used to evaluate seizure
each patient using EEG data from that patient. This approach is expected detection algorithms. Vidyaratne et al. [38] used the feature extraction
to give a higher seizure detection accuracy than the patient-independent algorithm harmonic wavelet packet transform (HWPT) and a classifier
model investigated in this study. We therefore carried out a comparative relevance vector machine (RVM) to achieve a seizure detection sensi­
experiment to investigate whether the optimized feature selection pro­ tivity of 0.96. Raghu et al. [103] used DWT-based sigmoid entropy in the
cedure developed in this study can help to improve the classic time and frequency domains to extract features, and applied the SVM
patient-specific approach. Fig. 7 illustrates that the procedure developed classifier, which achieved a sensitivity of 0.9421. Kaleem et al. [104]
in this study improved the accuracy of detection in the patient-specific detected seizures based on signal-derived empirical mode decomposi­
approach, with an average improvement of 0.0233 in the value for tion (EMD)-based dictionary approach, and obtained average values for
Acc, as shown in Fig. 7(a). The values of Sn, Sp, MCC and Kappa were accuracy, sensitivity and specificity of 0.929, 0.943 and 0.915, respec­
improved by 0.0346, 0.0141, 0.0509 and 0.0546, respectively, as shown tively. Xiang et al. [105] used fuzzy entropy to extract features and
in Fig. 7(c–f). A total of 638 features were extracted from the EEG sig­ achieved average values for accuracy, sensitivity and specificity of
nals, and our feature selection procedure selected at most six of these for 0.9831, 0.9827 and 0.9836, respectively. In comparison, a model
patient-specific seizure detection models, for all patients except chb11. trained with the proposed method for each patient can achieve average
Our procedure selected 51 features, and achieved a seizure detection values of accuracy, sensitivity and specificity of 0.9828, 0.9893 and
accuracy that was only 0.0071 lower than that of the “Raw” model, as 0.9764, respectively. These experimental results suggest that the pro­
shown in Fig. 7(b). The average values of Acc, Sn, Sp, MCC and Kappa posed method is superior. Although the accuracy reported by Bhatta­
were 0.9828, 0.9893, 0.9764, 0.9674 and 0.9655, respectively. The charyya et al. [106] was slightly better than in our study, their work was
value of Sn (0.9893) obtained in our study was higher than the value of patient-dependent, and cannot be easily applied to new patients.
Sn (0.96) reported in the prior study [38]. From a comparison with patient-independent studies that used the
CHB-MIT dataset, we found that our proposed method achieved better
results, as shown in Table 4. Our method required on 6 s to analyze and
3.8. Performance of semi-patient-specific seizure detection
process the raw data, whereas in Ref. [46], the reported sampling time
was 60 s. Although the values for the sensitivity and specificity of these
Haoqu and Gotman proposed a semi-patient-specific approach in
authors’ results were 0.93 and 0.94, the SMOTE over-sampling algo­
their study, with results that were better than non-specific but worse
rithm was used for data processing, and the sampling time was almost 10
than patient-specific models [102]. Their experiment used
times as long as ours. Before the SMOTE method was used, the sensi­
semi-patient-specific information including a seizure from one patient
tivity and specificity of the experimental results reported in Ref. [46]
and interictal EEG signals from many other patients to train a classifier
were both 0.88. In comparison, if the sampling time was extended to 60 s
for the seizure patient. The classifier was trained in this way for each
in our method, the values of sensitivity and specificity would reach
patient. We used the algorithm proposed in this work to carry out a
0.8896 and 0.9310. In Ref. [107], the author used a neural network
semi-patient-specific experiment and obtained a value of 0.8849 for the
classifier to analyze the dataset, and obtained a sensitivity value of 0.85;
accuracy of detection for the onset of epileptic seizures. This result was
however, the sampling time reported in this paper was 30 s. With an
higher than the non-patient-specific model and lower than the
equivalent sampling time, the sensitivity and specificity of our algorithm
patient-specific model. This conclusion was consistent with that of
reached 0.8967 and 0.9393, respectively, i.e. our method gives superior
Haoqu and Gotman, and our method achieved a higher accuracy. The
results.
non-specific method used EEG signals from all patients to train each
classifier. Although this method did not perform as well as the other two

9
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Fig. 7. Improving the classic patient-specific seizure detection model. An optimized feature selection procedure was applied to the classic model of patient-
specific seizure detection. The detection performance was measured based on (a) Acc, (b) FNum (number of features), (c) Sn, (d) Sp, (e) MCC and (f) Kappa. The
data series “Raw” refers to patient-specific seizure detection using all the extracted features, while the data series “FS” refers to detection models with features
selected using the feature selection procedure optimized in the previous sections.

