EEG epileptic seizure detection and classification based on dual-tree complex wavelet transform and machine learning algorithms

Introduction

According to the World Health Organization, epilepsy affects about 1% to 2% of the world's population; about 65 million people have epilepsy^[1–2]. Different examinations are used to confirm the epilepsy diagnosis, such as electroencephalogram (EEG) signals recorded from the patient's scalp, which is a collection of potential signals from many electrodes and they represent the spatial distribution of the potential fields in the brain. EEG is an invaluable tool for characterization of the spatial-temporal dynamics of neuronal activity in the brain; hence, diagnosis by manual observation of these signals is a very sensitive task that resulted in practical tools for the research area dedicated to the diagnosis of seizure detection^[3].

More than 120 million articles cited for epileptic seizure detection proved that this subject is still topical and should be addressed. Various epileptic seizure detection methods have been proposed in literature, and a brief review of some techniques related to our work was conducted and introduced in the next section. Gotman and Gloor were among the first to detect seizures using EEG signals^[4]. The first machine learning approach was developed using the CHB-MIT database^[5]. Some researchers have used a large database, extracted from Freiburg and CHB-MIT databases for seizure detection, and their proposed method allows achieving a high detection performance up to 100%^[6]. Some proposed methods are based on artificial neural networks allowing achieving a perfect classification rate for detecting epileptic seizures^[7]. Some are based on support vector machine (SVM) with time-frequency features and wavelet variance features allowing achieving correct classification rates up to 100%^[8]. Another seizure detection method is based on dual-tree complex wavelet transformation (DTCWT) and Fourier features allowing achieving a correct classification rate up to 100%^[9].

Various approaches for epileptic seizure detection have been proposed in literature. The processing of EEG signals requires that these data must be noiseless to establish reliable disease diagnosis; therefore, a filter step is necessary for this purpose. Some researchers proposed to use the Savitzky-Golay filter and have its performance verified^[10], while others proposed a multi-scale principal component analysis (MSPCA)^[11]. In general, all had adopted the objectives of the extraction characteristics and their classification, in order to differentiate between classes (i.e., seizure free: "interictal" class and seizure: "ictal" class). Frequency domain analysis^[11] and other time-frequency approaches, such as the short-term Fourier transform^[12] and the wavelets^{[13,14– 15]} have often been used to extract discriminant characteristics representing the EEG signals. Advanced detection of epileptic seizures used a feature extraction technique based on wavelets^[16]. Some features have shown high performance for characterization changes in EEG signals such as entropy, energy^{[17– 18]}, principal component analysis (PCA)^[19], and independent component analysis (ICA).

Based on these previous studies, the discrete wavelet transform (DWT) is one of the most widely used techniques for extracting EEG signal characteristics. For the same purpose, the DTCWT technique^[20], applies filter banks to handle the wavelet domain. This technique was recently applied for epileptic seizure detection which demonstrated its performance for extracting the characteristics of the time-frequency domain of EEG signals^[21]. After extracting features, the next step is to employ classification methods in order to determine the class of input features. Various classification methods are available. These range from rule-based decision-making^[4] and SVM^[22] to artificial neural network (ANN). For example, the classification of EEG signals was investigated using k-means classifier and multilayer perceptron neural network model^[14]. A supervised automatic learning method was proposed for detecting seizures using linear discriminant analysis (LDA)^[23].

The goal of this work is to build a reliable and accurate seizure detection method to facilitate and accomplish the diagnosis of epilepsy. Artifact eliminator is a valuable aid for diagnosis in EEG signal, using the Savitzky-Golay filter and the MSPCA. To overcome the non-linearity and non-stationary of EEG signals, the time-frequency techniques have been developed to eliminate the strict restrictions of classical frequency methods. Three popular signal decomposition techniques chosen in this work include empirical mode decomposition (EMD), DWT, and DTCWT. After crossing through a step of denoising the artifacts and extracting the frequency bands using EMD, DWT, and DTCWT, statistical values ​​were used to extract the relevant characteristics of these frequency bands. Then three classifiers k-nearest neighbor (k-NN, LDA, and SVM were established to distinguish between classes (i.e., interictal and ictal).

