1
EMD based Temporal and Spectral Features for the Classification of EEG Signals Using Supervised Learning Farhan Riaz1 , Ali Hassan1 , Saad Rehman1 , Imran Khan Niazi2,3 and Kim Dremstrup3
Abstract—This paper presents a novel method for feature extraction from electroencephalogram (EEG) signals using empirical mode decomposition (EMD). Its use is motivated by the fact that the EMD gives an effective time-frequency analysis of non-stationary signals. The intrinsic mode functions (IMF) obtained as a result of EMD give the decomposition of a signal according to its frequency components. We present the usage of upto third order temporal moments, and spectral features including spectral centroid, coefficient of variation and the spectral skew of the IMFs for feature extraction from EEG signals. These features are physiologically relevant given that the normal EEG signals have different temporal and spectral centroids, dispersions and symmetries when compared with the pathological EEG signals. The calculated features are fed into the standard support vector machine (SVM) for classification purposes. The performance of the proposed method is studied on a publicly available dataset which is designed to handle various classification problems including the identification of epilepsy patients and detection of seizures. Experiments show that good classification results are obtained using the proposed methodology for the classification of EEG signals. Our proposed method also compares favorably to other state-of-the-art feature extraction methods. Index Terms—Empirical Mode Decomposition, Feature Extraction, Classifiction.
I. I NTRODUCTION
E
LECTROENCEPHALOGRAM (EEG) is a set of electric potential differences that contain the information about the human brain activity. It exhibits the data regarding the volume currents that spread from a neural tissue throughout the conductive media of the brain. These measurements can be obtained using sensors placed on the sclap or using the intracranial electrodes. The EEG signals can be effectively used for various applications such as emotion recognition [9], brain computer interfaces (BCI) [13] etc. One of the most important applications of the analysis of EEG signals is its use in neuroscience to diagnose diseases and brain disorders. Epileptic seizure is one of the most common neurological disorders worldwide [14], [15]. Its detection is typically done by the physicians using a visual scanning of the EEG signals which is a time consuming process and may be inaccurate [10], 1 These authors are with National of Sciences and Technology, Islamabad, Email: farhan.riaz|alihassan|saadrehman@
University Pakistan.
ceme.nust.edu.pk 2 This author is with New Zealand College of Chiropractic, Auckland, New Zealand. Email:
[email protected] 3 These authors are with Department of Health Science and Technology, University of Aalborg, Denmark. Email: imrankn|
[email protected]
[16]. These inaccuracies are particularly significant for long time duration EEG signals [17]. The parameters extracted from the EEG signals using various signal processing methods are very useful for diagnostics. The spectral parameters based on the Fourier transform are useful for analysing the EEG signals and have shown good results on their classification [1], [2]. However, it is important to note that the Fourier domain does not exhibit any time-domain characteristics in the signal giving the features which are suboptimal for feature extraction from some signal processing scenarios [18]. Several other methods based on time-frequency domain have been developed for the detection of epileptic seizures from EEG signals. These methods include the use of short time Fourier transform (STFT) [3], [19]. Although good results are obtained using these methods, the STFT does not yield a multiresolution analysis of the signals. This is because of the fact that the STFT uses the filters of the same bandwidth for signal decomposition at all frequencies. This limitation is typically resolved using the wavelet analysis in which a multiresolution time-frequency analysis is facilitated by forming band pass filters with varying bandwidths [20]– [23]. Researchers have found the wavelet analysis to be a very useful tool for various signal processing applications [24]– [28]. Recently in the context of EEG signal processing, Neethu et al. [4] has proposed the so-called Wavelet CSP (Common Spatial Pattern) algorithm for processing of the EEG signals. Good results have been obtained using the proposed methodology. Nadia et al. [5] proposed a method of artifact removal from EEG signals using a wavelet ICA (Independent Component Analysis) based method giving good results on suppression of artifacts in EEG signals. Despite a wide usage of wavelets in the literature for signal processing, an underlying assumption in wavelet analysis is that the signal that is being analysed is stationary [10]. Previous studies on the analysis of EEG signals have shown that their frequency components change over a period of time making them nonstationary [6], [29], [30]. Hence, the signal processing methods which are more suitable for such signals are desired. More recently, new techniques for the analysis of nonstationary and non-linear signals have been proposed which are mainly based on empirical mode decomposition (EMD) [31]. The EMD is a time-frequency based method which decomposes a signals into a number of intrinsic mode functions (IMF) which are oscillatory components. This is an empirical method and thus is effective for a time-frequency analysis of the non-stationary signals. This characteristic of EMD has