3.10. Evaluation of the proposed method on a new corpus whom the causes of the disease were different. There are both eye
movements and artifacts in the EEG signals, and we can therefore treat
We also evaluated the proposed method on an independent dataset. this dataset as a real clinical dataset. We evaluated our method on this
The THU EEG corpus was released by the American Clinical Neuro­ corpus, and achieved values for seizure prediction of Acc, Sn, and Sp of
physiology Society, and contains many types of EEG signals for use in 0.8021, 0.80, and 0.8088, respectively. Although the performance of our
various studies [108]. We selected the THU EEG Seizure corpus, a subset method in this case was slightly lower than that for the CHB-MIT
of the THU EEG Corpus with files that are known to contain seizure dataset, this corpus contained more patients and more causes of epi­
events. There were 929 EEG files in this corpus, 217 of which were lepsy, the patients had a larger age range, and the EEG signals contained
samples of epileptic seizures. The EEG signals in this corpus were more noise and artifacts. This independent corpus was therefore similar
collected from 43 real epileptic patients, and had not been processed. to a realistic clinical dataset, and the proposed method demonstrated a
The corpus contained male and female patients of various ages, for reasonable seizure prediction performance.

10
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

Table 4
Comparison with other existing studies. The detection performance was measured based on Acc, Sn, Sp, MCC and Kappa. “-“ indicates that the value is not
mentioned in the literature. The column marked “Category” gives the type of approach used in each paper.
Study Category Acc Sn Sp MCC Kappa

Vidyaratne et al. patient-dependent - 0.9600 - - -

Raghu et al. patient-dependent - 0.9421 - – 0.94
Kaleem et al. patient-dependent 0.9291 0.9427 0.9155 – –
Xiang et al. patient-dependent 0.9831 0.9827 0.9836 – –
Bhattacharyya et al. patient-dependent 0.9941 0.9791 0.9957 – –
Fergus et al. patient-independent - 0.8800 0.8800 – –
Thodoroff et al. patient-independent - 0.8500 - - -
Current study patient-dependent 0.9828 0.9893 0.9764 0.9674 0.9656
Current study (6 s) patient-independent 0.8627 0.8032 0.9222 0.7504 0.7254
Current study (30 s) patient-independent 0.9180 0.8967 0.9393 0.8428 0.8356
Current study (60 s) patient-independent 0.9103 0.8896 0.9310 0.8275 0.8207

4. Conclusions References

Epileptic seizure is a nervous system disorder that does not directly [1] A.V. Bumanglag, R.S. Sloviter, No latency to dentate granule cell epileptogenesis
in experimental temporal lobe epilepsy with hippocampal sclerosis, Epilepsia 59
cause lethal damage to patients; however, seizures can substantially (2018) 2019–2034.
impact the quality of life of patients, and can give rise to major mental [2] C. Tang, Y. Gu, H. Wang, H. Wu, Y. Wang, Y. Meng, Z. Han, Y. Gu, W. Ma,
pressures. The accurate detection of seizures may provide vital infor­ Z. Jiang, Y. Song, M. Na, D. Lu, Z. Lin, Targeting of microRNA-21-5p protects
against seizure damage in a kainic acid-induced status epilepticus model via
mation to help both patients and their family members to maintain their PTEN-mTOR, Epilepsy Res. 144 (2018) 34–42.
daily routines and quality of life. [3] J. Corsini, L. Shoker, S. Sanei, G. Alarcon, Epileptic seizure predictability from
This study focused on the patient-independent seizure detection scalp EEG incorporating constrained blind source separation, IEEE (Inst. Electr.
