Electroencephalogram signal classification based on shearlet and contourlet transforms

Expert Systems With Applications 67 (2017) 140–147

Contents lists available at ScienceDirect

Expert Systems With Applications journal homepage: www.elsevier.com/locate/eswa

Electroencephalogram signal classification based on shearlet and contourlet transforms Paulo Amorim a, Thiago Moraes a, Dalton Fazanaro a, Jorge Silva a, Helio Pedrini b,∗ a b

Tridimensional Technology Division, Center for Information Technology Renato Archer, Campinas-SP 13069-901, Brazil Institute of Computing, University of Campinas, Campinas-SP 13083-852, Brazil

a r t i c l e

i n f o

Article history: Received 21 June 2016 Revised 25 September 2016 Accepted 25 September 2016 Available online 26 September 2016 Keywords: Epilepsy Electroencephalogram signals Shearlets Contourlets

a b s t r a c t Epilepsy is a disorder that affects approximately 50 million people of all ages, according to World Health Organization (2016), which makes it one of the most common neurological diseases worldwide. Electroencephalogram (EEG) signals have been widely used to detect epilepsy and other brain abnormalities. In this work, we propose and evaluate a novel methodology based on shearlet and contourlet transforms to decompose the EEG signals into frequency bands. A set of features are extracted from these timefrequency coefficients and used as input to different classifiers. Experiments are conducted on a public data set to demonstrate the effectiveness of the proposed classification method. The developed system can help neurophysiologists identify EEG patterns in epilepsy diagnostic tasks.

1. Introduction Electroencephalograms (EEGs) are recordings of the electrical activity of the brain. The monitoring methods are typically noninvasive, where electrodes are placed along the human scalp to identify abnormalities in EEG readings, such as epilepsy, sleep disorders, coma, encephalopathies, among other brain dysfunctions (Abou-Khalil & Misulis, 2006; Bronzino, 2000; Niedermeyer & da Silva, 2005). Although EEG interpretation is traditionally performed by a neurophysiologist through visual inspection of the signals, scientific and technological advances have driven the development of reliable computer-assisted EEG analysis. Imaging techniques, such as functional magnetic resonance imaging (fMRI), positron emission tomography (PET) and computed tomography (CT), have high-resolution spatial capabilities, however, EEG analysis is still a valuable tool for diagnosis and research related to neurological disorders. An important characteristic of EEGs is their temporal resolution on a millisecond scale, which allows for the detection of rapid changes in brain activity. Furthermore, EEG is relatively inexpensive compared to neuroimaging techniques. Time–frequency signal processing techniques, such as discrete wavelet transform (DWT) (Adeli, Zhou, & Dadmehr, 2003; Subasi, 2007), have been widely used to provide a quantitative measure of ∗

Corresponding author. Fax: +55 19 3521 5847. E-mail addresses: [email protected] (P. Amorim), [email protected] (T. Moraes), [email protected] (D. Fazanaro), [email protected] (J. Silva), [email protected] (H. Pedrini). http://dx.doi.org/10.1016/j.eswa.2016.09.037 0957-4174/© 2016 Elsevier Ltd. All rights reserved.

© 2016 Elsevier Ltd. All rights reserved.

the frequency distribution of the EEG and detect the presence of particular patterns. As an important tool for analyzing the human brain activity, EEG signals provide relevant information about epileptic processes. Small variations in EEG signals can show a certain type of brain abnormality. Effective signal processing methods for detecting such abrupt changes can help the diagnosis and treatment of patients with epilepsy. However, their development is still a very challenging task. This paper describes a novel work on epileptic EEG analysis through different time–frequency transforms, such as wavelets, shearlets and contourlets. Initially, the EEG signals are decomposed into frequency bands using each transform. Then, a set of features are extracted from the transform coefficients and used as input to different classifiers. The performance of the classification process is measured according to their capacity in detecting epileptic seizures from the data. The proposed method is applied to a set of EEG signals recorded from healthy subjects and epileptic patients. Its performance is compared to other classification approaches. To the best of our knowledge, this is the first work to employ shearlet and contourlet transforms to classify EEG signals. EEG signals contain non-stationary transient events and multiple frequency components that vary over time. Such signal characteristics motivated the use of time–frequency shearlet and contourlet transforms, which are appropriate to decompose the EEG components at different resolution levels. In contrast to the limited ability of wavelet transforms in decomposing signals only at horizontal, vertical and diagonal direction, shearlet and contourlet