2
TABLE I: A brief review of related work on EEG signal processing and classification Reference
Method
Dataset size
Objective
Year
Srinivasan et al. [1] Polat et al. [2] Tzallas et al. [3] Robinson et al. [4] Nadia et al. [5] Pachori et al. [6] Oweis et al. [7] Alam et al. [8] Petrantokanis et al. [9] Bajaj et al. [10] A. Subasi [11], [12]
time + frequency + ANN FFT + Decision trees STFT + ANN Wavelet CSP + Fisher Linear Discriminant Wavelet ICA EMD Hilbert Huang transform + Clustering EMD statistics + ANN Higher Order Crossings + SVM EMD + SVM Wavelets + ANN
200 200 500 2800 Unknown 500 500 500 960 500 500
Epilepsy Epilepsy Seizure Hand movement speed Noise removal Seizure Seizure Seizure + Epilepsy Emotion recognition Seizure Seizure
2005 2007 2009 2013 2012 2008 2011 2013 2010 2012 2007
motivated the researchers to use it for the analysis of EEG signals. In [32], the mean frequency of IMFs has been used for the classification of EEG signals. Oweis et al. [7] have used the analytic IMFs (Hilbert-Huang transform - HHT) for seizure classification in EEG signals. They have used the weighted frequencies in the IMFs to identify seizures in the EEG signals. Bajaj et al. [10] have used the amplitude modulation bandwidth and frequency modulation bandwidth of IMFs for the classification between seizure and nonseizure EEG signals. Alam et al. [8] have shown that the higher order statistics of the IMFs are adequate for the classification of EEG signals. Besides the strengths of feature extraction methods related to instantaneous frequencies (IF), it is important to note that the extraction of IF is more meaningful when the IMFs extracted from the EEG signals are monocomponent [38] (which is not true in our case). In this paper, we propose a novel feature extraction methodology for the classification of EEG signals involving three stages. The first stage of the algorithm involves the calculation of EMD of the EEG signal, giving a set of IMFs. The first three IMFs are selected for further processing. The second stage involves feature extraction which is done by calculating the temporal and spectral characteristics of the IMFs, which is the main contribution of this paper. For the calculation of spectral features, we have used power spectral density (PSD). The temporal and spectral features are obtained from the Hilbert transformed IMFs as using this transformation can remove the DC offset from the spectral content of the signals which is one of the sources of non-stationarity in the signals. The third stage involves the use of support vector machine (SVM) for the classification of EEG signals. The rest of the paper is organized as follows: We will describe the dataset used in this paper (Section II) followed by the proposed framework for classifying the EEG signals (Section III). Later, we present our experimental results in comparison with other recent methods (Section IV) and conclude the paper (Section V).
signals have been selected from continuous multichannel EEG recording after visual inspection of artifacts. The Sets A and B consist of surface EEG segments collected from five healthy volunteers in awaken and relaxed state with their eyes opened and closed respectively. Segments in Sets C, D and E are obtained from an archive of EEG signals of presurgical diagnosis. Five patients are selected who have achieved complete control of seizure after resection of one of the hippocampal formations. These resection sites are thus diagnosed as epileptogenic zone. Sets C and D consist of EEG epochs recorded during seizure free intervals (i.e., interictal) from epileptogenic zone and hippocampal formation of the opposite hemisphere, respectively. Set E contains signals corresponding to seizure attacks (i.e., ictal EEG), recorded using all the electrodes. The signals are recorded in a digital format at a sampling rate of 173.61 Hz [8]. Thus, the sample length of each segment is 173.61×23.6≈4097.
200
0
-200 0
500
1000
1500
2000
2500
3000
3500
4000
0
500
1000
1500
2000
2500
3000
3500
4000
0
500
1000
1500
2000
2500
3000
3500
4000
0
500
1000
1500
2000
2500
3000
3500
4000
0
500
1000
1500
2000
2500
3000
3500
4000
500
0
-500 500
0
-500 200
0
-200 2000
II. DATASET In this study, we have used an EEG dataset that is publicly available online [33]. The dataset consists of five subsets (denoted as sets A-E) each containing 100 single channel EEG signals, each one having a duration of 23.6 seconds. These
0
-2000
Fig. 1: Sample EEG signals from five different sets from rows 1 to 5 (A, B, C, D and E respectively).
