Elsevier

Knowledge-Based Systems

Volume 164, 15 January 2019, Pages 96-106
Knowledge-Based Systems

Epileptic seizure detection in EEG signals using sparse multiscale radial basis function networks and the Fisher vector approach

https://doi.org/10.1016/j.knosys.2018.10.029Get rights and content

Highlights

  • A novel multiscale radial basis function networks for EEG seizure detection.

  • High-resolution time-frequency images attained.

  • Discriminative texture features based on the Fisher vector encoding.

  • Two widely used independent datasets employed to test robustness.

  • High classification performance obtained.

Abstract

Detecting epileptic seizures in electroencephalography (EEG) signals is a challenging task due to nonstationary processes of brain activities. Currently, the epilepsy is mainly detected by clinicians based on visual observation of EEG recordings, which is generally time consuming and sensitive to bias. This paper presents a novel automatic seizure detection method based on the multiscale radial basis function (MRBF) networks and the Fisher vector (FV) encoding. Specifically, the MRBF networks are first used to obtain high-resolution time-frequency (TF) images for feature extraction, where both a modified particle swarm optimization (MPSO) method and an orthogonal least squares (OLS) algorithm are implemented to determine optimal scales and detect a parsimonious model structure. Gray level co-occurrence matrix (GLCM) texture descriptors and the FV, which contribute to high-dimensional vectors, are then adopted to achieve discriminative features based on five frequency subbands of clinical interests from TF images. Furthermore, the dimensionality of the original feature space can be effectively reduced by the t-test statistical tool before feeding compact features into the SVM classifier for seizure detection. Finally, the classification performance of the proposed method is evaluated by using two widely used EEG database, and is observed to provide good classification accuracy on both datasets. Experimental results demonstrate that our proposed method is a powerful tool in detecting epileptic seizures.

Introduction

Epilepsy is one of the most common neurological diseases around the world, which is mainly caused by sudden abnormal discharge of brain neurons [1], [2]. It has been reported that over 50 million people suffered from this disease, and the number was predicted to rise in the coming years [2]. Generally, expert physicians detect and diagnose the epileptic seizures by visual inspection of electroencephalogram (EEG) signals, which is a time consuming and troublesome process [3], [4]. Therefore, there is a high demand for developing an efficient automatic seizure detection system to improve efficiency of timely epilepsy diagnosis [5], [6].

The EEG is an effective tool to capture the variations in neural activity associated with epileptic seizures in the brain. Therefore, it is possible to detect seizures by directly analyzing EEG signals. The key point of EEG identification is to extract effective features to represent the characteristics of signals [7], [8], [9]. Due to the nonstationary property of EEG signals [10], [11], several EEG identification methods based on joint time-frequency (TF) analysis have been proposed for classification [12], [13], [14]. For instance, Tzallas et al. [15] employed the short time Fourier transform (STFT) to obtain the power spectrum density (PSD) of EEG signals, and extract features associated with the fractional energy of windows defined on the TF plane. Then those features were fed into an artificial neural network (ANN) for epileptic seizure classification. In the work of [16], Peker et al. analyzed EEG signals by using a dual-tree complex wavelet transformation (WT), and five statistic features derived from feature vectors were used to discriminate epileptic seizures in EEG signals. The classification results by complex-valued neural networks (CVANN) showed the effectiveness of the WT-based identification method for epilepsy diagnosis. In addition, image processing techniques have also been shown to be efficient methods to improve the classification performance by extracting new TF features [17]. Most existing image processing techniques can be used to enhance the resolution of TF representations and benefit TF features from different abnormalities [18]. For example, the texture feature coding method (TFCM), local binary pattern (LBP), and histogram of gray-scale image methods were generally used to extract TF image features for the signal classification [19], [20], [21]. Particularly, a gray level co-occurrence matrix (GLCM), as a well-known texture descriptor, is commonly used in various image processing applications such as lesion classification, texture classification and segmentation [22]. The Fisher vector (FV) encoding is an extended bags-of-visual words model for TF image representation, which leads to a more discriminative and compact image representation [23]. Alcin et al. combined GLCM texture descriptors with the FV encoding for extracting discriminative features from TF images and achieved promising results for EEG seizure classification [24]. Generally, identification methods mentioned above provide discriminative features for epileptic seizures. However, when signals are highly contaminated by noise, most of these methods based on conventional STFT and WT theories [25], [26], commonly suffer from the problem of tradeoff between time and frequency resolutions [27]. This deficiency commonly limits the effectiveness of the extracted features from EEG signals.