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147

transforms provide an effective method for overcoming such directional limitations. Therefore, features extracted from the timefrequency representation can explore variations in amplitude and frequency components of the EEG signals (such as spikes, slow and sharp waves), which are used to discriminate between epileptic and normal EEG signals. The text is organized as follows. Section 2 briefly presents some concepts and works related to the topic under investigation. Section 3 describes the proposed methodology for EEG signal classification. Experimental results are described and discussed in Section 4. Section 5 presents the conclusions and directions for future work. 2. Background This section briefly reviews some relevant concepts and works related to the classification of EEG signals based on time–frequency transforms. 2.1. Electroencephalography Hans Berger (1873–1941), a German physiologist and psychiatrist, recorded the first human electroencephalogram (EEG) in 1924 (Haas, 2003). Through his device, he measured the electrical activity of the brain and conducted experiments to record the effects of drugs on EEG signal. The brain activity is depicted as waveforms of varying amplitude and frequency measured in microvoltage. Information about amplitude, frequency and shape of recorded EEG waveforms is combined with age of the patient, state of sleep or alertness, and location on the scalp to determine abnormal rhythms. The basic waveforms present in EEG recordings for clinical practice include alpha, beta, theta, and delta rhythms (Deuschl & Eisen, 1999; Tatum, 2014). Delta waves usually occur at a frequency from 0.5 to 4 Hz in young children or adult during sleep, theta waves at a frequency from 4 to 7 Hz in children and young adults, alpha waves are commonly recorded at 7 to 15 Hz when the person is awake but with closed eyes, beta waves occur at a frequency from 15 to 30 Hz, usually associated with depression, anxiety, or the use of sedatives, whereas gamma waves are recorded from 30 to 80 Hz related to attention, memory associations in visual discrimination tasks, among other cognitive processes. The advances in digital technology have improved the use of EEG to include specialized techniques that provide new capabilities besides traditional practices in clinical diagnosis. Long-term EEG recordings are still used for seizure prediction in patients with epilepsy. Advanced measures of abnormal EEG signals have received attention as possible biomarkers for different brain disorders, such as Alzheimer’s disease (Montez et al., 2009), Parkinson’s disease, schizophrenia (Sekimoto, Kato, Watanabe, Kajimura, & Takahashi, 2011), and parasomnia (Pilon, Zadra, Joncas, & Montplaisir, 2006). Epilepsy can be found in approximately 1% of the world population (Mormann, Andrzejak, Elger, & Lehnertz, 2007; World Health Organization, 2016). Epileptic seizures are abnormalities in EEG recordings characterized by short and episodic neuronal synchronous discharges that occur with considerably increased amplitude (Adeli et al., 2003). In general, the following terms are related to the different stages of seizures: (i) preictal refers to the state immediately before the actual seizure, which can last from minutes to days; (ii) ictal refers to the actual seizure, where physical changes occur in the person’s body; (iii) postictal refers to the state shortly after the seizure; (iv) interictal refers to the period between seizures or convulsions. The diagnosis of epilepsy depends on the detection of spikes (epileptiform discharges) or other abnormalities in interictal EEG signals. The ictal pattern is generally rhythmic and

141

often shows an increase or decrease in frequency and an increase in amplitude. The traditional visual inspection of EEG signals requires highly trained medical professionals and is time consuming. In order to address these difficulties, automatic techniques have been proposed to analyze epileptic EEG signals. 2.2. Transforms for signal analysis Several transforms have been proposed for signal analysis (Hramov, Koronovskii, Makarov, Pavlov, & Sitnikova, 2015; Vet´ & Goyal, 2014). Some useful properties of sigterli, Kovacˇ evic, nal transforms include their compact representation of a signal, inversibility, availability of fast versions for computer computation, capacity of analyzing signals at each frequency independently, among others. In signal processing, time–frequency analysis consists in techniques that study a signal in both time and frequency domains simultaneously. The classical Fourier analysis assumes that signals are periodic or infinite in time, while many signals in practice have short duration and change substantially over their duration. Some examples of time–frequency transforms are the short-time Fourier transform and the wavelet transform. A wavelet transform (Chui, 2014; Mallat, 1999; Zhai, Zhang, Liu, 2008) is a mathematical function used to decompose a continuoustime signal into different scale components. The wavelet transforms have certain advantages compared to Fourier transform since they do not require the use of fixed data windows and are localized in both time and frequency, whereas the Fourier transform is only localized in frequency. Wavelet transforms are particularly effective for representing different aspects of non-stationary, such as repeated patterns and discontinuities. The wavelet transform of a continuous signal f(t) is defined as

1 FW (a, b) = √ a





∞

f (t ) ψ ∗

−∞

t −b a

 dt

(1)

where a is positive and defined a scale and b is a real number and defines a translation. ψ (t) is usually referred to as the mother. The operator ∗ denotes the complex conjugate. Depending on the choice of mother wavelet, there are several wavelet families such as Haar, Daubechies, Morlet, Coiflets, Symlets, Meyer, Shannon, Gaussian, Cauchy, Gabor, among others. Despite the advantages of wavelet transforms, a drawback is their limited capacity of representing directional features, usually determined only at horizontal, vertical and diagonal degrees. Some variations of wavelet transforms (Starck, Murtagh, & Fadili, 2010) have been proposed to overcome these limitations, such as directional wavelets, bandelets, brushlets, contourlets, curvelets, directionlets, phaselets and ridgelets. Shearlet transforms (Lim, 2010; Schwartz, Silva, Davis, & Pedrini, 2011) possess important properties, providing a general structure for analyzing and representing data with anisotropic information at multiple levels of decomposition. Consequently, certain signal singularities, such as corners and edges, can be identified and located in images. The continuous shearlet transform for a signal in two dimensions (image) f is defined as the mapping

f → SHψ f (a, s, b) =  f, ψa,s,b 

(2)

where ψ is a generating function, a > 0 is the scale parameter, s ∈ R is the shear parameter, b ∈ R2 is the translation parameter, and the analyzing elements ψ a, s, b (shearlet basis functions) are given by

ψa,s,b (x ) = a−3/4 ψ (A−1 S−1 (x − b)) 

where A =

a 0



0 √ a



and S =

1 0



s . 1

(3)

142

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147

Fig. 1. Main stages of the proposed methodology for EEG analysis.