3
0.5 Frequency
Power
0.5
Amplitude
Amplitude
0
0.5 Frequency
0 -1
0
1000
2000 3000 Samples
0 -1
0
1000
2000 3000 Samples
4000
1 0 -1
0
4000
1
0
1000
2000 3000 Samples
4000
Power Power 0.5 Frequency
0
0.5 Frequency
0 -1
0
1000
2000 3000 Samples
0 -1
0
1000
2000 3000 Samples
0 0
1000
2000 3000 Samples
4000
1
0
0.5 Frequency
1
0
0.5 Frequency
1
1 0.5 0
4000
1
-1
0
4000
1
0.5 Frequency
1
1
1
0
0.5
1
0.5
1
1
0
1
Amplitude
Amplitude
0
1
0
0
1
0.5
1
1
0.5 Frequency
Power
0
0
Amplitude
Power
1 0.5 0
0
1
0.5
Amplitude
0.5 Frequency
1
Amplitude
0
IMF3
0.5
Amplitude
Power
Power
0.5 0
Power
IMF2 1
Amplitude
Power
IMF1 1
1 0 -1
0
1000
2000 3000 Samples
4000
0
1000
2000 3000 Samples
4000
0
1000
2000 3000 Samples
4000
1 0 -1
1 0 -1
Fig. 2: Plots of temporal and spectral signals obtained from the IMFs after EMD decomposition. The first three rows contain the plots PSDs of the IMFs while the last three rows contain the plots of IMF amplitudes (Row 1 and 4 - Normal;Row 2 and 5 - Interictal;Row 3 and 6 - Seizure).
III. M ETHODS In this section, we describe the proposed methodology for the feature extraction from EEG signals. A. Empirical Mode Decomposition - EMD The EMD is a data dependent method of decomposing a signal into a number of oscillatory components, known as intrinsic mode functions (IMFs). EMD does not make any assumptions about the stationarity or linearity of the data. The aim of EMD is to decompose a signals into a number of IMFs, each one of them satisfying the two basic conditions: 1). the number of extrema or zero crossings must be the same or differ by at most one; 2). at any point, the average value of the envelope defined by local maxima and the envelope defined by the local minima is zero. Given that we have a signal x(t), the calculation of its IMFs involves the following steps [34]. 1) Identify all extrema (maxima and minima) in x(t). 2) Interpolate between minima and maxima, generating the envelopes el (t) and em (t). l (t) . 3) Determine the local mean as a(t) = em (t)+e 2
4) Extract the detail i.e. h1 (t) = x(t) − a(t). 5) Decide whether h1 (t) is an IMF or not based on two basic conditions for IMFs mentioned above. 6) Repeat step 1 to 4 until an IMF is obtained. Once the first IMF is obtained, define c1 (t) = h1 (t), which is the smallest temporal scale in x(t). A residual signal is obtained as r1 (t) = x(t) − c1 (t). The residue is treated as the next signal and the above mentioned process is repeated until the final residue is a constant (having no more IMFs). At the end of the decomposition, the original signal can be represented as follows: x(t) =
M X
cm (t) + rM (t)
m=1
where M is the number of IMFs, cm (t) is the mth IMF and rM (t) is the final residue. B. Analytic representation of IMFs After the extraction of IMFs from EEG signals, their analytic representation is obtained. This representation removes
4
the DC offset from the spectral component of the signals, which is an important aspect to compensate for the nonstationarity of the signals [35]. Given that we have an IMF cm (t), its analytic representation is given as:
feature extraction purposes. The discrimination power of the PSD features can be visually analysed by their respective plots for three IMFs from the normal and pathological EEG signals (Fig. 2). The PSD can be calculated as follows:
y(t) = cm (t) + iH{cm (t)} where H{cm (t)} is the Hilbert transform of cm (t), which is the mth IMF extracted from the signal x(t). After performing EMD of the signal, the IMFs are used for feature extraction purposes. C. Temporal Statistics of analytic IMFs Researchers have shown that the statistical features of IMFs are useful for discriminating between normal and pathological EEG signals [8]. Their use is motivated by the fact that the distribution of samples in the data are characterized by their asymmetry, dispersion and concentration around the mean. A visual analysis of the IMFs obtained from healthy and epilepsy patients during interictal and ictal periods after Hilbert transform (Fig. 2) reveals that they are quite different from one another. Interestingly, these differences are appropriately captured using the statistics of the IMFs. For an IMF, these statistics can be obtained by the following quantities: µt =
N 1 X yi N i=1
v u N u1 X t (yi − µt )2 σt = N i=1 βt =
3 N 1 X yi − µt N i=1 σt