Electron. Eng.) Trans. Biomed. Eng. 53 (2006) 790–799.
problem, which is more challenging to solve than the classic patient- [4] A. Shoeb, H. Edwards, J. Connolly, B. Bourgeois, S.T. Treves, J. Guttag, Patient-
specific seizure detection problem. Many existing studies have focused specific seizure onset detection, Epilepsy Behav. : E&B 5 (2004) 483–498.
on patient-specific seizure detection, which requires dozens of seizures [5] Y. Zhang, S. Yang, Y. Liu, Y. Zhang, B. Han, F. Zhou, Integration of 24 feature
types to accurately detect and predict seizures using scalp EEG signals, Sensors 18
to be recorded for a specific patient before a seizure detection model for
(2018).
that patient can be well trained. Our patient-independent seizure [6] G.L. Birbeck, R.D. Hays, X. Cui, B.G. Vickrey, Seizure reduction and quality of life
detection model can be easily deployed for a given patient without improvements in people with epilepsy, Epilepsia 43 (2002) 535–538.
requiring recordings of previous seizures. [7] X. Zhao, Z. Zhou, W. Zhu, H. Xiang, Role of conventional magnetic resonance
imaging in the screening of epilepsy with structural abnormalities: a pictorial
By integrating the proposed feature extraction algorithm (Min­ essay, Am. J. Nucl. Med. Mol. Imaging 7 (2017) 126–137.
MaxHist) with several existing algorithms, we generated a seizure [8] G.P. Winston, The role of magnetic resonance imaging techniques in the
detection model that achieved classification values of 0.8627, 0.7504 diagnosis, surgical treatment and biological understanding of epilepsy, Quant.
Imag. Med. Surg. 5 (2015) 186–187.
and 0.7254 for Acc, MCC and Kappa, respectively. This model also used [9] P. Fu, L. Wei, F. Zhang, J. Gao, J. Jing, W. Wu, H. Liu, Added value of NeuroGam
only 30 features, and we did not consider other possible features. It is software analysis in single photon emission computed tomography localization
possible that the integration of more feature extraction methods will diagnosis of epilepsy in interictal stage, Med. Sci. Mon. Int. Med. J. Exp. Clin. Res.
24 (2018) 1494–1501.
improve the experimental results, and further combinations of features [10] W.T. Kerr, S.T. Nguyen, A.Y. Cho, E.P. Lau, D.H. Silverman, P.K. Douglas, N.
may also be tried in the future. The optimized feature selection pro­ M. Reddy, A. Anderson, J. Bramen, N. Salamon, J.M. Stern, M.S. Cohen,
cedure in this study was also shown to improve the classic patient- Computer-aided diagnosis and localization of lateralized temporal lobe epilepsy
using interictal FDG-PET, Front. Neurol. 4 (2013) 31.
specific seizure detection model. [11] A. Kumar, F. Semah, H.T. Chugani, W.H. Theodore, Epilepsy diagnosis: positron
emission tomography, Handb. Clin. Neurol. 107 (2012) 409–424.
Data availability [12] C. Minardi, R. Minacapelli, P. Valastro, F. Vasile, S. Pitino, P. Pavone, M. Astuto,
P. Murabito, Epilepsy in children: from diagnosis to treatment with focus on
emergency, J. Clin. Med. 8 (2019).
The dataset used in this experiment was downloaded on May 4, [13] J. Nizard, J.P. Lefaucheur, M. Helbert, E. de Chauvigny, J.P. Nguyen, Non-
2018, and can be obtained from https://www.physionet.org/content/ch invasive stimulation therapies for the treatment of refractory pain, Discov. Med.
bmit/1.0.0/. 14 (2012) 21–31.
[14] L. Yang, C. Wilke, B. Brinkmann, G.A. Worrell, B. He, Dynamic imaging of ictal
oscillations using non-invasive high-resolution EEG, Neuroimage 56 (2011)
Declaration of competing interest 1908–1917.
[15] S. Koelstra, I. Patras, Fusion of facial expressions and EEG for implicit affective
tagging, Image Vis Comput. 31 (2013) 164–174.
None Declared. [16] S.R. Sinha, L. Sullivan, D. Sabau, D. San-Juan, K.E. Dombrowski, J.J. Halford, A.
J. Hani, F.W. Drislane, M.M. Stecker, American clinical neurophysiology society
Acknowledgements guideline 1: minimum technical requirements for performing clinical
electroencephalography, J. Clin. Neurophysiol. 33 (2016) 303–307.
[17] A. Emami, N. Kunii, T. Matsuo, T. Shinozaki, K. Kawai, H. Takahashi,
This work was supported by the Jilin Provincial Key Laboratory of Autoencoding of long-term scalp electroencephalogram to detect epileptic seizure
Big Data Intelligent Computing (20180622002JC), the Education for diagnosis support system, Comput. Biol. Med. 110 (2019) 227–233.