The shearlets ψ a, s, b form a collection of well-localized waveforms at various scales a, locations b and orientations s (Wang, Liu, & Yang, 2014). The contourlet transform (Do & Vetterli, 2005) is composed of a double filter bank structure that uses a Laplacian pyramid decomposition followed by directional filter banks applied on each bandpass sub-band. Such multiscale and directional decomposition stages provides directionality (basis elements are oriented at several directions), anisotropy (basis elements appear at various aspect ratios), multiresolution (signal is successively approximated from coarse to fine resolutions) and localization (basis elements in the representation are localized in both time and frequency). 3. Methodology This work describes and analyzes a method for classification of electroencephalogram (EEG) signals based on a set of features extracted from various time-frequency transforms. Different classifiers are employed and evaluated over a collection of signals. The main stages of the proposed methodology are illustrated in Fig. 1. Initially, we applied a second-order Butterworth band-pass filter to the EEG signals in order to maintain the frequency range of the signals between 0.5 and 40 Hz. A Butterworth filter was chosen due to its maximally flat magnitude response in the pass-band and better pulse response compared to other filters, such as Chebyshev and elliptic filters (Bianchi & Sorrentino, 2007). The filtered signals are input to three time-frequency transforms: wavelets, shearlets and contourlets. The following features (Subasi & Gursoy, 2010) were extracted from the transformed signals: i) Minimum of the absolute values of the coefficients in each subband. ii) Maximum of absolute values of the coefficients in each subband. iii) Mean of the absolute values of the coefficients in each subband. iv) Standard deviation of the coefficients in each sub-band. v) Average power of the wavelet coefficients in each sub-band. vi) Ratio of the absolute mean values of adjacent sub-bands. For the wavelet transforms, the features were extracted from all the detail coefficients and the final approximation coefficients. In this process, different levels of decomposition were evaluated for various wavelet families. The selection of appropriate wavelet basis

Table 1 Frequency bands corresponding to different levels of decomposition for wavelet transforms with a sampling frequency of 173.6 Hz. Decomposed signal

Frequency bands (Hz)

D1 D2 D3 D4 D5 A5

43.4–86.8 21.7–43.4 10.8–21.7 5.4–10.8 2.7–5.4 0.0–2.7

and number of decomposition levels is crucial for the analysis of signals through wavelet transforms. The number of decomposition levels is selected based on the dominant frequency components of the signal. According to the work by Subasi and Gursoy (2010), the levels are chosen such that those parts of the signal that correlate well with the frequencies required for the signal classification are retained in the wavelet coefficients. The frequency band [fm /2, fm ] of each level l is related to the sampling frequency fs of the original signal, which is given by fm = fs /2l+1 . The sampling frequency for the tested EEG signals is 173.6 Hz. Frequency bands corresponding to five decomposition levels for wavelet transforms are shown in Table 1. Thus, the signal is decomposed into five details D1–D5 and one approximation A5. For the shearlet and contourlet transforms, all their coefficients were used in the feature extraction. The EEG signals were converted from one-dimension to two-dimensions in order for the shearlet and contourlet transform to be applied. Each signal has 4097 samples, then an image with 64 × 64 size was created. Different patterns for epileptic and non-epileptic patients were observed in the images, as illustrated in Fig. 2 through a colormap. Pixels in red represent high amplitude, whereas pixels in blue represent low amplitude. The patterns were extracted from sets A to B, which are better described in Section 4.1. Different dimensionality reduction techniques, such as analysis of component analysis (PCA), independent component analysis (ICA) and linear discriminant analysis (LDA), were evaluated to reduce the feature space and provide a compact representation. Since a larger number of features can be incorporated into the proposed classification system, it is desirable to reduce the input space while maintaining the relevant information necessary for distinguishing among classes. Then, the reduced feature vectors can be used as

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147

143

Fig. 2. Image patterns generated from input EEG signals through (a) contourlets and (b) shearlets. The samples were extracted from five sets (A to E) described by Andrzejak et al. (2001).

input to a classification process. It is worth mentioning that the dimensionality reduction is an optional step in our method. Three classifiers were employed in the process: support vector machines (SVM), random forests (RF) and K-nearest neighbors (KNN). The EEG data sets composed of epileptic and non-epileptic classes were randomly divided into training and test as 60% and 40% with 5-fold cross-validation. 4. Experiments This section presents the experimental results obtained through the proposed EEG signal classification method, describes the data set used in the experiments, and performs a comparative analysis against other methods available in the literature. 4.1. EEG data set In our experiments, we used the publicly available EEG data set described by Andrzejak et al. (2001). The data set is composed of five sets, denoted from A to E, each one containing 100 single-channel EEG segments with 4097 samples of 23.6 s. The data sets A and B were recorded from healthy volunteers with eyes opened and closed, respectively. The data sets C and D were recorded from patients before an epileptic attack (during interictal periods) at hemisphere hippocampal formation and the epileptogenic zone, respectively. The data set E was recorded from patients during an epilepsy attack (ictal periods). All the data was acquired with a sampling rate of 173.61 Hz using 12-bit resolution. A bandpass filter of 40 Hz was applied to the data, as mentioned in the manuscript that describes the data set. Examples of EEG signals taken from different subjects are shown in Fig. 3. 4.2. Performance analysis Data samples can be classified as belonging or not to one specific class. Samples correctly classified and misclassified are labeled, respectively, as true positive (TP) and false negative (FN). On the other hand, samples of another class can be classified as true negative (TN) or false positive (FP). These outcomes are expressed in terms of sensitivity or true positive rate (TPR), and specificity or true negative rate (TNR) as