Where N is the number of samples in the IMF µt is the mean, σt is the variance and βt is skewness of the corresponding IMF. D. Spectral statistics of analytic IMFs One important strength of EMD is that it has the ability to perform a spectral analysis of the signals. Its importance in the design of automated systems for EEG is based on the fact that the epileptic seizures give rise to changes in certain frequency bands [36], [37]. A frequency based analysis can therefore be useful for feature extraction from EEG signals. A conceptual interpretation of EMD is that it decomposes a signal into a number of components (IMFs) which are responses to filters having narrow pass bands. The spectral features obtained from IMFs can thus give a rich clue about the physiology of the EEG signals. Traditionally when using EMD, this spectral analysis is done using the calculation of instantaneous frequencies (IF). However, it is well known that the calculation of IF has a physical meaning only for monocomponent signals [38]. In practice, when the EEG signals are subjected to EMD, we do not get monocomponent signals (first three rows of Fig. 2). As an alternative, we have resorted to the calculation of power spectral density (PSD) for
P (w) =
∞ X
ry [n]e−jwn
(1)
−∞
where ry [n] represents the autocorrelation of y[n], defined as ry [n] = E(y[m]y ∗ [m]). Visual analysis of the PSD of IMFs shows that the statistics of the PSD can be used as relevant features for feature extraction (Fig. 2). 1) Spectral centroid: The researchers have shown that the centroid frequencies of the IMFs extracted from EEG signals form distinct groups when supervised clustering is applied on the EEG signals [7]. These respective groups are indicative of the seizure and non-seizure EEG signals. The centroid frequency is therefore a distinctive feature that can be used for the characterization of EEG signals. P wP (w) (2) Cs = Pw w P (w) where P (w) is the amplitude of wth frequency bin in the spectrum. 2) Variation coefficient: Since the spectral variation in the IMFs is different for normal and pathological EEG signals, therefore it can be used for their characterization. This variation can be calculated as follows: P (w − Cs )2 P (w) σs2 = w P (3) w P (w) where Cs is the spectral centroid. 3) Spectral skew: Skewness is the third order moment and it measures the symmetry/asymmetry of a distribution. Visual inspection of the plot of PSD of IMFs (Fig. 2) shows that the skewness of the power of IMFs for the normal and pathological EEG signals differs thus potentially yielding a useful feature for the classification of EEG signals. Skewness of the PSD can be calculated as: P w−Cs 3 w ( σs ) P (w) P βs = (4) w P (w) After the extraction of temporal and spectral features of each IMF, its feature vector can be obtained by their concatenation as follows: F =
µt
σt
βt
Cs
σs
βs
The feature vectors obtained from several IMFs can than be used for classification purposes. E. Classification Feature extraction is followed by the classification of EEG signals using support vector machines (SVM). The SVM, originally proposed by Vapnik et al. [39] mainly consists of constructing an optimum hyperplane that maximizes the margin of separation margin between two different classes.
5
It uses a kernel to transform the input data to a higher dimensional space followed by an optimization step for the construction of an optimum hyperplane. This approach builds the classification models having excellent generalization capability and thus is used in a very wide range of pattern recognition applications [40]. Given a training set X1...N containing N training labeled samples and coefficients α1...N learned in the training step, the decision function of SVM is as follows: X D(x) = αi K(Xi , F ) + b (5) i
where K(.) is the kernel function and F is the input vector. For our implementation, we have used liner kernel for SVM classification. We have used Weka [41] (a data mining tool developed by the University of Waikato, New Zealand) for our classification experiments employing 10-fold coss validation for assessing the classifier performance. IV. E XPERIMENTAL R ESULTS The performance of the proposed methodology for feature extraction from EEG signals is studied using standard measures such as overall accuracy and area under receiver operating characteristics (ROC) curve.