[18] N.S. Frolov, V.V. Grubov, V.A. Maksimenko, A. Luttjohann, V.V. Makarov, A.
Department of Jilin Province (JJKH20180145KJ), and the startup grant N. Pavlov, E. Sitnikova, A.N. Pisarchik, J. Kurths, A.E. Hramov, Statistical
of the Jilin University. This work was also partially supported by the properties and predictability of extreme epileptic events, Sci. Rep. 9 (2019) 7243.
Bioknow MedAI Institute (BMCPP-2018-001), the High Performance [19] R. San-Segundo, M. Gil-Martin, L.F. D’Haro-Enriquez, J.M. Pardo, Classification
of epileptic EEG recordings using signal transforms and convolutional neural
Computing Center of Jilin University, and by the Fundamental Research
networks, Comput. Biol. Med. 109 (2019) 148–158.
Funds for the Central Universities, JLU. The insightful comments from [20] X. Fan, N. Gaspard, B. Legros, F. Lucchetti, R. Ercek, A. Nonclercq, Automated
the three anonymous reviewers were much appreciated. epileptic seizure detection based on break of excitation/inhibition balance,
Comput. Biol. Med. 107 (2019) 30–38.
[21] M. Zhou, C. Tian, R. Cao, B. Wang, Y. Niu, T. Hu, H. Guo, J. Xiang, Epileptic
seizure detection based on EEG signals and CNN, Front. Neuroinf. 12 (2018) 95.
[22] E.M. Ventouras, E.A. Monoyiou, P.Y. Ktonas, T. Paparrigopoulos, D.G. Dikeos, N.
K. Uzunoglu, C.R. Soldatos, Sleep spindle detection using artificial neural

11
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

networks trained with filtered time-domain EEG: a feasibility study, Comput. [53] U.R. Acharya, O. Faust, N. Kannathal, T. Chua, S. Laxminarayan, Non-linear
Methods Progr. Biomed. 78 (2005) 191–207. analysis of EEG signals at various sleep stages, Comput. Methods Progr. Biomed.
[23] N. Gursel Ozmen, L. Gumusel, Y. Yang, A biologically inspired approach to 80 (2005) 37–45.
frequency domain feature extraction for EEG classification, Comput. Math [54] L. Gupta, A.M. Upadhye, Non-linear alignment of neural net outputs for partial
Methods Med. 2018 (2018) 9890132. shape classification, Pattern Recogn. 24 (1991) 943–948.
[24] J. Birjandtalab, M. Baran Pouyan, D. Cogan, M. Nourani, J. Harvey, Automated [55] Y. Chen, Y. Li, R. Narayan, A. Subramanian, X. Xie, Gene expression inference
seizure detection using limited-channel EEG and non-linear dimension reduction, with deep learning, Bioinformatics 32 (2016) 1832–1839.
Comput. Biol. Med. 82 (2017) 49–58. [56] H. Sharma, R. Raj, M. Juneja, EEG signal based classification before and after
[25] S. Geng, W. Zhou, Q. Yuan, D. Cai, Y. Zeng, EEG non-linear feature extraction combined Yoga and Sudarshan Kriya, Neurosci. Lett. 707 (2019) 134300.
using correlation dimension and Hurst exponent, Neurol. Res. 33 (2011) [57] F. Han, B. Liu, J. Zhu, B. Zhang, Algorithm design for edge detection of high-
908–912. speed moving target image under noisy environment, Sensors 19 (2019).
[26] A.H. Ansari, P.J. Cherian, A. Caicedo, G. Naulaers, M. De Vos, S. Van Huffel, [58] T. Zoughi, R. Boostani, M. Deypir, A wavelet-based estimating depth of
Neonatal seizure detection using deep convolutional neural networks, Int. J. anesthesia, Eng. Appl. Artif. Intell. 25 (2012) 1710–1722.