Sensitivity = TPR =

TP TP + FN

(4)

Fig. 3. Examples of five different sets of EEG signals taken from different subjects.

Specificity = TNR =

TN TN + FP

(5)

The F-measure for a given class is calculated as the harmonic mean of the sensitivity and precision values for that specific class

F-measure = 2 ×

Sensitivity × Precision Sensitivity + Precision

(6)

which provides a global assessment of the classifier performance on each class The accuracy of the classifier is the percentage of the test data set correctly classified, expressed as

Accuracy =

TN + TP TN + TP + FN +FP

(7)

These metrics allow us to evaluate the performance of the classification techniques considered in the proposed methodology. 4.3. Results In several works (Lima, Coelho, & Chagas, 2009; Patnaik & Manyam, 2008; Subasi, 2005; 2007), only the sets A and E were

144

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147

employed to assess the classifier performance. Other works (Güler, Übeyli, & Güler, 2005; Übeyli, 2008) consider three out of the five sets, specifically sets A, D and E. In our study, we have considered the complete data set of 500 EEG segments (multiclass classification problem), making the classification problem more difficult to be solved, however, allowing for a more systematic analysis. Additionally, we conducted experiments by considering only the sets A and E (binary classification problem) to demonstrate the effectiveness of the shearlet and contourlet transforms in contrast with other existing approaches. Since the EEG signals usually do not have useful frequency components of the signal above 30 Hz, the number of decomposition levels in our approach was selected to be 5. In the wavelet-based approach, the signal was decomposed into the detail sub-bands D1–D5 and one approximation subband A5. The wavelet families used in our evaluation include Haar, Daubechies, Coiflets, Biorthogonals, Reverse Biorthogonals, and Symlets. For the shearlet transform, each element ψ a, s, b in Eq. (3) is locally supported on a box of size 25 × 25 , that is, a shearing filter with size of 32 × 32. A Laplacian pyramid scheme is used to decompose the signal into a low-pass and high-pass images. For the contourlet transform, a two-dimensional filter bank decomposes the image into several directional sub-bands at multiple scales (5 scales in our case). This process is performed by combining a Laplacian pyramid with a directional filter bank at each sub-band (scale). After the decomposition of the EEG signals, the six features mentioned in Section 3 were extracted from each sub-band and concatenated together, yielding a feature vector with a total of 36 attributes per sample for the wavelet transforms, 48 attributes for the shearlet transforms, and 48 attributes per sample for the contourlet transforms. The extracted features were reduced through three techniques for dimensionality reduction. This stage is particularly important when the number of features becomes excessively large due to their concatenation to form the final feature vector, as well as an increase in the number of wavelet decomposition levels or sample size. In our experiments, although the reduction in the data dimensionality demonstrated to be suitable for suppressing certain redundant or correlated attributes and improving convergence time, it was not worthwhile in terms of classification accuracy. Therefore, the following experiments do not consider dimensionality reduction in their configuration. In this work, the radial basis function (RBF) kernel was used as the kernel function of SVMs. The hyperparameters C = 1.0 and γ = 1/n of the SVM classifier were determined through a 5-fold cross-validation grid search performed for each of the 5 training sets used to calculate the accuracy of the classifier, where n is the number of features used in the classification. For the KNN classifier, the Dynamic Time Warping (DTW) was used as distance measure and K = 1. Furthermore, a forest of 10 trees was used in our experiments for the RF classifier. Table 2 shows the accuracy values corresponding to the classification results of the five data sets using different wavelet bases with KNN, SVM and RF classifiers. The best results for each classifier are highlighted in bold. As it can be observed from the table, the random forest (RF) classifier produced the highest accuracy rates. A comparison of the classification results on all five data sets for wavelet, shearlet and contourlet transforms is shown in Table 3. The wavelet bases used with KNN, SVM and RF are Coif1, Haar and Rbio2.2, respectively, since they achieved the best results for these classifiers. We performed the ANOVA test and obtained the p-values shown in the legend of Fig. 4. Since the p-values are less than

Table 2 Classification results using different wavelet bases. The five data sets are classified with KNN, SVM and RF. Wavelet

Haar Db2 Db3 Db4 Db5 Db6 Coif1 Coif2 Coif3 Coif4 Coif5 Bior1.1 Bior1.3 Bior1.5 Bior2.2 Bior2.4 Bior2.6 Bior3.1 Bior3.3 Bior3.5 Rbio1.1 Rbio1.3 Rbio1.5 Rbio2.2 Rbio2.4 Rbio2.6 Rbio3.1 Rbio3.3 Rbio3.5 Sym2 Sym3 Sym4