TABLE II: Different cases corresponding to different clinical objectives for classification purposes. Scenarios Case I
Case II Case III Case IV Case V
Grouping of Sets
Classifications
Sets A, B Sets C, D Set E Set A Set D Set E Set A Set E Sets A, B, C, D Set E Set D Set E
healthy interictal seizure healthy interictal seizure healthy seizure healthy seizure interictal seizure
wavelets [11]. This selection has been done based on a high degree of accuracy achieved by these methods in the classification of EEG signals. The features obtained using all
A. Classification objectives For the five sets of EEG recordings described in Section II, we have considered five different cases of classification problems in this paper. These cases are formulated based on their clinical relevance as well as their wide usage by various researchers for EEG signal classification [8], [10], [37], [42]. The Case I corresponds to classifying the EEG signals into three different categories. The EEG segments from Sets A and B are grouped together forming healthy class, Sets C and D are grouped as interictal class, and Set E signals belong to seizure class. In Case II, the signals in Sets A, D and E are classified as healthy, interictal and seizure classes respectively. In Case III, the signals from Sets A and E are classified as healthy and seizure classes respectively. In Case IV, Sets A, B, C and D are grouped together as the healthy class whereas the Set E is considered as seizure class. In Case V, Sets D and E are classified as interictal and seizure classes respectively. The Cases I and II are related to the discrimination between healthy and epilepsy patients and the detection of seizures. Cases III and IV are related to the detection of seizures and in addition, the discrimination of surface EEGs from the intracranial ones. Case V may be related to the detection of the onset of seizures in an automatic seizure detection system.
(a) Overall classification accuracy
B. Results In addition to the analysis of results produced by the methodology proposed in this paper, we have implemented several other descriptors of EEG signals. These methods include temporal statistics of IMFs from EMD [8], instantaneous frequency (IF) features [7], the frequency modulation (FM) and amplitude modulation (AM) bandwidth features [10] and
(b) Area under ROC curve
Fig. 3: Comparison of the classification results obtained using various features when SVMs are used for machine learning.
6
TABLE III: Overall classification performance obtained using different methods for the classification of EEG signals. The TP rates which are marked with ‘*’ indicate those predictions by the respective methods, which were significantly different from the predictions obtained using the proposed method since they give p-values of less than 0.05 in ANOVA tests. KNN Objectives
Case I
Case II
Case III
Case IV
Case V
Features EMD IF BW Wavelets Proposed EMD IF BW Wavelets Proposed EMD IF BW Wavelets Proposed EMD IF BW Wavelets Proposed EMD IF BW Wavelets Proposed
TP Rate 0.64* 0.66* 0.69* 0.70* 0.75 0.74* 0.73* 0.74* 0.74* 0.79 0.97 0.99 0.98 0.96* 0.99 0.92 0.93 0.92 0.92 0.94 0.89* 0.88* 0.87* 0.89* 0.93
ROC area 0.71 0.73 0.75 0.82 0.86 0.81 0.80 0.82 0.80 0.83 0.97 0.99 0.98 0.96 0.99 0.87 0.89 0.86 0.88 0.90 0.89 0.89 0.87 0.89 0.92
Decision Tree TP Rate 0.68* 0.66* 0.72* 0.78* 0.82 0.76* 0.69* 0.75* 0.78* 0.82 0.98 0.98 0.96 0.99 0.99 0.95 0.95 0.93 0.96 0.95 0.93* 0.94 0.86* 0.95 0.96
the descriptors are classified using four different classifiers i.e., 1-nearest neighbor (1NN), decision trees, artificial neural networks (ANN) and support vector machine (SVM) based classifiers for a rich analysis. The test bench is kept consistent for all these methods to ensure a fair comparison of their performance. The objective of assessing the discrimination power of the descriptors between healthy and epilepsy patients and additionally, the detection of seizures is achieved when the Cases I and II are considered in the classification experiments. These are the most challenging cases where an EEG signal has to be classified into three distinct classes. Our experiments show that the proposed feature set has outperformed the other methods which have been considered in this paper (Fig. 3) by a significant margin. The best accuracy has been achieved using the SVM based classifier. It is important to note that when 1NN classifier is used for classification, the proposed features show better performance as compared to all the other features indicating that the separability of the instances for different classification objectives is best maintained using the proposed feature set. A detailed analysis of the results is performed using the ANOVA (analysis of variance) test between the labels predicted by the proposed method and all the other methods. Our experiments show that the classification results obtained using the proposed method are significantly