Neural Syst. 29 (2019) 1850011. [59] J.S. Richman, J.R. Moorman, Physiological time-series analysis using
[27] Y. Li, W. Cui, M. Luo, K. Li, L. Wang, Epileptic seizure detection based on time- approximate entropy and sample entropy, Am. J. Physiol. Heart Circ. Physiol. 278
frequency images of EEG signals using Gaussian mixture model and gray level Co- (2000) H2039–H2049.
occurrence matrix features, Int. J. Neural Syst. 28 (2018) 1850003. [60] Y. Song, P. Lï� , A new approach for epileptic seizure detection: sample entropy
[28] Q. Yuan, W. Zhou, L. Zhang, F. Zhang, F. Xu, Y. Leng, D. Wei, M. Chen, Epileptic based feature extraction and extreme learning machine, 03, J. Biomed. Sci. Eng.
seizure detection based on imbalanced classification and wavelet packet (2010) 12, 06.
transform, Seizure 50 (2017) 99–108. [61] U.R. Acharya, S.V. Sree, A.P. Alvin, R. Yanti, J.S. Suri, Application of non-linear
[29] H. Chao, L. Dong, Y. Liu, B. Lu, Emotion recognition from multiband EEG signals and wavelet based features for the automated identification of epileptic EEG
using CapsNet, Sensors 19 (2019). signals, Int. J. Neural Syst. 22 (2012) 1250002.
[30] H. Chao, H. Zhi, L. Dong, Y. Liu, Recognition of emotions using multichannel EEG [62] X. Chen, N.-D. Jin, A. Zhao, Z.-K. Gao, L.-S. Zhai, B. Sun, The experimental signals
data and DBN-GC-based ensemble deep learning framework, Comput. Intell. analysis for bubbly oil-in-water flow using multi-scale weighted-permutation
Neurosci. 2018 (2018) 9750904. entropy, Phys. Stat. Mech. Appl. 417 (2015) 230–244.
[31] A. Sharmila, P. Geethanjali, A review on the pattern detection methods for [63] A.A. Bruzzo, B. Gesierich, M. Santi, C.A. Tassinari, N. Birbaumer, G. Rubboli,
epilepsy seizure detection from EEG signals, Biomed. Tech. (2019). Permutation entropy to detect vigilance changes and preictal states from scalp
[32] N. Padfield, J. Zabalza, H. Zhao, V. Masero, J. Ren, EEG-based brain-computer EEG in epileptic patients. A preliminary study, Neurol. Sci. 29 (2008) 3–9, official
interfaces using motor-imagery: techniques and challenges, Sensors 19 (2019). journal of the Italian Neurological Society and of the Italian Society of Clinical
[33] A.J. Gabor, R.R. Leach, F.U. Dowla, Automated seizure detection using a self- Neurophysiology.
organizing neural network, Electroencephalogr. Clin. Neurophysiol. 99 (1996) [64] H.E. Hurst, Long-term storage capacity of reservoirs, Trans. Am. Soc. Civ. Eng.
257–266. 116 (1951) 770–799.
[34] M.E. Saab, J. Gotman, A system to detect the onset of epileptic seizures in scalp [65] B.A. Carreras, B.P. van Milligen, M.A. Pedrosa, R. Balbin, C. Hidalgo, D.
EEG, Clinical neurophysiology, Off. J. Int. Feder. Clin. Neurophysiol. 116 (2005) E. Newman, E. Sanchez, M. Frances, I. Garcia-Cortes, J. Bleuel, M. Endler,
427–442. C. Riccardi, S. Davies, G.F. Matthews, E. Martines, V. Antoni, A. Latten,
[35] J.L. Song, W.F. Hu, R. Zhang, Automated detection of epileptic EEGs using a novel T. Klinger, Self-similarity of the plasma edge fluctuations (vol 5, pg 3632, 1998),
fusion feature and extreme learning machine, Neurocomputing 175 (2016) Phys. Plasmas 10 (2003), 4591-4591.
383–391. [66] M. Nurujjaman, R. Narayanan, A.S. Iyengar, Comparative study of nonlinear
[36] N. Kannathal, M.L. Choo, U.R. Acharya, P.K. Sadasivan, Entropies for detection of properties of EEG signals of normal persons and epileptic patients, Nonlinear
epilepsy in EEG, Comput. Methods Progr. Biomed. 80 (2005) 187–194. Biomed. Phys. 3 (2009) 6.
[37] N. Kannathal, U.R. Acharya, C.M. Lim, P.K. Sadasivan, Characterization of EEG–a [67] K. Ivanova, M. Ausloos, Application of the detrended fluctuation analysis (DFA)
comparative study, Comput. Methods Progr. Biomed. 80 (2005) 17–23. method for describing cloud breaking, Physica A 274 (1999) 349–354.