Accuracy KNN

SVM

RF

62.5 62.5 59.5 53.0 53.5 53.5 65.0 55.0 60.5 58.0 58.0 62.5 64.5 63.5 60.5 60.0 60.5 55.5 55.5 54.5 62.5 54.0 53.0 62.5 56.0 58.5 60.0 62.0 50.0 62.5 59.5 58.5

65.6 63.2 59.6 54.8 54.2 55.0 64.2 57.6 55.2 50.0 47.7 63.4 65.2 64.0 57.0 63.4 63.6 55.2 52.0 56.0 63.6 61.4 53.8 63.4 58.2 52.2 64.8 57.6 55.4 62.5 55.0 54.5

58.2 62.5 54.4 49.1 51.6 49.1 61.5 49.6 45.2 44.3 44.6 57.4 55.1 56.6 59.4 55.3 54.3 20.0 39.2 34.4 57.5 58.1 48.4 63.7 53.6 52.2 61.2 58.8 53.8 62.5 45.0 59.0

Table 3 Classification results on all five data sets for wavelet, shearlet and contourlet transforms. Transform

Wavelets

Shearlets

Contourlets

Classifier

Accuracy

Specificity

Sensitivity

F-measure

KNN SVM RF KNN SVM RF KNN SVM RF

0.645 0.680 0.700 0.680 0.740 0.795 0.655 0.675 0.815

0.670 0.680 0.700 0.680 0.750 0.800 0.640 0.690 0.817

0.650 0.680 0.700 0.680 0.740 0.800 0.660 0.680 0.815

0.650 0.680 0.700 0.670 0.750 0.790 0.640 0.670 0.813

Fig. 4. Post-hoc analysis of Tukey’s test for the evaluated methods.

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147 Table 4 Confusion matrices for each data set classified with KNN, SVM and RF for their best accuracy. The features were generated through shearlet transform. True class

A

B

C

D

E

A

B

C

D

E

Table 6 Confusion matrices for each data set (A and E) classified with KNN, SVM and RF for their best accuracy. The features were generated through shearlet transform.

Classified as

KNN

SVM

RF

True class

Classified as

KNN

SVM

RF

A B C D E A B C D E A B C D E A B C D E A B C D E

80.00 10.00 10.00 0.00 0.00 15.00 82.50 0.00 2.50 0.00 10.00 2.50 57.50 30.00 0.00 17.50 0.00 47.50 30.00 5.00 0.00 5.00 5.00 0.00 90.00

87.50 7.50 5.00 0.00 0.00 12.50 87.50 0.00 0.00 0.00 0.00 2.50 57.50 40.00 0.00 2.50 0.00 42.50 55.00 0.00 0.00 10.00 2.50 2.50 85.00

85.00 15.00 0.00 0.00 0.00 12.50 85.00 2.50 0.00 0.00 0.00 5.00 77.50 15.00 2.50 0.00 0.00 32.50 60.00 7.50 0.00 5.00 0.00 5.00 90.00

A

A E A E

100.00 0.00 2.50 97.50

100.00 0.00 0.00 100.00

100.00 0.00 0.00 100.00

Table 5 Confusion matrices for each data set classified with KNN, SVM and RF for their best accuracy. The features were generated through contourlet transform. True class

145

Classified as

KNN

SVM

RF

A B C D E A B C D E A B C D E A B C D E A B C D E

85.00 7.50 7.50 0.00 0.00 12.50 75.00 2.50 10.00 0.00 15.00 17.50 47.50 20.00 0.00 17.50 10.00 37.50 30.00 5.00 0.00 5.00 2.50 2.50 90.00

85.00 12.50 2.50 0.00 0.00 25.00 70.00 5.00 0.00 0.00 12.50 7.50 55.00 25.00 0.00 5.00 10.00 40.00 45.00 0.00 0.00 12.50 0.00 5.00 82.50

85.00 12.50 0.00 2.50 0.00 15.50 82.50 0.00 0.00 0.00 0.00 7.50 82.50 10.00 0.00 0.00 7.50 25.00 62.50 5.00 0.00 2.50 0.00 2.50 95.00

a significance level of 0.05, this indicates that there is a significant difference in means between the methods. A post-hoc analysis based on Tukey’s HSD (honest significant difference) test is illustrated in the figure. Tables 4 and 5 show the confusion matrices for each data set (tests conducted on five data sets) classified with KNN, SVM and RF for their best accuracy, where the features were extracted from image patterns generated through shearlet and contourlet transforms, respectively.

