312
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013
Detection of Seizure and Epilepsy Using Higher Order Statistics in the EMD Domain S. M. Shafiul Alam and M. I. H. Bhuiyan, Member, IEEE
Abstract—In this paper, a method using higher order statistical moments of EEG signals calculated in the empirical mode decomposition (EMD) domain is proposed for detecting seizure and epilepsy. The appropriateness of these moments in distinguishing the EEG signals is investigated through an extensive analysis in the EMD domain. An artificial neural network is employed as the classifier of the EEG signals wherein these moments are used as features. The performance of the proposed method is studied using a publicly available benchmark database for various classification cases that include healthy, interictal (seizure-free interval) and ictal (seizure), healthy and seizure, nonseizure and seizure, and interictal and ictal, and compared with that of several recent methods based on time–frequency analysis and statistical moments. It is shown that the proposed method can provide, in almost all the cases, 100% accuracy, sensitivity, and specificity, especially in the case of discriminating seizure activities from the nonseizure ones for patients with epilepsy while being much faster as compared to the time–frequency analysis-based techniques. Index Terms—Electroencephalogram (EEG), empirical mode decomposition (EMD), epileptic seizure, neural network.
I. INTRODUCTION EG contains a set of electric potential differences developed as a result of volume currents spreading from an active neural tissue throughout the conductive media of the brain. These measurements can be obtained either using sensors on the scalp or by placing special intracranial electrodes. Since its first recording by German psychiatrist Berger [1], EEG has been considered a successful tool in neuroscience to diagnose diseases and disorders. Epilepsy is one of the most serious neurological disorders. About 50 million people world-wide are suffering from epilepsy and 85% of those live in developing countries. Each year, 2.4 million new cases are estimated to occur globally [2]. In most of the adult patients, it occurs in the mesial temporal structures such as hippocampus, amygdala, and parahippocampal gyrus [3]. It is characterized by recurrent seizures, i.e., transient impairments of sensation, thinking, and motor control, caused by sudden, usually brief, excessive electrical discharges in a group of brain cells. The EEG records can easily display these
E
Manuscript received October 10, 2011; revised July 31, 2012 and November 10, 2012; accepted December 21, 2012. Date of publication January 3, 2013; date of current version March 8, 2013. The authors are with the Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology, Dhaka 1000, Bangladesh (e-mail: [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2012.2237409
electrical discharges as a rapid change in potential differences. Thus, neurologists invariably use EEG records to investigate suspected seizure phenomena [4]–[12]. Detection of a seizure attack is traditionally carried out by viewing EEG records of long duration that may even last for several days. Considering the length of the records to be observed and the huge volume of patients, it is often difficult and time consuming for an expert neurologist to locate a seizure. An automatic seizure detection system can considerably reduce the volume of data to be observed. The system needs to be highly sensitive, even though that may generate lots of false alarms as the neurologist can easily discard them [12]. The purpose of the system is not to replace the neurologist, rather to relieve him off the burden of time-consuming observation by providing alarms and in addition, reduce the effect of misinterpretation, as the manual seizure detection is highly subjective. If he suspects, he can review the alarm and ignore if it is a false one. Also, such a system can be integrated into an implantable device for detecting the onset of seizures and henceforth, trigger an alarm and initiate treatment through neurostimulation or drug delivery. This is important since often the patient is not in a state to push the alarm and focal drug delivery or neurostimulation might be more effective with reduced side effects related to antiepileptic drugs. Various algorithms have been introduced for detecting seizure from EEG records [4]–[17]. A multidimensional probability evolution-based technique is introduced by McSharry et al., which gives fewer false positives as compared to using variances [4]. However, the difference is quite small and, thus, does not negate the significance of linear statistics such as variance in identifying seizures. In [5], seizures are detected by thresholding variances calculated from local windows using an arbitrary threshold for classifying normal and seizure activities. In [6], chaotic features that include largest Lyapunov exponent (LLE) and correlation dimension (CD) obtained from the wavelet subbands of the EEG signals are shown to be effective in differentiating the signals of various classes including those of seizures. Concurrently, Lyapunov spectra have been used in [7] to make a multiway classification using multiclass support vector machines (SVMs). The approximate entropy in conjunction with autoregressive model parameters extracted from the Fourier transforms of the EEG signals is employed in linear and nonlinear classifiers by Liang et al. [8]. A two-way classification using fractal dimension and artificial neural networks (ANN) is carried out in [9]. In [10] and [11], linear statistical measures obtained from the EEG signals are used as features, later employed in a linear classifier to distinguish normal and seizure activities. Various time–frequency analysis (TFA) techniques such
2168-2194/$31.00 © 2013 IEEE
ALAM AND BHUIYAN: DETECTION OF SEIZURE AND EPILEPSY USING HIGHER ORDER STATISTICS IN THE EMD DOMAIN
313
Fig. 1. Sample EEG signals from five datasets (first row) and corresponding first four IMFs. The second, third, fourth, and fifth rows correspond to the first, second, third, and fourth IMFs, respectively.