ROC area 0.77 0.74 0.81 0.85 0.88 0.85 0.83 0.85 0.85 0.88 0.97 0.99 0.97 0.99 0.99 0.93 0.89 0.93 0.92 0.94 0.91 0.94 0.87 0.94 0.95
ANN TP Rate 0.72* 0.73* 0.77* 0.68* 0.82 0.79* 0.75* 0.78* 0.72* 0.81 0.98 0.97 0.99 0.98 0.99 0.93 0.90* 0.94 0.93 0.95 0.88* 0.90* 0.89* 0.89* 0.94
ROC area 0.85 0.86 0.88 0.86 0.92 0.89 0.80 0.89 0.88 0.93 0.99 0.97 1 0.99 1 0.92 0.90 0.95 0.97 0.98 0.90 0.92 0.90 0.91 0.98
SVM TP Rate 0.63* 0.62* 0.71* 0.58* 0.83 0.71* 0.70* 0.72* 0.57* 0.85 0.98 0.99 0.98 0.89* 0.99 0.95 0.95 0.94 0.93* 0.96 0.92 0.92 0.92 0.83* 0.93
ROC area 0.72 0.72 0.78 0.67 0.94 0.81 0.81 0.83 0.71 0.91 0.98 0.99 0.98 0.89 0.99 0.9 0.91 0.89 0.83 0.98 0.92 0.92 0.94 0.83 0.96
different as compared to the other methods, elucidating on the superiority of the proposed method (state-of-the-art methods showing statistically significant results are marked with asterisk in Table III). The superiority of the PSD features over IF can be established due to the fact that the IF based features make an underlying assumption that the IMFs are monocomponent (which is not true in our case). The wavelets are based on fixed bandpass filters that cannot cater for the non-stationarity in EEG signals. The higher order statistics features are based only on the temporal characteristics of the signals and they are complemented with the PSD based features is our implementation. These are the main reasons for getting the performance enhancement for the classification of EEG signals in the proposed method as compared to the other methods that have been considered in this paper. In order to discriminate between surface and intracranial EEGs and additionally, the detection of seizures, the Cases III and IV are considered for our classification experiments. The classification results show that the proposed feature set exhibits similar performance as compared to the other methods considered in this paper. For Case III, the proposed method shows the same performance as that of IF using all the classifiers except when ANN are used for classification (the proposed method is about 2% better when using ANN). The same performance as that of the proposed method is repeated by the BW and
7
wavelet based features using different classifiers. However, it is important to note that the proposed method performs better or similar as compared to all the other methods using any machine learning method. For the classification experiments related to the Case IV, the best performance is shown by the proposed method in combination with an SVM classifier. Similar overall accuracy was observed when wavelets were used in combination with the decision trees for classification. The results produced by the rest of the features are also good (>90%). ANOVA statistical significance test shows that the results produced by the proposed method are not significantly different as compared to the other methods. To detect the onset of seizures in an automatic seizure detection system, the Case V can be considered for the classification experiments. The best performance is obtained using the proposed method in combination with the decision trees giving about 96% classification accuracy. Among the other methods, wavelets in combination with decision trees give about 95% classification accuracy. It should be noted that when 1NN classifier is used, the results obtained for Case V using the proposed method are better with higher statistical significance, elucidating on the better discrimination power of the proposed feature extraction method in the feature space. From the results, it is clear that most state-of-the-art methods show good performance for the Cases III, IV and V but for the Cases I and II, they are deficient. The proposed method achieves significantly higher classification accuracies and respective areas under the ROC curves for the Cases I and II while maintaining better discrimination capability for the rest of the Cases. V. C ONCLUSION In this paper, we have proposed a method for the detection of seizures and epilepsy in the EEG signals. The foundation of this method lies on the extraction of temporal and spectral features from Empirical Mode Decomposition (EMD) of the EEG signals. The usage of EMD is motivated by the fact that EEG signals are non-stationary and EMD is a data dependent method exhibiting a better adaptability towards non-stationarity in the EEG signals. Previously, researchers have shown that temporal statistics are useful for classifying the EEG signals [8]. We have extended their work in the spectral domain where the power spectrum density (PSD) has been observed to