The choice of MSPCA and Savitzky-Golay for artifact removal is justified by their proven superiority in different biomedical signals kinds, such as ECG^[24], EMG^[25], and EEG^[26]. The plethora of EMD, DWT, and DTWCT signal decomposition methods used in different fields of analysis justifies the choice in this work. This step is followed by the extraction of the statistical characteristics to extract the necessary information with reduction of used data dimensions. The aforementioned modulators were combined with three popular machine-learning techniques to find a powerful model for detecting epileptic seizures. Therefore, this work aims to develop an EEG signal classification system for seizure detection using CHB-MIT (EEG scalp) databases, and to ensure the robustness of the proposed model in which two databases have been extracted and realized based on EEG segments with and without overlapping procedure. The results proved that the proposed model is suitable for epileptic seizure detection. Fig. 1 illustrates the work plan presented in this study.

Figure  1.  Diagram for proposed seizure detection approach.

MSPCA: multi-scale principal component analysis; DWT: discrete wavelet transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet transformation; SVM: support vector machine; k-NN: k-nearest neighbor; LDA: linear discriminant analysis.

DownLoad: Full-Size Img PowerPoint

Materials and methods

The lack of visual difference between the free seizure and seizure EEG signals makes the epileptic seizure detection a difficult task. Therefore, three proposed steps were employed, with EEG signal denoising techniques followed by features extraction methods to enable faster and more effective seizure detection. Hence, the classification using the extracted characteristics is performed using different machine learning. The EEG database used in our study is presented with a brief description.

EEG database

All tests and simulations were conducted based on the database collected at the Boston Children's Hospital (CHB) and used to validate the experimental results. This "CHB-MIT" database includes 23 different subsets containing EEG records of pediatric patients. It contains 182 seizures. All signals were sampled at 256 Hz per second with a resolution of 16 bits. Most files contain 23 EEG signals. The international 10-20 EEG electrodeposition and nomenclature system was used for these recordings^[5].

EEG signal pre-processing

Biomedical signals such as EEG were contaminated by artifacts and noise, and thus a pre-processing step is needed to clean the data. Many signal-filtering algorithms have been proposed to improve the performance of health devices (EEG, ECG, EMG...) besieged by artifacts. Savitzky-Golay filter and multi-scale principal component analysis have been used in recent research and their performance has been verified.

Savitzky-Golay filter

During the pre-processing of biomedical signals, the most challenging problem is to extract high resolution EEG signals from noisy measurements and retain the EEG waveform sharpness. The spectral quality of a signal [signal-to-noise (S/N) ratio] may be improved by increasing filter width or smoothing the signal multiple times. Hence, The Savitzky-Golay filter is a low-pass filter used for attenuation noise, without damaging or destroying the spectral properties of data. This filter is defined as a weighted moving average that performs a least-squares-fit convolution procedure.

Therefore, the S–Golay algorithm is characterized by the selected data interval, in other words, the window size, a polynomial function is fitted to the chosen size window and the data point (noisy signal) is calculated using the polynomial coefficients. The set of data $n\{{x}_{j},{y}_{j}\}$ points, where j=1,2,…,n; x is an independent variable and y is the observed value, can be represented with convolution coefficients called c and C_i and expressed by:

Hence, S-Golay filter requires the inputs which include: the input noisy signal (x), the polynomial order (k) and its frame size (f). One of the advantages of this filter is to retain distribution characteristics, which are often flattened with other smoothing methods^[10,27].

Multi-scale principal component analysis (MSPCA)

The MSPCA approach is proposed to achieve the objective of eliminating noise without sacrificing a significant amount of fast changes in the signal and to provide a dimension reduction with select components relevant in the different scales. The advantages of this method come from the combination of the PCA, which is responsible for extracting the relationship between several variables, and the wavelets, that divide the stochastic processes of the determinists and approximately decorrelate the auto-correlation between the variables. The MSPCA calculates the PCA of all wavelet coefficients for all wavelet decomposition levels, and then establishes a relationship between the results obtained and the suitable levels^[24]. Fig. 2 illustrates the steps of the MSPCA method and the following algorithm explains it:

Figure  2.  The procedure for MSPCA.

DWT: discrete wavelet transform; PCA: principal component analysis.