Parametric modeling approaches of the PSD estimation can give significant improvement of TF resolution in processing nonstationary signals, which can overcome the deficiency of time-frequency resolution from the STFT and WT methods [28], [29]. The PSD estimation based on time-varying (TV) parametric models can produce high TF resolution and further extract effective nonstationary signal features. Particularly, a robust basis function expansion method of the TF estimation was investigated in our previous work [29]. Specifically, the multiscale radial basis functions (MRBF) with a modified particle swarm optimization (MPSO) were used to approximate TV parameters in the model, which can rapidly capture abrupt changes of nonstationary signals and produce accurate TF estimation results. The MRBF-MPSO method was further applied to classify the epileptic seizure activities, and the experimental results obtained by the support vector machine (SVM) classifier indicated the good classification performance. However, two main deficiencies are involved in our recent method [29]. First, although multiple basis functions combined with the MPSO algorithm can provide a satisfactory approximation for TV parameters, the MPSO method may produce a suboptimal model with possible redundant or spurious terms in the MRBF-based expansion model when dealing with nonstationary signals [30]. Thus, the estimated TV parameters may be biased and perhaps produce the estimation of low resolution TF images. Second, we extracted the direct fractional spectral energy as features, and this feature extraction method omitted the significant information of time variation in the TF images, which may lead to an inaccurate classification result of nonstationary seizure activities.

In this paper, we extend our previous work on epileptic seizure detection [29], and propose a novel MRBF-MPSO-OLS classification method, together with the combination of GLCM texture descriptors and the FV encoding. The proposed framework mainly includes three steps. First, high-resolution TF images from EEG signals are obtained by using the MRBF-MPSO-OLS method, where the MPSO algorithm is applied to hunt for optimal model parameters, and the orthogonal least squares (OLS) algorithm, which has been proven to be an efficient method for constructing a parsimonious model [31], is applied to remove redundant or insignificant terms in MRBF-based expansion models. Second, we extracted the GLCM texture features from high-resolution TF images, and the FV encoding is then applied to the GLCM descriptors to produce high-dimensional discriminative features. Finally, in order to improve the classification performance for seizure activities, some redundant features can be efficiently eliminated by the t-test statistical tool prior to the seizure classification tasks. The selected discriminative features are fed into a SVM classifier for seizure classification. Our proposed classification scheme is validated on two publicly available EEG datasets and compared the classification performance to the results from the state-of-the-art studies. Experimental results indicate that the proposed classification framework outperforms most existing studies in terms of the classification accuracy. One contribution of the proposed scheme is that the OLS algorithm combined with the MRBF-MPSO method is applied to produce a parsimonious model with good generalization performance, which can achieve high-resolution TF images from EEG signals. Another contribution is that the GLCM descriptors and the FV encoding are adopted to extract features from these TF representations, which takes advantage of the TF variation to produce more discriminative features, and further leads to a better classification performance against the existing feature extraction studies including the conventional GLCM descriptors. As a result, the proposed scheme can be regarded as a powerful tool to detect epileptic seizure activities.

The remainder of this paper is organized as follows. The Section 2 describes EEG dataset used. In Section 3, the proposed automatic epileptic seizure detection method is introduced: a high-resolution TF image method generated by MRBF-MPSO-OLS algorithm in Section 3.1, the texture feature extraction based on GLCM and FV encoding in Section 3.2, and the classification procedure in Section 3.3. The experimental results are shown in Section 4. The discussion about the classification performance are given in the next section. Finally, the conclusion of this work is drawn in Section 6.

Section snippets

Database

In this study, two independent EEG datasets are employed to evaluate the effectiveness of the proposed seizure detection approach. Specifically, dataset-1 was recorded at the Bonn University [32], and the other which referred to as dataset-2 was acquired from Neurology and Sleep Centre Hauz Khas, New Delhi [33].

Methodology

An automatic epileptic seizure detection method based on TF images is implemented in this work. The proposed MRBF-MPSO-OLS method can produce high-resolution TF images from nonstationary EEG series. The GLCM texture descriptors, the FV encoding and the SVM classifier are then incorporated for classifying EEG signals between healthy samples and epileptic seizures. A detail illustration of the proposed framework is given in Fig. 1.