[38] L.S. Vidyaratne, K.M. Iftekharuddin, Real-time epileptic seizure detection using [68] T. Higuchi, Approach to an irregular time series on the basis of the fractal theory,
EEG, IEEE Trans. Neural Syst. Rehabil. Eng. 25 (2017) 2146–2156, a publication Phys. Nonlinear Phenom. 31 (1988) 277–283.
of the IEEE Engineering in Medicine and Biology Society. [69] A. Petrosian, Kolmogorov complexity of finite sequences and recognition of
[39] S.N. Baldassano, B.H. Brinkmann, H. Ung, T. Blevins, E.C. Conrad, K. Leyde, M. different preictal EEG patterns, in: Proceedings Eighth IEEE Symposium on
J. Cook, A.N. Khambhati, J.B. Wagenaar, G.A. Worrell, B. Litt, Crowdsourcing Computer-Based Medical Systems, IEEE, 1995, pp. 212–217.
seizure detection: algorithm development and validation on human implanted [70] R. Esteller, G. Vachtsevanos, J. Echauz, B. Litt, A comparison of waveform fractal
device recordings, Brain 140 (2017) 1680–1691. dimension algorithms, IEEE Trans. Circuits Syst. I: Fund. Theor. Appl.v 48 (2001)
[40] G.R. Minasyan, J.B. Chatten, M.J. Chatten, R.N. Harner, Patient-specific early 177–183.
seizure detection from scalp electroencephalogram, J. Clin. Neurophysiol. 27 [71] B. Hjorth, EEG analysis based on time domain properties, Electroencephalogr.
(2010) 163–178, official publication of the American Electroencephalographic Clin. Neurophysiol. 29 (1970) 306–310.
Society. [72] M. Alirezanejad, R. Enayatifar, H. Motameni, H. Nematzadeh, Heuristic filter
[41] R. Hussein, H. Palangi, R. Ward, Z. Wang, Epileptic Seizure Detection: A Deep feature selection methods for medical datasets, Genomics (2019).
Learning Approach, 2018. [73] A. Asif, M. Majid, S.M. Anwar, Human stress classification using EEG signals in
[42] H. Stefan, S. Rampp, R.C. Knowlton, Magnetoencephalography adds to the response to music tracks, Comput. Biol. Med. 107 (2019) 182–196.
surgical evaluation process, Epilepsy Behav. : E&B 20 (2011) 172–177. [74] A. Ahmadi, S. Davoudi, M.R. Daliri, Computer Aided Diagnosis System for
[43] P. Holler, E. Trinka, Y. Holler, High-frequency oscillations in the scalp multiple sclerosis disease based on phase to amplitude coupling in covert visual
electroencephalogram: mission impossible without computational intelligence, attention, Comput. Methods Progr. Biomed. 169 (2019) 9–18.
Comput. Intell. Neurosci. 2018 (2018) 1638097. [75] S. Khoshnoud, M.A. Nazari, M. Shamsi, Functional brain dynamic analysis of
[44] C. Yeon, D. Kim, K. Kim, E. Chung, Visual evoked potential recordings in mice ADHD and control children using nonlinear dynamical features of EEG signals,
using a dry non-invasive multi-channel scalp EEG sensor, J Vis Exp (2018). J. Integr. Neurosci. 17 (2018) 11–17.
[45] A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J. [76] D. Endres, S. Maier, B. Feige, N.A. Posielski, K. Nickel, D. Ebert, A. Riedel,
E. Mietus, G.B. Moody, C.K. Peng, H.E. Stanley, PhysioBank, PhysioToolkit, and A. Philipsen, E. Perlov, L. Tebartz van Elst, Altered intermittent rhythmic delta
PhysioNet: components of a new research resource for complex physiologic and theta activity in the electroencephalographies of high functioning adult
signals, Circulation 101 (2000) E215–E220. patients with autism spectrum disorder, Front. Hum. Neurosci. 11 (2017) 66.