as smoothed pseudo-Wigner–Ville and reduced interference are used in conjunction with an ANN in [12] and [13] and a high accuracy in detection is reported. Recently, the empirical mode decomposition (EMD) has drawn the attention of researchers in nonlinear signal analysis for being intuitive and adaptive to signals, while requiring no assumption in regard to stationarity and linearity [14]. Since the EEG signals exhibit nonstationary behavior, a number of methods have been developed to detect seizures in the EMD domain. The mean frequencies of the intrinsic mode function (IMF) obtained by the EMD of an EEG signal are shown to be effective in discriminating the ictal periods from the nonictal ones, that is, seizure activities from the nonseizure ones. The energy of an IMF and minimum distance duration are used in [15] for seizure detection. In [16], features are extracted from the IMFs using Mann–Whitney Test and Lambda of Wilks criterion and used in a linear discriminant analyzer for classifying the EEG signals. In [17], chaotic features such as the LLE and CD computed from the various IMFs are shown to be effective in distinguishing the EEG signals of various classes. Notice that in [14] and [17], the features are shown to have the ability to discriminate and not further used for classification. In this paper, higher order moments that include variance, kurtosis, and skewness, extracted from the IMFs of the EEG signals are used as features in an ANN to classify the EEG signals. The appropriateness of these features is studied using a large database. The performance of the proposed method for classifying EEG signals is evaluated for original EEG signals as well as the first four IMFs, and compared with those of stateof-the-art methods.
the data itself [18]. For an N -point data, X{x1 , x2 , . . ., xN }, the decomposition is carried out as follows. 1) Set input as h = X and hold = h. 2) The local maxima and minima of hold are identified. 3) Envelops of local maxima em ax and that of local minima em in are obtained using cubic spline interpolation. 4) The values of the mean of em ax and em in are calculated as m = (em ax + em in )/2 and subsequently subtracted from hold as hnew = hold − m. 5) Set hold = hnew . Go to step 2. Steps 2–5 are repeated until |hnew − hold |2 2
IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013
Detection of Seizure and Epilepsy Using Higher Order Statistics in the EMD Domain S. M. Shafiul Alam and M. I. H. Bhuiyan, Member, IEEE
Abstract—In this paper, a method using higher order statistical moments of EEG signals calculated in the empirical mode decomposition (EMD) domain is proposed for detecting seizure and epilepsy. The appropriateness of these moments in distinguishing the EEG signals is investigated through an extensive analysis in the EMD domain. An artificial neural network is employed as the classifier of the EEG signals wherein these moments are used as features. The performance of the proposed method is studied using a publicly available benchmark database for various classification cases that include healthy, interictal (seizure-free interval) and ictal (seizure), healthy and seizure, nonseizure and seizure, and interictal and ictal, and compared with that of several recent methods based on time–frequency analysis and statistical moments. It is shown that the proposed method can provide, in almost all the cases, 100% accuracy, sensitivity, and specificity, especially in the case of discriminating seizure activities from the nonseizure ones for patients with epilepsy while being much faster as compared to the time–frequency analysis-based techniques. Index Terms—Electroencephalogram (EEG), empirical mode decomposition (EMD), epileptic seizure, neural network.