E

Table 7 Confusion matrices for each data set (A and E) classified with KNN, SVM and RF for their best accuracy. The features were generated through contourlet transform. True class

Classified as

KNN

SVM

RF

A

A E A E

100.00 0.00 0.00 100.00

100.00 0.00 0.00 100.00

100.00 0.00 0.00 100.00

E

Tables 6 and 7 show the confusion matrices for each data set (tests conducted only on sets A and E) classified with KNN, SVM and RF for their best accuracy, where the features were extracted from image patterns generated through shearlet and contourlet transforms, respectively. As it can be observed from the results, our methodology based on shearlet and contourlet transforms achieves high accuracy rates to classify electroencephalogram signals. The overall accuracy rates for classifying the five data sets (A to E) were 79.5% and 81.5% using the RF classifier for shearlet and contourlet transforms, respectively, whereas the overall accuracy rates for two data sets (A and E) were 100% using SVM and RF classifiers for both transforms. The method is straightforward to implement, such that it can be used as a tool for supporting the identification and treatment of epilepsy patients. Table 8 reports classification rates for some existing method for the evaluated database. It is important to mention that the methods consider the problem as a binary classification (normal or seizure) through one of following combinations of the EEG data sets (A, B, C, D or E): (i) classification of sets A and E, where segment A is normal and E is seizure; (ii) classification of sets A, D and E, where segment A is considered normal and segments (D, E) are seizure; (iii) classification of sets A, B, C, D and E, where segments (A, B) are normal and (C, D, E) are seizure. 5. Conclusions and future work Electroencephalogram (EEG) signals have been extensively used for detection of epileptic seizures. There still exists a demand for automatic and effective methods for scanning and identifying abnormal activities in EEG recordings to support neurophysiologists. In this work, we introduced and evaluated a novel classification approach based on shearlet and contourlet transforms to decompose EEG signals into frequency bands. Several features were extracted from these time-frequency coefficients and used as input to three types of classifiers. Experiments were conducted on EEG recordings to demonstrate the effectiveness of the proposed classification method. Properties of multiresolution, time-frequency localization, compact support and anisotropy make the shearlets and contourlets an accurate representation of EEG signal information. From the obtained results, it is possible to observe that the shearlet and contourlet transforms are capable of achieving high classification accuracy rates. The method is robust and straightforward to implement, contributing to valuable clinical information.

146

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147 Table 8 Classification rates (in percentage) for existing methods in the literature for the evaluated EEG database. Author

Method

Data set

Accuracy (%)

Nigam and Graupe (2004) Kannathal, Choo, Acharya, and Sadasivan (2005) Güler et al. (2005) Srinivasan, Eswaran, and Sriraam (2005) Ataee, Avanaki, Shariatpanahi, and Khoee (2006) Jahankhani, Kodogiannis, and Revett (2006) Subasi (2007) Tzallas, Tsipouras, and Fotiadis (2007) Polat and Günes¸ (2007) Übeyli (2008) Tzallas, Tsipouras, and Fotiadis (2009) Tzallas et al. (2009) Guo, Rivero, and Pazos (2010) Liang, Wang, and Chang (2010) Liang et al. (2010) Guo et al. (2010) Orhan, Hekim, and Ozer (2011) Iscan, Dokur, and Demiralp (2011) Nicolaou and Georgiou (2012) Nicolaou and Georgiou (2012) Nicolaou and Georgiou (2012) Alam and Bhuiyan (2013) Alam and Bhuiyan (2013) Zainuddin, Huong, and Pauline (2013) Xie and Krishnan (2013) Das, Bhuiyan, and Alam (2014) Das et al. (2014) Das et al. (2014) Kumar, Dewal, and Anand (2014) Kumar et al. (2014) Chen (2014) Fu, Qu, Chai, and Zou (2015) Fu et al. (2015) Das, Bhuiyan, and Alam (2016) Das et al. (2016)

Non-linear preprocessing filter + Neural networks Entropy + Adaptive neuro-fuzzy inference system Lyapunov exponents + Recurrent neural networks Time-frequency features + Recurrent neural networks Wavelet features + Neural network Wavelet features + Neural networks Wavelet features + Expert system Time-frequency analysis + Neural networks Fourier features + Decision tree Wavelet features + Expert model Time-frequency analysis and power spectral density + Neural networks Time-frequency analysis and power spectral density + Neural networks Multiwavelet-entropy features + Neural networks Spectral analysis and principal component analysis + Genetic algorithms Spectral analysis and principal component analysis + Genetic algorithms Multiwavelet-entropy features + Neural networks Wavelet features + k-means clustering + Neural networks Cross correlation and power spectral density + Support vector machines Permutation entropy and support vector machines Permutation entropy and support vector machines Permutation entropy and support vector machines Time-frequency analysis and higher order statistics + Neural networks Time-frequency analysis and higher order statistics + Neural networks Wavelet features + Neural networks Wavelet variances + Nearest neighbors Dual-tree complex wavelets + Support vector machines Dual-tree complex wavelets + Support vector machines Dual-tree complex wavelets + Support vector machines Wavelet features and approximate entropy + Neural networks Wavelet features and approximate entropy + Neural networks Dual-tree complex wavelet-Fourier features + Nearest neighbors Hilbert marginal spectrum analysis + Support vector machines Hilbert marginal spectrum analysis + Support vector machines Dual-tree complex wavelets + inverse Gaussian + Support vector machines Dual-tree complex wavelets + inverse Gaussian + Support vector machines

A-E A-E A-E A-E A-E A-E A-E A/B/C/D - E A-E A/D - E A-E A/E - E A/B/C/D - E A/D - E D-E A-E A-E A-E D-E C-E A-E AB/CD - E A-E A-E A-E AB/CD - E A-E A/D - E A/B/C/D - E A-E A-E A/B/C/D - E A-E AB/CD - E A-E

