Elsevier

Knowledge-Based Systems

Volume 118, 15 February 2017, Pages 217-227
Knowledge-Based Systems

An automatic detection of focal EEG signals using new class of time–frequency localized orthogonal wavelet filter banks

https://doi.org/10.1016/j.knosys.2016.11.024Get rights and content

Highlights

  • Proposed an automatic detection of focal and non-focal EEG signals.

  • Design of a new class of time-frequency localized wavelet filter banks.

  • Obtained the highest classification accuracy of 94.25%.

  • Computationally faster than all existing methods.

Abstract

It is difficult to detect subtle and vital differences in electroencephalogram (EEG) signals simply by visual inspection. Further, the non-stationary nature of EEG signals makes the task more difficult. Determination of epileptic focus is essential for the treatment of pharmacoresistant focal epilepsy. This requires accurate separation of focal and non-focal groups of EEG signals. Hence, an intelligent system that can detect and discriminate focal–class (FC) and non–focal–class (NFC) of EEG signals automatically can aid the clinicians in their diagnosis. In order to facilitate accurate analysis of non-stationary signals, joint time–frequency localized bases are highly desirable. The performance of wavelet bases is found to be effective in analyzing transient and abrupt behavior of EEG signals. Hence, we employ a novel class of orthogonal wavelet filter banks which are localized in time–frequency domain to detect FC and NFC EEG signals automatically. We classify EEG signals as FC and NFC using the proposed wavelet based system. We compute various entropies from the wavelet coefficients of the signals. These entropies are used as discriminating features for the classification of FC and NFC of EEG signals. The features are ranked using Student’s t-test ranking algorithm and then fed to Least Squares-Support Vector Machine (LS–SVM) to classify the signals. Our proposed method achieved the highest classification accuracy of 94.25%. We have obtained 91.95% sensitivity and 96.56% specificity, respectively, using this method.

The classification of FC and NFC of EEG signals helps in localization of the affected brain area which needs to undergo surgery.

Introduction

Epilepsy is a chronic disorder identified by repeated seizures. Epilepsy is the most common neurological disorder and worldwide more than 60 million patients are suffering, of which 80% live in developing countries [1]. The epilepsy can be categorized mainly in two types, namely, generalized and partial (or focal) epilepsy. In generalized epilepsy, a continuous region contributes to epilepsy whereas focal epilepsy is primarily due to discrete points which act as the epileptic foci [2]. There are two widely used techniques to observe brain activity, electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI). The fMRI measures electrical activity of brain indirectly by recording changes in blood flow [3]. On the other hand, EEG signals directly measure the brain’s electrical activity as voltages. These signals can be acquired by placing electrodes on the skull (extracranial recording) or inside the skull (intracranial recording) [4]. EEG signals can be used to detect a number of neurological disorders such as sleep disorder [5], epilepsy [6], autism[7], attention deficit hyperactivity disorder (ADHD), depression [8] etc. The fMRI has a better spatial resolution than that of EEG. However, its time resolution is inferior than that of EEG. Further, diagnosis through EEG is much less expensive than fMRI and this can be a significant factor in treatment in third world countries [9]. The EEG signals can be used to detect epilepsy and epileptogenic areas which contribute to the epileptic seizures. In generalized epilepsy, the epileptic seizure EEG signals have impulsive nature and higher energy than that of seizure-free EEG signals. The seizure EEG signals are generated because of synchronous activities of neurons in the brain [10]. In case of partial epilepsy, the recording contains two classes of EEG signals: focal–class (FC) of EEG signals and non–focal–class (NFC) of EEG signals. Signals belonging to the FC are referred to as focal EEG signals whereas signals belonging to the NFC are referred to as non–focal EEG signals. It may be noted that focal EEG signals are different than epileptic seizure EEG signals. Focal EEG signals are recorded from the areas of brain where first ictal EEG signals are detected, and non–focal EEG signals are recorded from brain areas that are not involved at seizure onset [11]. The epileptic seizure EEG signals are recorded at the time when the patient suffers from an epileptic seizure. The FC of EEG signals aid in the detection of focal epilepsy. The automatic classification of EEG signals such as normal versus epileptic and seizure versus inter-ictal has been studied by many researchers [6]. The most challenging classification problem is FC versus NFC. Almost 100% accuracy has been obtained for the classification of seizure from normal and seizure-free EEG signals [6]. However, 100% accuracy has not yet been achieved for classification of focal and non–focal EEG signals.