DownLoad: Full-Size Img PowerPoint

1) Perform wavelet decomposition at level I for each column of data matrix X.

2) For 1≤i≤I, perform the PCA of the detailed matrix G_iX, choose an appropriate number of significant principal components p_i or suppress the detail.

3) Perform the PCA of the approximation matrix H_IX and select a suitable number of principal components p_I+1.

4) The main characteristics of the original matrix X put in a new matrix $\dot X$, by inverting the DWT^T wavelet transformation.

5) Finally, execute the PCA of $\dot X$^[28].

EEG signal decomposition

To characterize the EEG signal and determine its class, some methods have often been used to extract discriminant characteristics representing EEG signals from frequency analysis and time-frequency approaches. These, such as DWT, DTCWT, and EMD, are recently used in the literature and will be used in this work and represented in this section with their proven superiority to reach this goal.

Empirical mode decomposition (EMD)

EMD is a procedure used to decompose signals into different simple oscillatory modes called intrinsic function mode (IMF). Hence, the signal x(t) can be represented by the sum of these IMFs components and by a residue as follows^[29]:

where M is the number of IMFs and r_M(t) is the residue. The IMFs of a signal must satisfy two necessary conditions. Firstly, the difference between the total numbers of extrema and zero-crossings is at most one. Secondly, in all points, the average value of the envelope defined by the local maxima and that by the local minima is zero or very close to zero^[30]. The following steps represent the EMD algorithm:

1) Extract local maxima and local minima from signal of x(t),

2) Connect between all minima and all maxima to obtain two envelopes e_min(t) and e_max(t),

3) Compute the average between them as: $m\left(t\right)=\dfrac{{[e}_{min}\left(t\right)+{e}_{max}\left(t\right)]}{2}$,

4) Extract the detail from x(t) as: d(t)=x(t)-m(t)

5) Examine, if d(t) is an IMF or not by verifying the above conditions for IMFs,

6) Reiterate on the residual d(t).

After having the first IMF D₁(t) of the signal, to get other IMFs, generate the residual r(t) as follows: r(t)=x(t)–D₁(t). The process described above will be continued until IMF can no longer be obtained from the final residue.

Discrete wavelet transform (DWT)

DWT is one of the most efficient time-frequency tools for non-stationary signal analysis. The DWT procedure makes it possible to present an input signal in sets of functions, called wavelets, by scaling and shifting the mother function into wavelets, that is to say, to decompose a signal into two components called sub-bands: a low frequency component called "approximation coefficients A", and a high frequency component called "detailed coefficients D". The low frequency component can further be decomposed into approximations and details. After each decomposition, the frequency resolution is doubled and the temporal resolution is halved. After choosing the decomposition level j, the components of the signal s(k) can be represented by the following expression and Fig. 3^[31]:

Figure  3.  DWT decomposition tree.

DownLoad: Full-Size Img PowerPoint

Regarding Fig. 3, h and g are low-pass and high-pass filters, respectively. s is the input original signal where Ca1, Ca2, and Ca3 represent the approximation coefficients of h(1), h(2) and h(3) and Cd1, Cd2 and Cd3 represent the detail coefficients of g(1), g(2) and g(3) at level 1, 2 and 3, respectively.

Dual-tree complex wavelet transform (DTCWT)

DTCWT, a DWT enhancement, uses double wavelet filters low and high at each scale to obtain two parts: a real and an imaginary complex wavelet coefficient. This transformation is invariant by directional and selective shifting in two or more dimensions, which are very important in applications such as pattern recognition and signal analysis.

Hence, the DTCWT can be represented by the following expression^[20]:

where the complex wavelet is ψ(t) while the complex scaling function is φ(t). With reference to Fig. 4, s represents the input signal, and the output coefficients of the top tree h and those of the bottom tree g are considered as the real and imaginary parts, respectively. Therefore, ${h}_{0}^{\left(1\right)}, {h}_{0}$, ${g}_{0}^{\left(1\right)}$, and g₀ are a low-pass filter and ${h}_{1}^{\left(1\right)}$, ${g}_{1}^{\left(1\right)}$, h₁, and g₁ are high-pass filters. Consequently, the corresponding wavelets ψ_h(t) and ψ_g(t) form approximately a Hilbert pair, while the scaling function φ_g(t) is approximately the Hilbert transform of φ_h(t).