97.20 92.22 96.94 99.60 94.00 97.00 95.00 97.73 98.72 96.89 100.00 100.00 98.27 98.67 98.74 99.85 99.60 100.00 79.94 88.00 93.55 80.00 100.00 98.87 100.00 83.5 100.00 96.8 94.00 100.00 100.00 98.80 99.85 96.28 100.00

As directions for future work, we intend to investigate the use of other time–frequency transforms and design new features for improving the classification process. Acknowledgements The authors are grateful to FAPESP - São Paulo Research Foundation (Grant 2011/22749-8) and CNPq (Grant 307113/2012-4) for their financial support. References Abou-Khalil, B., & Misulis, K. E. (2006). Atlas of EEG & seizure semiology. Elsevier, Butterworth Heinehmann Edition. Adeli, H., Zhou, Z., & Dadmehr, N. (2003). Analysis of EEG records in an epileptic patient using wavelet transform. Journal of Neuroscience Methods, 123, 69–87. Alam, S. S., & Bhuiyan, M. I. H. (2013). Detection of seizure and epilepsy using higher order statistics in the EMD domain. IEEE Journal of Biomedical and Health Informatics, 17, 312–318. Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., & Elger, C. E. (2001). Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E, 64, 061907. Ataee, P., Avanaki, A. N., Shariatpanahi, H. F., & Khoee, S. M. (2006). Ranking features of wavelet-decomposed EEG based on significance in epileptic seizure prediction. In 14th European signal processing conference (pp. 1–4). IEEE. Bianchi, G., & Sorrentino, R. (2007). Electronic filter simulation & design. McGraw-Hill New York. Bronzino, J. D. (20 0 0). Principles of electroencephalography: The biomedical engineering handbook (2nd ed.). Boca Raton: CRC Press LLC. Chen, G. (2014). Automatic EEG seizure detection using dual-tree complex wavelet– Fourier features. Expert Systems with Applications, 41, 2391–2394. Chui, C. K. (2014). An introduction to wavelets: 1. Academic Press. Das, A. B., Bhuiyan, M. I. H., & Alam, S. S. (2014). A statistical method for automatic detection of seizure and epilepsy in the dual tree complex wavelet transform domain. In International conference on informatics, electronics & vision (pp. 1–6). IEEE.

Das, A. B., Bhuiyan, M. I. H., & Alam, S. S. (2016). Classification of EEG signals using normal inverse Gaussian parameters in the dual-tree complex wavelet transform domain for seizure detection. Signal, Image and Video Processing, 10, 259–266. Deuschl, G., & Eisen, A. (1999). Recommendations for the practice of clinical neurophysiology (guidelines of the international federation of clinical neurophysiology). Electroencephalography and Clinical Neurophysiology - Supplement. Do, M. N., & Vetterli, M. (2005). The contourlet transform: An efficient directional multiresolution image representation. IEEE Transactions on Image Processing, 14, 2091–2106. Fu, K., Qu, J., Chai, Y., & Zou, T. (2015). Hilbert marginal spectrum analysis for automatic seizure detection in EEG signals. Biomedical Signal Processing and Control, 18, 179–185. Güler, N. F., Übeyli, E. D., & Güler, I. (2005). Recurrent neural networks employing lyapunov exponents for EEG signals classification. Expert Systems with Applications, 29, 506–514. Guo, L., Rivero, D., & Pazos, A. (2010). Epileptic seizure detection using multiwavelet transform based approximate entropy and artificial neural networks. Journal of Neuroscience Methods, 193, 156–163. Haas, L. (2003). Hans Berger (1873–1941), Richard Caton (1842–1926), and electroencephalography. Journal of Neurology, Neurosurgery & Psychiatry, 74, 9. Hramov, A. E., Koronovskii, A. A., Makarov, V. A., Pavlov, A. N., & Sitnikova, E. (2015). Mathematical methods of signal processing in neuroscience. In Wavelets in neuroscience (pp. 1–13). Springer. Iscan, Z., Dokur, Z., & Demiralp, T. (2011). Classification of electroencephalogram signals with combined time and frequency features. Expert Systems with Applications, 38, 10499–10505. Jahankhani, P., Kodogiannis, V., & Revett, K. (2006). EEG signal classification using wavelet feature extraction and neural networks. In IEEE John Vincent Atanasoff 2006 international symposium on modern computing (pp. 120–124). IEEE. Kannathal, N., Choo, M. L., Acharya, U. R., & Sadasivan, P. (2005). Entropies for detection of epilepsy in EEG. Computer Methods and Programs in Biomedicine, 80, 187–194. Kumar, Y., Dewal, M., & Anand, R. (2014). Epileptic seizures detection in EEG using DWT-based ApEn and artificial neural network. Signal, Image and Video Processing, 8, 1323–1334. Liang, S.-F., Wang, H.-C., & Chang, W.-L. (2010). Combination of EEG complexity and spectral analysis for epilepsy diagnosis and seizure detection. EURASIP Journal on Advances in Signal Processing, 2010, 1.