[46] P. Fergus, D. Hignett, A. Hussain, D. Al-Jumeily, K. Abdel-Aziz, Automatic [77] M.-W. Huang, C.-W. Chen, W.-C. Lin, S.-W. Ke, C.-F. Tsai, SVM and SVM
epileptic seizure detection using scalp EEG and advanced artificial intelligence ensembles in breast cancer prediction, PloS One 12 (2017) e0161501-e0161501.
techniques, BioMed Res. Int. 2015 (2015) 986736. [78] T. Chen, C. Guestrin, XGBoost: a scalable tree boosting system, in: Proceedings of
[47] J. Gotman, Automatic recognition of epileptic seizures in the EEG, the 22nd ACM SIGKDD International Conference on Knowledge Discovery and
Electroencephalogr. Clin. Neurophysiol. 54 (1982) 530–540. Data Mining, ACM, San Francisco, California, USA, 2016, pp. 785–794.
[48] W.T. Blume, G.B. Young, J.F. Lemieux, EEG morphology of partial epileptic [79] I. Guyon, J. Weston, S. Barnhill, V. Vapnik, Gene selection for cancer
seizures, Electroencephalogr. Clin. Neurophysiol. 57 (1984) 295–302. classification using support vector machines, Mach. Learn. 46 (2002) 389–422.
[49] I. Ginzburg, Prediction of the moments in advection-diffusion lattice Boltzmann [80] S. Sahran, D. Albashish, A. Abdullah, N.A. Shukor, S. Hayati Md Pauzi, Absolute
method. I. Truncation dispersion, skewness, and kurtosis, Phys. Rev. E 95 (2017), cosine-based SVM-RFE feature selection method for prostate histopathological
013304. grading, Artif. Intell. Med. 87 (2018) 78–90.
[50] J.A. Blanco, A.C. Vanleer, T.K. Calibo, S.L. Firebaugh, Single-trial cognitive stress [81] X. Feng, R. Zhang, M. Liu, Q. Liu, F. Li, Z. Yan, F. Zhou, An accurate regression of
classification using portable wireless electroencephalography, Sensors 19 (2019). developmental stages for breast cancer based on transcriptomic biomarkers,
[51] J. Bai, Y. Li, Y. Yi, W. Cheng, H. Du, PAPR reduction based on tone reservation Biomarkers Med. 13 (2019) 5–15.
scheme for DCO-OFDM indoor visible light communications, Optic Express 25 [82] S. Zhang, X. Li, M. Zong, X. Zhu, R. Wang, Efficient kNN classification with
(2017) 24630–24638. different numbers of nearest neighbors, IEEE Trans. Neural Network. Learn Syst.
[52] A. Gramfort, D. Strohmeier, J. Haueisen, M.S. Hamalainen, M. Kowalski, Time- 29 (2018) 1774–1785.
frequency mixed-norm estimates: sparse M/EEG imaging with non-stationary [83] J. Hu, Automated detection of driver fatigue based on AdaBoost classifier with
source activations, Neuroimage 70 (2013) 410–422. EEG signals, Front. Comput. Neurosci. 11 (2017) 72.
[84] S.K. Palei, S.K. Das, Logistic regression model for prediction of roof fall risks in
bord and pillar workings in coal mines: an approach, Saf. Sci. 47 (2009) 88–96.

12
S. Yang et al. Computers in Biology and Medicine 119 (2020) 103671

[85] L. Rutkowski, M. Jaworski, L. Pietruczuk, P. Duda, The CART decision tree for [98] F. Hasanzadeh, M. Mohebbi, R. Rostami, Prediction of rTMS treatment response
mining data streams, Inf. Sci. 266 (2014) 1–15. in major depressive disorder using machine learning techniques and nonlinear
[86] Z.-H. Zhou, Ensemble Methods: Foundations and Algorithms, Chapman and Hall/ features of EEG signal, J. Affect. Disord. 256 (2019) 132–142.
CRC, 2012. [99] R.Q. Ge, M.L. Zhou, Y.X. Luo, Q.H. Meng, G.Q. Mai, D.L. Ma, G.Q. Wang, F.
[87] Y. Freund, R.E. Schapire, A decision-theoretic generalization of on-line learning F. Zhou, McTwo: a two-step feature selection algorithm based on maximal
and an application to boosting, J. Comput. Syst. Sci. 55 (1997) 119–139. information coefficient, BMC Bioinf. 17 (2016).
[88] M. Maniruzzaman, M. Jahanur Rahman, B. Ahammed, M.M. Abedin, H.S. Suri, [100] A. Blumer, A. Ehrenfeucht, D. Haussler, M.K. Warmuth, Occam’s razor, Read.