The underlying physiology behind epileptic seizures is that the hyper synchronous activity of neurons causes improper changes in the sensory and motor activities. Hence, an unexpected and vast surge of energy takes place instead of manageable energy discharge of cells in the brain. This is caused by malfunctioning of the brain’s electrophysiological system [6]. The visual inspection of EEG signals can allow experienced neurophysiologists to identify epileptogenic areas. However, this involves subjectivity, consumes more time and is prone to errors. It is also observed that 60% of focal epilepsy patients become pharmacoresistant [12]. In such cases, surgical intervention is needed to remove epileptogenic areas. Hence an accurate determination of these areas is very important. Automatic detection of EEG signals associated with FC allows determination of epileptogenic focus by signal processing methods [13]. For determination of the epileptic focus, a process of brain mapping is performed where an electrode strip is surgically inserted in the subdural region. Following this, the EEG signals originating from each electrode are observed. These EEG signals would then be classified as focal or non–focal with the proposed methodology. Since the location of each electrode is known, if the EEG signals from electrode placed at a particular point are classified as focal, then that point is one of the epileptic foci. An automated classification system uses recorded EEG signals as an input, from which discriminating features are obtained and employed to train a classifier. Hence, the features employed for the discrimination and the classification techniques to be used are primary considerations in such systems [6], [14].

EEG signals are associated with the class of non-stationary signals [15]. The Fourier transform is not suitable for the analysis of EEG signals as it fails to capture time evolution of frequency components. In time–frequency analysis of signals, we attempt to localize the frequency components in time. For non-stationary signals and sharp transitions, the spectral contents vary with time. For the analysis of these signals, the Fourier transform cannot be considered a promising tool. To overcome the limitations of Fourier transform, Gabor proposed short time Fourier transform (STFT) where time localization is obtained by taking Fourier transform of the windowed signal. This gives rise to time–frequency representation of the signal. In STFT, due to the fixed size of the window, the frequency and time localizations are fixed. Further, the analysis of the signal depends on the type of the window chosen. These limitations of STFT can be overcome by using wavelet transform (WT). The WTs have been proven to be a powerful aid for the analysis of non-stationary and transient signals. Unlike the STFT, the WT allows to improve localization in one domain at the cost of localization in the other domain. The time localization can be traded with the frequency localization in case of WT. Thus, the WT offers good frequency resolution at low frequencies and good time resolution at high frequencies [16], [17]. Hence, WT is an effective tool for the analysis of EEG signals. The WT has been proven to be an excellent tool for the detection and classification of various biomedical signals including EEG [17], [18].

A typical two–channel perfect reconstruction filter bank (PRFB) is shown in Fig. 1. A filter bank (FB) is said to be perfect reconstruction if the output X^ is an exact replica of the input. Here, filters G1(z) and G0(z) represent analysis highpass and lowpass filters, respectively and F1(z) and F0(z) denote synthesis highpass and lowpass filters, respectively.

Wavelet filter banks (WFBs) and PRFBs are closely connected [19]. Iterations of PRFBs yield smooth wavelets only if they satisfy a necessary condition termed as regularity or vanishing moment (VM) condition. Thus, the design of WFBs can be regarded equivalent to design of PRFBs provided the latter satisfy regularity condition. It is to be noted that PRFBs cannot yield regular L2(R) stable wavelets if they do not satisfy regularity condition. Regularity condition is achieved by imposing VMs on lowpass filters of the underlying PRFB. In wavelet literature, a lowpass filter (LPF) is said to be regular if its transfer function G(z), which is a polynomial in the variable z, contains the factor (z+1)m, m ≥ 1. Equivalently, a lowpass filter is said to be m–regular if it has m zeroes at z=1 [20].

Daubechies filters are obtained from the factorization of the Lagrange half-band polynomial (LHBP) which has maximum number of zeros at z=1 (VMs) [16]. Smooth orthogonal wavelets given by Daubechies [16] can be generated from iterations of these regular filters [20]. The LHBP has maximum number of VMs, however it does not provide any freedom to optimize filter coefficients of the underlying orthogonal FB. By reducing the number of zeros at z=1, some degrees of freedom can be obtained. The freedoms can be used to tailor attributes of the filter like time–frequency localization, energy compaction and frequency selectivity [21], [22]. There are several criteria to design WFBs such as orthogonality, regularity or VMs, frequency selectivity and time–frequency localization [23]. Daubechies wavelets are designed so as to have maximum VMs. However, time–frequency localization is not considered in the design. In this work, we employ a new class of time–frequency localized orthogonal WFBs. The EEG signals are decomposed into various subbands using the new class of WFBs. The disriminating features are extracted from the wavelet subbands of EEG signals.