exhibit good discrimination power for the classification of EEG signals (Fig. 2). Since the temporal and spectral characteristics of the signals represent complementary information, we have proposed to concatenate these features in order to extract the salient characteristics of EEG signals. The proposed feature set is embedded in a pattern recognition framework where we have used SVMs for classification purposes (results using various classifiers is also presented for comparison). We have used a publicly available EEG dataset for our experiments. Several classification objectives can be effectively studied using this dataset. We have considered five different cases, referring to different objectives from a clinical perspective. Our experiments have shown that in two cases which
indicate a three class pattern recognition problem (Case I and II), we get significant performance improvements after the use of hybrid features for the extraction of relevant information from the EEG signals. Under our experimental setup, the proposed method increases the classification performance by more than 10% for the Cases I and II (when using an SVM classifier) as compared to the respective benchmark methods for these cases. For the rest of the three Cases our methodology shows similar performance to the other methods that have been considered in this paper. In the classification experiments related to Case V, the proposed features in comnbination with 1NN classifier shows significantly different results as compared to the other feature extraction methods. In the future, we intend to extend this work by researching on the possibility to extract more discriminative features to further improve the classification results for the Cases I and II. Furthermore, we intend to explore the possibility to do the hardware implementation of the proposed features for designing an automatic seizure detection system for effective usage by the patients. R EFERENCES [1] V. Srinivasan, C. Eswaran, Sriraam, and N, “Artificial neural network based epileptic detection using time-domain and frequency-domain features,” Journal of Medical Systems, vol. 29, no. 6, pp. 647–660, 2005. [2] K. Polat and S. G¨unes¸, “Classification of epileptiform eeg using a hybrid system based on decision tree classifier and fast fourier transform,” Applied Mathematics and Computation, vol. 187, no. 2, pp. 1017–1026, 2007. [3] A. Tzallas, M. Tsipouras, and D. Fotiadis, “Epileptic seizure detection in eegs using time-frequency analysis,” Information Technology in Biomedicine, IEEE Transactions on, vol. 13, no. 5, pp. 703–710, 2009. [4] N. Robinson, A. Vinod, K. K. Ang, K. P. Tee, and C. Guan, “Eegbased classification of fast and slow hand movements using waveletcsp algorithm,” Biomedical Engineering, IEEE Transactions on, vol. 60, no. 8, pp. 2123–2132, 2013. [5] N. Mammone, F. La Foresta, and F. C. Morabito, “Automatic artifact rejection from multichannel scalp eeg by wavelet ica,” Sensors Journal, IEEE, vol. 12, no. 3, pp. 533–542, 2012. [6] R. B. Pachori and P. Sircar, “Eeg signal analysis using fb expansion and second-order linear tvar process,” Signal Processing, vol. 88, no. 2, pp. 415–420, 2008. [7] R. J. Oweis and E. W. Abdulhay, “Seizure classification in eeg signals utilizing hilbert-huang transform,” Biomedical engineering online, vol. 10, no. 1, p. 38, 2011. [8] S. Alam and M. Bhuiyan, “Detection of seizure and epilepsy using higher-order statistics in the emd domain,” Journal of Biomedical and Health Informatics, IEEE, vol. 17, no. 2, 2013. [9] P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion recognition from eeg using higher order crossings,” Information Technology in Biomedicine, IEEE Transactions on, vol. 14, no. 2, pp. 186–197, 2010. [10] V. Bajaj and R. Pachori, “Classification of seizure and non-seizure eeg signals using empirical mode decomposition,” Transactions on Information Technology in Biomedicine, IEEE, vol. 16, no. 6, pp. 1135– 1142, 2012. [11] A. Subasi, “Application of adaptive neuro-fuzzy inference system for epileptic seizure detection using wavelet feature extraction,” Computers in Biology and Medicine, vol. 37, no. 2, pp. 227–244, 2007. [12] A. Subasi, “Eeg signal classification using wavelet feature extraction and a mixture of expert model,” Expert Systems with Applications, vol. 32, no. 4, pp. 1084–1093, 2007. [13] Y. Li, J. Long, T. Yu, Z. Yu, C. Wang, H. Zhang, and C. Guan, “An eegbased bci system for 2-d cursor control by combining mu/beta rhythm and p300 potential,” Biomedical Engineering, IEEE Transactions on, vol. 57, no. 10, pp. 2495–2505, 2010. [14] R. Tetzlaff and V. Senger, “The seizure prediction problem in epilepsy: Cellular nonlinear networks,” Circuits and Systems Magazine, IEEE, vol. 12, no. 4, pp. 8–20, 2012.