M. Biswas, A. El-Baz, P. Bangeas, G. Tsoulfas, J.S. Suri, Statistical characterization Mach. Learn. 24 (1987) 377–380.
and classification of colon microarray gene expression data using multiple [101] M. Shah, M. Marchand, J. Corbeil, Feature selection with conjunctions of decision
machine learning paradigms, Comput. Methods Progr. Biomed. 176 (2019) stumps and learning from microarray data, IEEE Trans. Pattern Anal. Mach. Intell.
173–193. 34 (2012) 174–186.
[89] K. Li, G. Zhou, J. Zhai, F. Li, M. Shao, Improved PSO_AdaBoost ensemble [102] H. Qu, J. Gotman, A patient-specific algorithm for the detection of seizure onset in
algorithm for imbalanced data, Sensors 19 (2019). long-term EEG monitoring: possible use as a warning device, IEEE Trans. Biomed.
[90] C. Jia, Q. Yang, Q. Zou, NucPosPred: predicting species-specific genomic Eng. 44 (1997) 115–122.
nucleosome positioning via four different modes of general PseKNC, J. Theor. [103] S. Raghu, N. Sriraam, Y. Temel, S.V. Rao, A.S. Hegde, P.L. Kubben, Performance
Biol. 450 (2018) 15–21. evaluation of DWT based sigmoid entropy in time and frequency domains for
[91] J.H. Friedman, Greedy function approximation: a gradient boosting machine, automated detection of epileptic seizures using SVM classifier, Comput. Biol.
Ann. Stat. 29 (2001) 1189–1232. Med. 110 (2019) 127–143.
[92] J.H. Friedman, Stochastic gradient boosting, Comput. Stat. Data Anal. 38 (2002) [104] M. Kaleem, D. Gurve, A. Guergachi, S. Krishnan, Patient-specific seizure detection
367–378. in long-term EEG using signal-derived empirical mode decomposition (EMD)-
[93] N.S. Malan, S. Sharma, Feature selection using regularized neighbourhood based dictionary approach, J. Neural. Eng. 15 (2018).
component analysis to enhance the classification performance of motor imagery [105] J. Xiang, C. Li, H. Li, R. Cao, B. Wang, X. Han, J. Chen, The detection of epileptic
signals, Comput. Biol. Med. 107 (2019) 118–126. seizure signals based on fuzzy entropy, J. Neurosci. Methods 243 (2015) 18–25.
[94] M. Rawashdeh, S. Lewis, M. Zaitoun, P. Brennan, Breast lesion shape and margin [106] A. Bhattacharyya, R.B. Pachori, A multivariate approach for patient-specific EEG
evaluation: BI-RADS based metrics understate radiologists’ actual levels of seizure detection using empirical wavelet transform, IEEE (Inst. Electr. Electron.
agreement, Comput. Biol. Med. 96 (2018) 294–298. Eng.) Trans. Biomed. Eng. 64 (2017) 2003–2015.
[95] A. Gupta, P. Singh, M. Karlekar, A novel signal modeling approach for [107] P. Thodoroff, J. Pineau, A. Lim, Learning robust features using deep learning for
classification of seizure and seizure-free EEG signals, IEEE Trans. Neural Syst. automatic seizure detection, in: D.-V. Finale, F. Jim, K. David, W. Byron, W. Jenna
Rehabil. Eng. 26 (2018) 925–935. (Eds.), Proceedings of the 1st Machine Learning for Healthcare Conference,
[96] Z. Zhang, T. Wen, W. Huang, M. Wang, C. Li, Automatic epileptic seizure PMLR, Proceedings of Machine Learning Research, 2016, pp. 178–190.
detection in EEGs using MF-DFA, SVM based on cloud computing, J. X Ray Sci. [108] A. Harati, S. Lopez, I. Obeid, J. Picone, M. Jacobson, S. Tobochnik, The TUH EEG
Technol. 25 (2017) 261–272. CORPUS: a big data resource for automated EEG interpretation, in: 2014 IEEE
[97] H. Namazi, V.V. Kulish, J. Hussaini, J. Hussaini, A. Delaviz, F. Delaviz, S. Habibi, Signal Processing in Medicine and Biology Symposium (SPMB), IEEE, 2014,
S. Ramezanpoor, A signal processing based analysis and prediction of seizure pp. 1–5.
onset in patients with epilepsy, Oncotarget 7 (2016) 342–350.

13