In the literature, different EEG signal processing techniques such as Fourier transform, WTs and empirical mode decomposition (EMD) have been used for epileptic seizure detection [6], [24], [25], [26].

For classification of EEG signals as ictal, inter-ictal or normal, Singh et al. [27] extracted features from EEG rhythms using Fourier transform. Guo et al. [28] employed Daubechies wavelets along with artificial neural networks and line length feature for seizure detection. Acharya et al. [29] used wavelet features for identification of EEG signals associated with seizure. However, none of them explored the classification of FC and NFC of EEG signals. Sharma et al. [13] employed Daubechies discrete wavelet transform (DWT) and used wavelet entropy features for identification of FC and NFC of EEG signals [25], [26]. The EMD has also attracted attention of researchers in the analysis of non-stationary signals [30], [31], [32], [33]. Sharma et al. [34] applied EMD with average variance of instantaneous frequencies (AVIF) and average sample entropy (ASE) as features to classify FC and NFC of EEG signals. The authors employed intrinsic mode function (IMF) with different entropy features for the same purpose [35]. The EMD is an iterative method that extracts IMFs from the given signal. It provides only finite frequency components which cannot be specified in advance, unlike wavelet decomposition. If one intends to observe spectral evolution change with respect to time, the WT is more suitable than EMD. Further, the WT provides a trade-off between time and frequency localization. Hence, for controlling the time–frequency localization, the WT is a more appropriate tool than EMD. On the other hand, the EMD is more suitable than WT when one needs to estimate instantaneous frequency of a signal accurately. The limitation of WT is that the performance of wavelet analysis depends upon type and time–frequency localization of wavelets chosen. In this study, we are using time–frequency localized wavelet filter, that outperforms others. One should use WT when there is apriori information about data. In this case, since we know the data priori, WT is preferred over EMD.

In the proposed work, we employ a new class of time–frequency localized orthogonal WFB that has minimal frequency localization and very good time localization. The entropy features such as Shannon, Tsallis, Fuzzy, Rényi, Permutation entropies and higher order statistics (HOS) are extracted from wavelet coefficients of EEG signals to classify FC and NFC of EEG signals. The proposed method gives better accuracy and sensitivity in the automatic detection of FC and NFC of EEG signals than that of the existing methods in the literature. The process flow is sketched in Fig. 2. First, the EEG signals are applied to Butterworth filter of 6th order and 60 Hz cut-off frequency in order to remove noise and other artifacts [36]. Then the signals are subjected to time–frequency localized orthogonal WFBs. Following this, the various features have been extracted from the wavelet subbands of the signals. These extracted features have been ranked and applied to least squares-support vector machine (LS–SVM) classifier for automated classification of FC and NFC of EEG signals.

The remaining paper is organized as follows: Section 2 presents details of the database used by us. The proposed method is explained in Section 3. Results are presented in Section 4 while discussions and comparisons with previous works are provided in Section 5. Conclusions and suggestions for future work are given in Section 6.

Section snippets

Dataset used

The EEG signals of FC and NFC used for this work were obtained from Bern–Barcelona EEG database [11]. The studied dataset is the largest publicly available database for intracranial EEG signals of pharmacoresistant epilepsy patients. Apart from being a public database, it is also voluminous which is necessary to provide statistical significance of computed features. Many researchers [13], [34], [35], [36], [37] have used the same database for classification of FC and NFC of EEG signals. The

Methodology

The process flow for method proposed in this paper is depicted in Fig 2. The required steps of the proposed methodology are explained as follows:

Results

In this work, orthogonal WFBs with joint time–frequency localization have been designed to extract wavelet features that are used for the identification of EEG signals as FC and NFC. Table 2 shows time–frequency localization properties of three different wavelet filters, which are designed by us in examples 1, 2 and 3. Here, L denotes length of the filter and m denotes number of VMs. The subband signals obtained from four level wavelet decomposition by employing the WFBs designed in example 1

Discussion

Recently, some other methods have been proposed to detect EEG signals of FC and NFC automatically. Table 5 presents the comparison of works carried out on automatic classification of FC and NFC of EEG signals, employing the same database. The records in Table 5 have been sorted in increasing order of accuracy achieved. As shown in the table, only a few studies have been done for the detection of focal EEG signals. It is also evident from the table that among all existing methods, our proposed

Conclusion

We have proposed an automated system for identification of FC and NFC of EEG signals by employing orthogonal wavelet filter banks which are localized in time–frequency domain. We have decomposed EEG signals using time–frequency localized filter banks to obtain wavelet coefficients. The filter banks have been designed using the proposed optimization technique.