Figure  4.  An illustrative scheme of the DTCWT at level 3.

DownLoad: Full-Size Img PowerPoint

Discriminative feature extraction

The classification requires a necessary step whose characteristic extraction aims to obtain the most important information and preserve the small size of the data. Therefore, after signal decomposition using the DWT, EMD and DTCWT methods, the most prominent statistical features have been used for this study to obtain the class of the input EEG signal:

Mean of coefficients in each sub-band:

Average power of coefficients in each sub-band:

Standard deviation of coefficients in each sub-band:

Signal classification

Frequency band decomposition is performed to represent an EEG signal in feature vector and to obtain the most important information of the data to determine its class. The identification of signals relied on classification algorithms to estimate the desired class. For this purpose, three popular classifiers are used, namely LDA, k-NN, and SVM, which are briefly explained in this section.

Linear discriminant analysis (LDA)

LDA, also called Fisher's LDA, is a simple technique used in machine learning, statistics, and pattern recognition to find a linear separating two or more classes, and in general, it gives a good class estimation. Fisher's LDA requires very little computation. LDA aims to use hyperplanes to separate data from different classes, and the separator hyperplane is obtained by looking for the projection that maximizes the distance between classes and minimizes the interclass variance between them. For a class number N (N>2), several hyperplanes are used^[32–33]. In this work, the linear discriminant type was used for regularized linear discriminant analysis.

Support vector machine (SVM)

With excellent performance results, SVM is applied in various fields. Contrary to LDA, which constructs a probabilistic model for each class using all data points, SVM is to minimize a linear separation error by focusing on the neighboring points of the linear separation surface, that is to say, the points located far from the separation surface are ignored in the learning process. SVM is based on the Vapnik-Chervonenkis theory and the principle of structural risk minimization. Its purpose is to determine the minimizing of the training set error by maximizing the boundary of the separating hyper-plane between the data^{[34– 35]}. In this work, the Kernel function used to differentiate between the two classes is linear kernel with the Kernel scale parameter chosen to be 1.

k-nearest neighbor (k-NN)

The k-NN is a machine-learning technique used for the classification task, which aims to find the nearest k-neighbors among the learning set, and the categories of k closest neighbors, which are used to weight the candidates in the category. The closest neighbors to the sample are expressed as distance such as Euclidean distance, Chebyshev distance, and Manhattan distance. The performance of k-NN algorithm strongly depends on two factors: an appropriate similarity function and an appropriate value for k. If k is too large, large classes will overwhelm small classes, whereas if k is too small, the advantage of the k-NN algorithm cannot be traced^{[36– 37,19]}. In this research, 3-NN algorithm (3 nearest neighbors) and Euclidean distance metric were used to classify the EEG signals.

Database preparation

The CHB-MIT database is one of the EEG databases used in recent research for epileptic seizure detection. To prove the effectiveness of the proposed method, two EEG datasets have been extracted from the CHB-MIT database. Hence, the extraction of shorter EEG segments based on no overlapping process represents the first used database, while the second based on overlapping process. The segment extraction process of these two datasets is clearly explained in this section. The performance evaluation criteria to assess the performance and the effectiveness of the proposed method and the experimental results obtained will be shown as follows.

The developed seizure detection model is based on EEG segment extraction with an 8-second rectangular window. The seizure time is known and given in the description database for each patient. The first extracted dataset contains 400 EEG segments extracted approximately from all the recorded signals in the CHB-MIT database with 200 no seizure segments ("interictal") and 200 seizure segments ("ictal").

According to a previous study^[6], the choice of the 8 seconds window size (with 2 048 samples) provides a more robust performance than that of shorter or longer windows. To prove the effectiveness of the proposed seizure detection model, the overlapped database has been created with an overlapping process based on the first extracted database that will be presented as follows.