I. INTRODUCTION EG contains a set of electric potential differences developed as a result of volume currents spreading from an active neural tissue throughout the conductive media of the brain. These measurements can be obtained either using sensors on the scalp or by placing special intracranial electrodes. Since its first recording by German psychiatrist Berger [1], EEG has been considered a successful tool in neuroscience to diagnose diseases and disorders. Epilepsy is one of the most serious neurological disorders. About 50 million people world-wide are suffering from epilepsy and 85% of those live in developing countries. Each year, 2.4 million new cases are estimated to occur globally [2]. In most of the adult patients, it occurs in the mesial temporal structures such as hippocampus, amygdala, and parahippocampal gyrus [3]. It is characterized by recurrent seizures, i.e., transient impairments of sensation, thinking, and motor control, caused by sudden, usually brief, excessive electrical discharges in a group of brain cells. The EEG records can easily display these
E
Manuscript received October 10, 2011; revised July 31, 2012 and November 10, 2012; accepted December 21, 2012. Date of publication January 3, 2013; date of current version March 8, 2013. The authors are with the Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology, Dhaka 1000, Bangladesh (e-mail: [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2012.2237409
electrical discharges as a rapid change in potential differences. Thus, neurologists invariably use EEG records to investigate suspected seizure phenomena [4]–[12]. Detection of a seizure attack is traditionally carried out by viewing EEG records of long duration that may even last for several days. Considering the length of the records to be observed and the huge volume of patients, it is often difficult and time consuming for an expert neurologist to locate a seizure. An automatic seizure detection system can considerably reduce the volume of data to be observed. The system needs to be highly sensitive, even though that may generate lots of false alarms as the neurologist can easily discard them [12]. The purpose of the system is not to replace the neurologist, rather to relieve him off the burden of time-consuming observation by providing alarms and in addition, reduce the effect of misinterpretation, as the manual seizure detection is highly subjective. If he suspects, he can review the alarm and ignore if it is a false one. Also, such a system can be integrated into an implantable device for detecting the onset of seizures and henceforth, trigger an alarm and initiate treatment through neurostimulation or drug delivery. This is important since often the patient is not in a state to push the alarm and focal drug delivery or neurostimulation might be more effective with reduced side effects related to antiepileptic drugs. Various algorithms have been introduced for detecting seizure from EEG records [4]–[17]. A multidimensional probability evolution-based technique is introduced by McSharry et al., which gives fewer false positives as compared to using variances [4]. However, the difference is quite small and, thus, does not negate the significance of linear statistics such as variance in identifying seizures. In [5], seizures are detected by thresholding variances calculated from local windows using an arbitrary threshold for classifying normal and seizure activities. In [6], chaotic features that include largest Lyapunov exponent (LLE) and correlation dimension (CD) obtained from the wavelet subbands of the EEG signals are shown to be effective in differentiating the signals of various classes including those of seizures. Concurrently, Lyapunov spectra have been used in [7] to make a multiway classification using multiclass support vector machines (SVMs). The approximate entropy in conjunction with autoregressive model parameters extracted from the Fourier transforms of the EEG signals is employed in linear and nonlinear classifiers by Liang et al. [8]. A two-way classification using fractal dimension and artificial neural networks (ANN) is carried out in [9]. In [10] and [11], linear statistical measures obtained from the EEG signals are used as features, later employed in a linear classifier to distinguish normal and seizure activities. Various time–frequency analysis (TFA) techniques such
2168-2194/$31.00 © 2013 IEEE
ALAM AND BHUIYAN: DETECTION OF SEIZURE AND EPILEPSY USING HIGHER ORDER STATISTICS IN THE EMD DOMAIN
313
Fig. 1. Sample EEG signals from five datasets (first row) and corresponding first four IMFs. The second, third, fourth, and fifth rows correspond to the first, second, third, and fourth IMFs, respectively.
as smoothed pseudo-Wigner–Ville and reduced interference are used in conjunction with an ANN in [12] and [13] and a high accuracy in detection is reported. Recently, the empirical mode decomposition (EMD) has drawn the attention of researchers in nonlinear signal analysis for being intuitive and adaptive to signals, while requiring no assumption in regard to stationarity and linearity [14]. Since the EEG signals exhibit nonstationary behavior, a number of methods have been developed to detect seizures in the EMD domain. The mean frequencies of the intrinsic mode function (IMF) obtained by the EMD of an EEG signal are shown to be effective in discriminating the ictal periods from the nonictal ones, that is, seizure activities from the nonseizure ones. The energy of an IMF and minimum distance duration are used in [15] for seizure detection. In [16], features are extracted from the IMFs using Mann–Whitney Test and Lambda of Wilks criterion and used in a linear discriminant analyzer for classifying the EEG signals. In [17], chaotic features such as the LLE and CD computed from the various IMFs are shown to be effective in distinguishing the EEG signals of various classes. Notice that in [14] and [17], the features are shown to have the ability to discriminate and not further used for classification. In this paper, higher order moments that include variance, kurtosis, and skewness, extracted from the IMFs of the EEG signals are used as features in an ANN to classify the EEG signals. The appropriateness of these features is studied using a large database. The performance of the proposed method for classifying EEG signals is evaluated for original EEG signals as well as the first four IMFs, and compared with those of stateof-the-art methods.
the data itself [18]. For an N -point data, X{x1 , x2 , . . ., xN }, the decomposition is carried out as follows. 1) Set input as h = X and hold = h. 2) The local maxima and minima of hold are identified. 3) Envelops of local maxima em ax and that of local minima em in are obtained using cubic spline interpolation. 4) The values of the mean of em ax and em in are calculated as m = (em ax + em in )/2 and subsequently subtracted from hold as hnew = hold − m. 5) Set hold = hnew . Go to step 2. Steps 2–5 are repeated until |hnew − hold |2 2