The proposed system uses fast DWT and does not involve any iterations unlike some of the methods given in Table 5. The method is

References (68)

  • A. Delorme et al.

    Enhanced detection of artifacts in EEG data using higher-order statistics and independent component analysis

    Neuroimage

    (2007)
  • R.J. Croft et al.

    Removal of ocular artifact from the EEG: a review

    Clin. Neurophysiol.

    (2000)
  • T. Kalayci et al.

    Wavelet preprocessing for automated neural network detection of EEG spikes

    Eng. Med. Biol. Mag. IEEE

    (1995)
  • A. Rényi

    On measures of entropy and information

    Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1: Contributions to the Theory of Statistics

    (1961)
  • C. Bandt et al.

    Permutation entropy: a natural complexity measure for time series

    Phys. Rev. Lett.

    (2002)
  • W. Zhu et al.

    Detection of cancer-specific markers amid massive mass spectral data

    Proc. National Acad. Sci.

    (2003)
  • U.R. Acharya et al.

    Computer-aided diagnosis of diabetic subjects by heart rate variability signals using discrete wavelet transform method

    Knowl. Based Syst.

    (2015)
  • A. Janecek et al.

    On the relationship between feature selection and classification accuracy

    FSDM

    (2008)
  • World health organization, epilepsy, (http://www.who.int/mental_health/neurology/epilepsy/en/index.html)(last accessed...
  • R.B. Daroff et al.

    Bradley’s Neurology in Clinical Practice, 2-Volume Set: Expert Consult - Online and Print

    (2012)
  • L. Greenfield et al.

    Reading EEGs: A Practical Approach

    (2012)
  • G. Zhu et al.

    Analysis and classification of sleep stages based on difference visibility graphs from a single-channel EEG signal

    IEEE J. Biomed. Health Inform.

    (2014)
  • S. Bhat et al.

    Automated diagnosis of autism: in search of a mathematical marker

    Rev. Neurosci.

    (2014)
  • U.R. Acharya et al.

    Computer-aided diagnosis of depression using EEG signals

    Eur. Neurol.

    (2015)
  • F.M. Zanzotto et al.

    Comparing EEG/ERP-like and fMRI-like techniques for reading machine thoughts

    International Conference on Brain Informatics

    (2010)
  • R.G. Andrzejak et al.

    Nonrandomness, nonlinear dependence, and nonstationarity of electroencephalographic recordings from epilepsy patients

    Phys. Rev. E

    (2012)
  • S. Pati et al.

    Pharmacoresistant epilepsy: from pathogenesis to current and emerging therapies

    Cleve Clin. J. Med.

    (2010)
  • R. Sharma et al.

    An integrated index for the identification of focal electroencephalogram signals using discrete wavelet transform and entropy measures

    Entropy

    (2015)
  • I. Daubechies

    Orthonormal bases of compactly supported wavelets

    Commun. Pure Appl. Math.

    (1988)
  • M. Unser et al.

    A review of wavelets in biomedical applications

    Proc. IEEE

    (1996)
  • S. Maheshwari et al.

    Automated diagnosis of glaucoma using empirical wavelet transform and correntropy features extracted from fundus images

    IEEE J. Biomed. Health Inform.

    (2016)
  • M. Vetterli et al.

    Wavelets and filter banks: theory and design

    IEEE Trans. Signal. Process.

    (1992)
  • M. Sharma, A.V. Vanmali, V.M. Gadre, Wavelets and Fractals in Earth System Sciences, CRC Press, Taylor and Francis...
  • M. Sharma et al.

    Optimal duration–bandwidth localized antisymmetric biorthogonal wavelet filters

    Signal Process.

    (2016)
  • Cited by (149)

    View all citing articles on Scopus
    View full text