8
[15] F. Semah, M.-C. Picot, C. Adam, D. Broglin, A. Arzimanoglou, B. Bazin, D. Cavalcanti, and M. Baulac, “Is the underlying cause of epilepsy a major prognostic factor for recurrence?,” Neurology, vol. 51, no. 5, pp. 1256–1262, 1998. [16] H. R. Mohseni, A. Maghsoudi, and M. B. Shamsollahi, “Seizure detection in eeg signals: A comparison of different approaches,” in Conf Proc IEEE Eng Med Biol Soc, pp. 6724–7, 2006. [17] H. Mohseni, A. Maghsoudi, and M. Shamsollahi, “Seizure detection in eeg signals: A comparison of different approaches,” in Engineering in Medicine and Biology Society, 2006. EMBS ’06. 28th Annual International Conference of the IEEE, vol. Supplement, pp. 6724–6727, 2006. [18] A. V. Oppenheim, R. W. Schafer, J. R. Buck, et al., Discrete-time signal processing, vol. 5. Prentice Hall Upper Saddle River, 1999. [19] A. Hyv¨arinen, P. Ramkumar, L. Parkkonen, and R. Hari, “Independent component analysis of short-time fourier transforms for spontaneous eeg/meg analysis,” Neuroimage, vol. 49, no. 1, pp. 257–271, 2010. [20] M. S. Crouse, R. D. Nowak, and R. G. Baraniuk, “Wavelet-based statistical signal processing using hidden markov models,” Signal Processing, IEEE Transactions on, vol. 46, no. 4, pp. 886–902, 1998. [21] S. G. Mallat, “A theory for multiresolution signal decomposition: the wavelet representation,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 11, no. 7, pp. 674–693, 1989. [22] S. Mallat, A wavelet tour of signal processing. Access Online via Elsevier, 1999. [23] S. Mallat, A wavelet tour of signal processing: the sparse way. Access Online via Elsevier, 2008. [24] C. Li, C. Zheng, and C. Tai, “Detection of ecg characteristic points using wavelet transforms,” Biomedical Engineering, IEEE Transactions on, vol. 42, no. 1, pp. 21–28, 1995. [25] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of eeg records in an epileptic patient using wavelet transform,” Journal of neuroscience methods, vol. 123, no. 1, pp. 69–87, 2003. [26] G. Tzanetakis and P. Cook, “Musical genre classification of audio signals,” Speech and Audio Processing, IEEE transactions on, vol. 10, no. 5, pp. 293–302, 2002. [27] D. Sinha and A. H. Tewfik, “Low bit rate transparent audio compression using adapted wavelets,” Signal Processing, IEEE Transactions on, vol. 41, no. 12, pp. 3463–3479, 1993. [28] R. Sarikaya and J. H. Hansen, “High resolution speech feature parametrization for monophone-based stressed speech recognition,” Signal Processing Letters, IEEE, vol. 7, no. 7, pp. 182–185, 2000. [29] B. Boashash, Time frequency analysis. Access Online via Elsevier, 2003. [30] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of eeg records in an epileptic patient using wavelet transform,” Journal of neuroscience methods, vol. 123, no. 1, pp. 69–87, 2003. [31] P. Flandrin, G. Rilling, and P. Goncalves, “Empirical mode decomposition as a filter bank,” Signal Processing Letters, IEEE, vol. 11, no. 2, pp. 112–114, 2004. [32] R. B. Pachori, “Discrimination between ictal and seizure-free eeg signals using empirical mode decomposition,” Research Letters in Signal Processing, vol. 2008, p. 14, 2008. [33] “Eeg time series download page @ONLINE,” 2012. [34] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.C. Yen, C. C. Tung, and H. H. Liu, “The empirical mode decomposition and the hilbert spectrum for nonlinear and non-stationary time series analysis,” Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, vol. 454, no. 1971, pp. 903–995, 1998. [35] T. I. Netoff and S. J. Schiff, “Decreased neuronal synchronization during experimental seizures,” The Journal of neuroscience, vol. 22, no. 16, pp. 7297–7307, 2002. [36] A. Tzallas, M. Tsipouras, and D. Fotiadis, “Automatic seizure detection based on time-frequency analysis and artificial neural networks,” Computational Intelligence and Neuroscience, vol. 2007, p. 18, 2007. [37] H. Adeli, S. Ghosh-Dastidar, and N. Dadmehr, “A wavelet-chaos methodology for analysis of eegs and eeg subbands to detect seizure and epilepsy,” Biomedical Engineering, IEEE Transactions on, vol. 54, no. 2, pp. 205–211, 2007. [38] B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal. i. fundamentals,” Proceedings of the IEEE, vol. 80, pp. 520– 538, Apr 1992. [39] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik, “Gene selection for cancer classification using support vector machines,” Machine learning, vol. 46, no. 1-3, pp. 389–422, 2002. [40] H. Byun and S.-W. Lee, “A survey on pattern recognition applications of support vector machines,” International Journal of Pattern Recognition and Artificial Intelligence, vol. 17, no. 03, pp. 459–486, 2003.
[41] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, and I. H. Witten, “The weka data mining software: an update,” ACM SIGKDD Explorations Newsletter, vol. 11, no. 1, pp. 10–18, 2009. [42] I. Guler and E. D. Ubeyli, “Multiclass support vector machines for eegsignals classification,” Information Technology in Biomedicine, IEEE Transactions on, vol. 11, no. 2, pp. 117–126, 2007.