Overlapping procedure

While the first database was created (with 400 segments), to evaluate the performance and better prove the effectiveness of the proposed seizure detection method, another database will be created based on the first realized with an overlapping process. Hence, new segments will be created by overlapping between two segments belonging to the same EEG signal in the same class. Let suppose that:

- N is the segment number of the first database already created (without overlapping) called B. N1 and N2 is the number of segments in "interictal" class B1 and "ictal" class B2 respectively,

- Nc is the segment number of the second database with overlapping process called Bc. Nc1 and Nc2 are the numbers of segments in "interictal" class Bc1 and "ictal" class Bc2 respectively.

Therefore, the overlap process is presented as follows:

- n represents the segment samples number (2 048), i is segment index (i^th segment), p is the point offset for overlap, fe is the sampling frequency 256 Hz, and sc is the segment seconds number (8 seconds).

Example: let suppose i and i+1 are two segments of the same class to be overlapped where $X{1}_{i}, \;X{2}_{i},$$X{3}_{i}, \;X{4}_{i}, \;X{5}_{i}, \;X{6}_{i}, \;X{7}_{i}, \;X{8}_{i}$ and $X{1}_{i+1}, \;X{2}_{i+1}, \;X{3}_{i+1}, \;$$X{4}_{i+1}, \;X{5}_{i+1}, \; X{6}_{i+1}, \;X{7}_{i+1}, X{8}_{i+1}$, are represent the sample of segments of $i$ and $i+1,$ respectively, and $j$ the new obtained segment with a gap between the segment is $p=6$.

Finally, the new segment can be presented by the following expression:

The size of new database (overlapped) will be increased compared to the first database.

$N=N1+N2$ and $Nc=Nc1+Nc2$; hence, $Nc1=$$\left(N1 \times 2\right)-1$ and $Nc2=$$\left(N2 \times 2\right)-1$.

Therefore, the overlapped database size will be 798 segments, with 399 segments in both classes and each length segment 8 seconds. The experimental results of the proposed approach using the two extracted databases will be presented by following the performance criteria to validate the model accuracy.

Feature selection process

The sampling rates of the databases are 256 Hz; the polynomial order was chosen to be 2 on S-Golay filter and the frame length was selected to be 17, proved to be the parameter' choice^[10]. In MSPCA, a five-level decomposition was used to generate the sub-bands signals. The Kaiser rule was responsible for eliminating the number of loadings^[26]. After the artifact removal procedure with S-Golay and MSPCA, the features have been extracted from the sub-band decomposed signals using EMD, DWT, and DTCWT methods. The Daubechies wavelets of different orders (2, 3, 4, 5, and 6) are investigated for the analysis of epileptic EEGs. This family of wavelets has orthogonal property and is efficient as a filter implementation^[38]. The fourth order Daubechies wavelet is found to be most appropriate for analysis of EEG data^[6,39]. Hence, DB4 wavelet function was used to decompose the EEG signals with DWT. The decomposition levels in all methods were selected to be 3. Hence, 4, 8 and 3 sub-band signals were obtained using DWT, DTCWT, and EMD respectively. Therefore, in each sub-band signals three statistical characteristics have been extracted to obtain 12, 24, and 9 features for DWT, DTCWT, and respectively. However, although the number of extracted features was considerably doubled in DTCWT, the performance in the testing step improved.

Performance evaluation criteria

To prove the robustness of the proposed method, cross-validation (CV) is a procedure mainly applied to estimate the proficiency of machine learning algorithm. CV is used to estimate how the model should work when used to make predictions of unused data at the step of preparation and model learning. In a 10 times CV, the database is randomly separated into 10 different mutually exclusive folds of the same size, where nine folds will be used to training the data and the remaining one will be used to test and predict the classification. The CV accuracy is expressed by:

where represents the number of folds used, and ${A}_{i}$ is the accuracy measure of each, $i=1,\dots , k.$ In addition, there are three performance measures to evaluate machine learning methods using true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), which can be expressed by:

Results

The previous sections have shown that the first goal in this work is to find a model to classify EEG signal segments to develop automatic seizure detection systems using the children's EEG database CHB-MIT. However, shorter EEG segments with 8 seconds have been extracted from this database that makes it possible to develop a practical and robust seizure detection system to diagnose epilepsy. Ictal record signals represent a small amount of data compared to interictal record signals. Therefore, a subsampling paradigm of the class with larger proportions (interictal) is used to circumvent the imbalance between the two classes^[40], in order to have better accuracy of seizure detection. In this work, segments of interictal data were randomly selected to produce a comparable and balanced number of ictal and interictal segments.