P. Amorim et al. / Expert Systems With Applications 67 (2017) 140–147 Lim, W.-Q. (2010). The discrete shearlet transform: A new directional transform and compactly supported shearlet frames. IEEE Transactions on Image Processing, 19, 1166–1180. Lima, C. A., Coelho, A. L., & Chagas, S. (2009). Automatic EEG signal classification for epilepsy diagnosis with relevance vector machines. Expert Systems with Applications, 36, 10054–10059. Mallat, S. (1999). A wavelet tour of signal processing. Academic Press. Montez, T., Poil, S.-S., Jones, B. F., Manshanden, I., Verbunt, J. P., van Dijk, B. W., et al. (2009). Altered temporal correlations in parietal alpha and prefrontal theta oscillations in early-stage Alzheimer disease. Proceedings of the National Academy of Sciences, 106, 1614–1619. Mormann, F., Andrzejak, R. G., Elger, C. E., & Lehnertz, K. (2007). Seizure prediction: The long and winding road. Brain, 130, 314–333. Nicolaou, N., & Georgiou, J. (2012). Detection of epileptic electroencephalogram based on permutation entropy and support vector machines. Expert Systems with Applications, 39, 202–209. Niedermeyer, E., & da Silva, F. L. (2005). Electroencephalography: Basic principles, clinical applications, and related fields. Lippincott Williams & Wilkins. Nigam, V. P., & Graupe, D. (2004). A neural-network-based detection of epilepsy. Neurological Research, 26, 55–60. Orhan, U., Hekim, M., & Ozer, M. (2011). EEG signals classification using the k-means clustering and a multilayer perceptron neural network model. Expert Systems with Applications, 38, 13475–13481. Patnaik, L. M., & Manyam, O. K. (2008). Epileptic EEG detection using neural networks and post-classification. Computer Methods and Programs in Biomedicine, 91, 100–109. Pilon, M., Zadra, A., Joncas, S., & Montplaisir, J. (2006). Hypersynchronous delta waves and somnambulism: Brain topography and effect of sleep deprivation. Sleep, 29, 77. Polat, K., & Günes¸ , S. (2007). Classification of epileptiform EEG using a hybrid system based on decision tree classifier and fast Fourier transform. Applied Mathematics and Computation, 187, 1017–1026. Schwartz, W. R., Silva, R. D. d., Davis, L. S., & Pedrini, H. (2011). A novel feature descriptor based on the shearlet transform. In 18th IEEE international conference on image processing (pp. 1033–1036). IEEE. Sekimoto, M., Kato, M., Watanabe, T., Kajimura, N., & Takahashi, K. (2011). Cortical regional differences of delta waves during all-night sleep in Schizophrenia. Schizophrenia Research, 126, 284–290.

147

Srinivasan, V., Eswaran, C., & Sriraam, N. (2005). Artificial neural network based epileptic detection using time-domain and frequency-domain features. Journal of Medical Systems, 29, 647–660. Starck, J.-L., Murtagh, F., & Fadili, J. (2010). Sparse image and signal processing: Wavelets, curvelets, morphological diversity. Cambridge University Press. Subasi, A. (2005). Epileptic seizure detection using dynamic wavelet network. Expert Systems with Applications, 29, 343–355. Subasi, A. (2007). EEG signal classification using wavelet feature extraction and a mixture of expert model. Expert Systems with Applications, 32, 1084–1093. Subasi, A., & Gursoy, M. I. (2010). EEG signal classification using PCA, ICA, LDA and support vector machines. Expert Systems with Applications, 37, 8659–8666. Tatum, W. O. (2014). Ellen R. Grass lecture: Extraordinary EEG. The Neurodiagnostic Journal, 54, 3–21. Tzallas, A., Tsipouras, M., & Fotiadis, D. (2007). Automatic seizure detection based on time-frequency analysis and artificial neural networks. Computational Intelligence and Neuroscience, 2007, 80510, 13 pages. Tzallas, A. T., Tsipouras, M. G., & Fotiadis, D. I. (2009). Epileptic seizure detection in EEGs using time-frequency analysis. IEEE Transactions on Information Technology in Biomedicine, 13, 703–710. Übeyli, E. D. (2008). Wavelet/mixture of experts network structure for EEG signals classification. Expert Systems with Applications, 34, 1954–1962. ´ J., & Goyal, V. K. (2014). Foundations of signal processing. CamVetterli, M., Kovacˇ evic, bridge University Press. Wang, X.-Y., Liu, Y.-C., & Yang, H.-Y. (2014). An efficient remote sensing image denoising method in extended discrete shearlet domain. Journal of Mathematical Imaging and Vision, 49, 434–453. World Health Organization (2016). Epilepsy - fact sheet #999. Xie, S., & Krishnan, S. (2013). Wavelet-based sparse functional linear model with applications to EEGs seizure detection and epilepsy diagnosis. Medical & Biological Engineering & Computing, 51, 49–60. Zainuddin, Z., Huong, L. K., & Pauline, O. (2013). Reliable epileptic seizure detection using an improved wavelet neural network. The Australasian Medical Journal, 6, 308–314. Zhai, J.-H., Zhang, S.-F., & Liu, L.-J. (2008). Image recognition based on wavelet transform and artificial neural networks. In International conference on machine learning and cybernetics: 2 (pp. 789–793). IEEE.