Classification of seizure based on the time-frequency image of EEG signals using HHT and SVM

Highlights

• A method using combined HHT and SVM is proposed for seizure detection.

• The time-frequency image based on HHT is employed for signal analysis.

• The statistical features of histogram are extracted for SVM classification.

• The best average classification accuracy over ten trails is 99.125%.

Abstract

The detection of seizure activity in electroencephalogram (EEG) signals is crucial for the classification of epileptic seizures. However, epileptic seizures occur irregularly and unpredictably, automatic seizure detection in EEG recordings is highly required. In this work, we present a new technique for seizure classification of EEG signals using Hilbert–Huang transform (HHT) and support vector machine (SVM). In our method, the HHT based time-frequency representation (TFR) has been considered as time-frequency image (TFI), the segmentation of TFI has been implemented based on the frequency-bands of the rhythms of EEG signals, the histogram of grayscale sub-images has been represented. Statistical features such as mean, variance, skewness and kurtosis of pixel intensity in the histogram have been extracted. The SVM with radial basis function (RBF) kernel has been employed for classification of seizure and nonseizure EEG signals. The classification accuracy and receiver operating characteristics (ROC) curve have been used for evaluating the performance of the classifier. Experimental results show that the best average classification accuracy of this algorithm can reach 99.125% with the theta rhythm of EEG signals.

Introduction

Epilepsy is one of the most common disorders that characterized by recurrent discharge from the cerebral cortex. As the seizures are episodic in their occurrences, detection and prediction of epilepsy from electroencephalogram (EEG) signals is a tedious and time-consuming process. It requires an expert's effort in analyzing the entire length of the EEG recordings to detect the seizure activity. Hence, new innovations for automatic detection of seizure have been sought by many researchers for a long time. However, the EEG is a highly complex signal which is nonlinear and nonstationary, the detection of the seizure from the EEG signal using traditional spectral analysis methods based on Fourier transform has some limitations, where the Fourier transform assumes that the signal being analyzed should be stationary.

Recently, many nonlinear and nonstationary methods [1], [2], [3], [4] including short time Fourier transform (STFT) [3], wavelet analysis [2], Wigner–Ville distribution (WVD) have been proposed to extract new parameters for seizure classification in EEG signals. However, these methods also have some weaknesses, for example the major limitation of STFT is unavoidable in trade-off between time and frequency resolutions, the wavelet theory is limited by the fundamental uncertainty principle, the difficulty of the WVD is the severe cross terms as indicated by the existence of negative power for some frequency ranges [5].

More recently, a novel nonlinear and nonstationary method for analyzing signals, namely Hilbert–Huang transform (HHT), has been proposed to process high dimensional EEG signals [6], [7]. A typical application of the HHT technique for seizure classification is presented in this work. As a kind of time-frequency analysis method, HHT can reveal the information of signal both in time and frequency domain. HHT is mainly applied in analyzing both nonlinear and nonstationary signal, as the core of the technique to decompose signals is posteriori and enables the extraction of the inner scales of each signal, so it has a great advantage in EEG signal processing. Compared with the above-mentioned time-frequency analysis methods, HHT is based on empirical mode decomposition (EMD) [6] which is an intuitive, direct and adaptive decomposing method using the basis of the decomposition derived from the signal. The EMD technique decomposes a dataset into a finite and often small number of intrinsic mode functions (IMFs) that admit well-behaved Hilbert transforms [8]. Recently, the applications of EMD and HHT methods are well presented by many researchers in their works [9], [10], [11], [12], [13]. Martis et al. used features of spectral peaks, spectral entropy and spectral energy computed from the IMFs for automated diagnosis of seizure [9]. Oweis and Abdulhay employed the weight frequency of IMF using HHT to detect seizure from EEG signal [10]. Bajaj and Pachori presented a method which used HHT for separation of the rhythms of the EEG signal and applied the method to classification of seizure and nonseizure EEG signals [13].

Recently, some machine learning classification techniques for seizure classification have been presented by researchers [16], [17], [18]. Among these techniques, support vector machine (SVM) has shown a good performance in classification [16]. SVM was initially developed as a binary classifier and thus it has a great advantage in binary classification problems such as seizure detection. Compared with other algorithms based on empirical risk minimization (ERM) principle, such as rule-based classifier and artificial neural network (ANN) [17], [18], SVM is based on structure risk minimization (SRM) principle. It constructs an optimal separating hyper-plane in the feature space and makes the learning machine get the global optimum [15].

This paper discusses an automatic detection of epileptic seizures based on the time-frequency image using combined HHT and SVM. The statistical features including mean, variance, skewness and kurtosis of pixel intensity in the histogram of segmented grayscale time-frequency image (TFI) have been extracted. A hypothesis testing with lower p-values indicates that the four waveforms (theta, alpha, beta and gamma) with features of mean, variance and skewness are highly determinant. The optimal features of theta, alpha, beta and gamma waves are fed into the SVM with radial basis function (RBF) kernel (RBF-SVM) for classification of seizure and nonseizure EEG signals. Experiments are conducted in ten time independent trails to test the performance of the proposed method. The results show that the proposed method can achieve the best average classification accuracy of 99.125% and provide better classification accuracy than some approaches studied previously.

The rest of the paper is organized as follows: the methodology of whole experiment including HHT method, time-frequency image processing method, feature extraction and SVM classifier are introduced in Section 2. The experimental results and discussion for the classification of seizure and nonseizure EEG signals are given in Section 3. Finally, Section 4 concludes the paper.