Based on the segment model, the purpose of this study was to develop an EEG signal classification model. Therefore, the contribution is to have a powerful and suitable model using three popular methods to extract discriminant characteristics such as EMD, DWT, and DTCWT, followed by a dimension reduction employing three different statistical characteristics and three commonly used methods of automatic learning such as LDA, k-NN, and SVM.

To prove the performance results of the average precision, extraction methods were evaluated and the results are presented in detail and illustrated in Fig. 5. The parameter values of all selected techniques are explained in the preview section because they produce the most robust results. In Fig. 5, the plots are shown in the left column present the S-Golay results to apply a series of data points to denoise the signals without deforming its input, while the plots are shown in the right column present the MSPCA results, which shows the advantage of PCA in extracting necessary information and the effectiveness of extracting deterministic features from stochastic processes using wavelets. With the two filter processes compared, it can be found that MSPCA outperformed the S-Golay filter in the removal of artifacts; MSPCA achieved a 100% accuracy with the three classification methods while S-Golay 98%.

Figure  5.  Classification accuracy of the proposed method using the signal decomposition transforms for feature extraction considered in this study (i.e., EMD, DWT, DTCWT) with EEG signal pre-processing techniques: MSPCA (right column) and S-Golay (left column).

DownLoad: Full-Size Img PowerPoint

With the three EEG signal decomposition methods compared, the results indicate that using the two denoising methods, DTCWT was superior to DWT and much better than EMD where with S-Golay (DTCWT, 99%; DWT, 94%; and EMD, 91%). Fig. 5 also shows that the DTCWT gave a fairly high classification result with 100% accuracy classification rates, which is attributed to the approximate shift invariant property, a very important factor in epileptic seizure detection. On aspects of the performance of feature extraction methods, the experimental results show the superiority of the standard deviation and the average power, which both gave a rather high classification accuracy when compared to the mean of the sub-band features. Hence the mean static feature with DTCWT outperforms that with DWT, which ensures the high effectiveness of DTCWT. In addition, the result provides that k-NN and SVM gave 100% classification accuracy when compared to LDA.

Discussion

The overall results obtained in this study show that combination of the DTCWT with SVM and k-NN improves the accuracy rates of the classification especially by using the standard static deviation and average power characteristics with the two denoising processes. As shown in Fig. 5 which (MSPCA+DTCWT+STD+SVM), (MSPCA+DTCWT+STD+k-NN, (S-Golay+DTCWT+STD+SVM) and (S-Golay+DTCWT+STD+k-NN achieve a 100% classification accuracy rate. Hence, it can be concluded that the proposed model is reliable and robust for seizure detection with the DTCWT combined with SVM and k-NN.

To compare the results of the proposed approach with seizure detection methods realized in literature, Table 1 provides that the proposed approach has achieved an overall perfect accuracy of 100% in seizure detection using EEG segments.

Table  1.  Comparison of the proposed method against some state-of-the-art methods

Method	Feature extraction	Classifier	Accuracy
Fergus et al[23] (2014)	Band-pass filter, peak and median frequency, RMS, Entropy based-features	k-NN	88%
Raffiudin et al[16] (2011)	Wavelet-based features+statistics-based features+IQR+MAD	LDA	96%
Alickovic et al[6] (2018)	Multiscale PCA, EMD, DWT, WPD, statistics-based features	RF, SVM, MLP, and k-NN	99.7%
Swami et al[21] (2016)	DTCWT	GRNN	98%
Gandhi et al[18] (2011)	DWT, energy, standard deviation and entropy based-features	SVM and PNN	99%
Acharya et al[13] (2012)	Approximate entropy, sample entropy, HOS, decision tree, fuzzy-based features	GMM, k-NN, SVM, and RBPNN	99%
Swami et al[22] (2014)	WPT, energy, standard deviation and entropy based-features	SVM	99%
Proposed method	MSPCA, DTCWT	SVM and k-NN	100%
MSPCA: multi-scale principal component analysis; DWT: discrete wavelettransform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet transformation; SVM: support vector machine; k-NN:k-nearest neighbor; LDA: linear discriminant analysis.