As shown in Fig. 1, the TF images are first obtained by the

Experimental results

For the classification cases in two datasets, we initially transformed the EEG segments into TF images by the proposed MRBF-MPSO-OLS method. The TF images obtained were converted into 8-bit gray-scale images, and then separated to five subband images (i.e. delta, theta, alpha, beta, gamma) based on general medical knowledge [43]. The GLCM descriptors and the FV from each sub-image have been extracted for seizure detection. Specifically, the distance parameter of the GLCM was set to 1, and the

Discussion

In this study, four different TF analysis approaches are implemented to produce TF images from EEG signals, and the discriminative FV encoding features are extracted based on the TF images for seizure classification. Fig. 2, Fig. 3 present the TF images of EEG segments by using STFT, WT, MRBF-MPSO method and the proposed MRBF-MPSO-OLS method, respectively. The TF image by the proposed method indicate that the frequency component of seizure activities is distributed in the range from zero to

Conclusion

In this paper, a novel classification framework is proposed for automatic detection of epileptic seizures by means of a sparse MRBF network and the FV encoding based on the GLCM texture features. The TV parametric models by the proposed MRBF-MPSO-OLS scheme can first provide high-resolution TF images. The seizure detection framework further takes the advantage of the GLCM texture descriptors and the FV encoding to extract compact features from TF images. The initial redundant or spurious

Acknowledgments

This study was supported by the grant from the National Natural Science Foundation of China (1671042, 61403016 and 61473196), as well as Beijing Talents foundation (016000021223TD07), Open Fund Project of Fujian Provincial Key Laboratory of Information Processing and Intelligent Control in Minjiang University (No. MJUKF201702), the Specialized Research Fund for the Doctoral Program of Higher Education (20131102120008), Project Sponsored by the Scientific Research Foundation for the Returned

References (56)

  • HassanA.R.

    Computer-aided obstructive sleep apnea detection using normal inverse gaussian parameters and adaptive boosting

    Biomed. Signal Process. Control

    (2016)
  • LiY. et al.

    High-resolution time-frequency analysis of eeg signals using multiscale radial basis functions

    Neurocomputing

    (2016)
  • HongX. et al.

    The system identification and control of hammerstein system using non-uniform rational b-spline neural network and particle swarm optimization

    Neurocomputing

    (2012)
  • SwamiP. et al.

    A novel robust diagnostic model to detect seizures in electroencephalography

    Expert Syst. Appl.

    (2016)
  • LiY. et al.

    Forecasting short-term subway passenger flow under special events scenarios using multiscale radial basis function networks

    Transp. Res C

    (2017)
  • LiY. et al.

    Learning brain connectivity sub-networks by group-constrained sparse inverse covariance estimation for alzheimer’s disease classification

    Front. Neuroinform.

    (2018)
  • WangL. et al.

    Automatic epileptic seizure detection in eeg signals using multi-domain feature extraction and nonlinear analysis

    Entropy

    (2017)
  • HassanA.R. et al.

    A decision support system for automatic sleep staging from eeg signals using tunable q-factor wavelet transform and spectral features

    J. Neurosci. Methods

    (2016)
  • LiY. et al.

    A multiwavelet-based time-varying model identification approach for time–frequency analysis of eeg signals

    Neurocomputing

    (2016)
  • SharmaM. et al.

    A new approach to characterize epileptic seizures using analytic time-frequency flexible wavelet transform and fractal dimension

    Pattern Recognit. Lett.

    (2017)
  • HassanA.R. et al.

    Epileptic seizure detection in eeg signals using tunable-q factor wavelet transform and bootstrap aggregating

    Comput. Methods Programs Biomed.

    (2016)
  • JiaJ. et al.

    Automated identification of epileptic seizures in eeg signals based on phase space representation and statistical features in the ceemd domain

    Biomed. Signal Process. Control

    (2017)
  • KumarY. et al.

    Epileptic seizure detection using dwt based fuzzy approximate entropy and support vector machine

    Neurocomputing

    (2014)
  • TawfikN.S. et al.

    A hybrid automated detection of epileptic seizures in eeg records

    Comput. Electr. Eng.

    (2016)
  • ZhuG. et al.

    Epileptic seizure detection in eegs signals using a fast weighted horizontal visibility algorithm

    Comput. Methods Programs Biomed.

    (2014)
  • HassanA.R. et al.

    Automatic identification of epileptic seizures from eeg signals using linear programming boosting

    Comput. Methods Programs Biomed.

    (2016)
  • HassanA.R. et al.

    A decision support system for automated identification of sleep stages from single-channel eeg signals

    Knowl.-Based Syst.

    (2017)
  • Krook-MagnusonE. et al.

    Beyond the hammer and the scalpel: selective circuit control for the epilepsies

    Nature Neurosci.

    (2015)
  • Cited by (106)

    View all citing articles on Scopus
    View full text