 | Show Table

DownLoad: CSV

The obtained results using the overlapped database ensure the reliability and stability of this seizure detection model as illustrated in Table 2 and Fig. 6. The combination of DTCWT and MSPCA offers advantages over the DWT and EMD and keeps the same 100% accuracy performance. The important difference between EMD and wavelet-based methods lies in the process of feature extraction. IMF signals are the product of EMD methods and have the same length as the original EEG signals, which have given no reduction in the number of samples. On the other hand, the coefficients characterizing different sub-bands produced by wavelet-based methods (DWT and DTCWT) have a much smaller number of samples than the EEG signals input. However, the DWT suffers from certain disadvantages such as the oscillatory nature, the aliasing of the spectrum with limited directional information, and the lack of lag invariance. The DTCWT has solved these problems by its non-oscillating property with approximate invariant shift amplitude where this property is very important for the detection of epileptic seizures in EEG signals.

Table  2.  Classification accuracy of the proposed overlapping method

	Overlapping		Features
			Average power				Mean				Standard deviation
			2	4	6		2	4	6		2	4	6
DTCWT	SVM	S-Golay	98	98	96		75	80	78		97	98	96
		MSPCA	100	100	100		87	71	71		100	100	100
	k-NN	S-Golay	98	97	96		86	88	89		99	98	96
		MSPCA	100	100	100		93	92	93		100	100	100
	LDA	S-Golay	93	92	93		70	81	71		99	93	92
		MSPCA	91	87	91		76	74	75		98	96	97
DWT	SVM	S-Golay	94	92	92		75	78	75		93	94	94
		MSPCA	100	100	99		76	71	72		100	100	99
	k-NN	S-Golay	95	93	96		84	89	90		95	92	96
		MSPCA	100	100	99		88	90	91		100	100	100
	LDA	S-Golay	98	90	89		80	81	81		91	91	92
		MSPCA	90	91	90		76	72	75		98	93	98
EMD	SVM	S-Golay	94	95	95		66	61	64		93	94	92
		MSPCA	100	100	99		65	70	63		100	100	99
	k-NN	S-Golay	94	94	94		75	76	84		96	92	94
		MSPCA	99	100	99		92	91	78		99	100	100
	LDA	S-Golay	80	81	83		50	62	60		88	90	91
		MSPCA	84	88	87		63	72	67		99	97	97
MSPCA: multi-scale principal component analysis; DWT: discrete wavelettransform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet transformation; SVM: support vector machine; k-NN:k-nearest neighbor; LDA: linear discriminant analysis.

 | Show Table

DownLoad: CSV

Figure  6.  The sensitivity and specificity of the proposed method using the signal decomposition transforms for feature extraction considered in this study (i.e., EMD, DWT, DTCWT) with EEG signal pre-processing techniques (MSPCA and S-Golay).

MSPCA: multi-scale principal component analysis; DWT: discrete wavelet transform; EMD: empirical mode decomposition; DTCWT: dual-tree complex wavelet transformation; SVM: support vector machine; k-NN: k-nearest neighbor; LDA: linear discriminant analysis; Sens: sensitivity; Spec: specificity.

DownLoad: Full-Size Img PowerPoint

In conclusion, we have proposed a robust automatic approach for epileptic seizure detection and classification that was assessed on two EEG datasets generated from the CHB-MIT database. The proposed approach is based on two denoising signal methods (i.e., S-Golay and MSPCA) combined with three popular feature extraction techniques (i.e., DTCWT, DWT, and EMD). Three different machine-learning techniques (namely SVM, k-NN, and LDA) have been used to distinguish between the seizure-free and seizure segments based on the feature extraction methods already mentioned. The experimental results have shown that the employment of DTCWT has been very successful because it outperforms the other two methods (DWT and EMD). The results obtained have also clearly shown that MSPCA+DTCWT+SVM and k-NN offers a perfect classification with 100% accuracy. The findings demonstrate the effectiveness of the proposed method compared to existing methods. After completing the seizure detection task, future studies may take more interest in the prediction of epileptic seizures.