FEATURE EXTRACTION AND INTERPRETATION OF EEC SIGNALS

Abstract

Epilepsy is a chronic noncommunicable neurological disorder of the brain which is It associated with the recurrent, unprovoked seizures resulting from a sudden disturbance of brain function. Usually long term monitoring of EEG of epileptic patients is needed for the diagnosis of seizures. The visual examination of epileptic seizures from this long term EEG recordings are highly labour intensive, very tedious, expensive, and time consuming. Also a clinician may miss to identify a seizure from long term recoding during the examination of EEG recordings. This motivates us to develop a computer based system for epileptic seizure detection, which not only facilitates the long-term examination of EEG for epileptic seizures detection but could also assist the clinician to control seizures by giving therapeutic agents as early as possible and would improve the quality of life and safety for the affected patients. Common epilepsy treatments are medication and surgery which may have not only severe side effects but also fail to control the seizures satisfactorily of epileptic patients. For the treatment of epilepsy, patients undergo for continuous intracranial EEG monitoring. The success of treatment through surgery heavily depends on the accurate identification, localization of the brain regions involved in epilepsy and cause for origination of the partial seizure. Therefore, a reliable EEG classification system could assist the clinicians to streamline the focus of partial seizure which could be helpful in presurgical evaluations for epilepsy and their possible treatment plan. Based on the above background, major objectives of the present research work are framed as (i) Feature extraction from the EEG signals for evaluation of Epilepsy; (ii) Design and implementation of various feature selection methods to identify the optimal feature set of extracted features for classification of EEG signals; (iii) Design and development of binary classifiers to achieve accurate detection of epileptic seizure for interpretation of EEG signals; and (iv) Design and development of the computer based multiclass EEG classification system. Feature extraction approach is designed to obtain the important features of EEG signals which extract the maximum possible discriminatory information in terms of extracted feature set. For this work, features such as sample entropy, approximate entropy, and spectral entropy are computed to measure the complexity of these signals. The epileptic EEG signals are rhythmic in nature and their amplitude is in increasing order during the seizure periods. The above mentioned entropy features are highly suitable for the signals which have the periodicity and also have the varying magnitudes. After the suitable analysis of these features it is found that these entropy features are capable enough to provide the significant differences between the normal and epileptic subjects. Application of feature selection step prior to the classification is required to reduce the dimensionality of features set and to improve the classification accuracy. A feature set is normally composed of several features. Different combination of these features may receive different classification accuracies and garbage features may appear in each set. Feature selection methods such as filter methods and wrapper methods are used to eliminate the irrelevant features from the feature set to evaluate the optimal subset of features for classification. The different binary classifiers have been designed to improve the statistical parameters of classifiers to detect the epileptic seizures. The works reported for the analysis of physiological signals in time-frequency tells us that different frequency sub-bands of these signals provide more useful information of physiological signals as compared to the whole signals. Therefore, EEG signals are decomposed into different sub-bands through wavelet filters to design the Relative wavelet energy (RWE) based approach for detection of epileptic seizures. In this approach RWE is computed from each sub-band of EEG signals using fifth level decomposition through bi-orthogonal wavelet filter The RWE is the measure of wavelet energy associated with different frequency sub-bands of signals and also their corresponding degree of importance. Based on analysis of RWE feature of each sub-band, four features are selected to classify the EEGs into two classes. Statistical parameters of different classifiers are computed using these features. It is observed that the accuracy obtained from this approach is not improved significantly Therefore, Approximate entropy based approach is proposed, where some features like average power, standard deviation and approximate entropy are calculated from these sub-bands to form a feature vector to detect the epileptic seizure using different classifiers. Box plots are used to provide the graphical summary of data sets through their quartile. These plots are also used to analyze the features of each sub-band of signals to get the optimum feature vector for classification. To overcome the limitations of approximate entropy a feature has been derived. Further, this feature has been utilized to form a feature vector to improve the seizure detection rate of binary classifiers. Literature survey on the analysis of EEG signals using approximate entropy, found that the vector simiTarity is defined on the basis of hard sensitive boundary of Heaviside function. This shows that the approximate entropy (ApEn) is highly sensitive to parameter selection and may be invalid in case of small parameters. This caveat of ApEn is rectified by proposing a fuzzy function where the vector similarity is defined on the basis of soft and continuous boundary. This function will provide the continuity as well as the validity of entropy at small parameters. Physiological signals are decomposed into different sub-bands using discrete wavelet transform (DWT). Daubechies order four (db4) wavelet filter is suitable for analyzing EEG signals because of its orthogonality property and efficient filter implementation. Therefore, in this work Daubechies family of wavelet filters is chosen among the various wavelet filters and db4 is employed as basis for wavelet decomposing of the EEG signals. From the analysis of fuzzy approximate entropy (fApEn), it is observed that the quantitative value of fApEn drops significantly during a seizure interval which proves that the signals during seizure intervals are more regular than the data in interictal intervals. The fApEn values are calculated from these sub-band signals to extract the features and to form feature vector thereof. This feature vector is used as input to SVM for classifying the EEG data sets into normal/interictal versus ictal data sets. Data sets used in the above work have only 500 EEG signal samples of different classes. Only 50 or 60% feature data of different classes are used to train the classifiers and the remaining ratio of percentage is used for detection of epileptic seizures. To overcome the limitation of limited data used for proper training of classifiers, the framing of data set is used and an approach based on fApEn and the largest Lyapunov exponent (LL) is proposed to design the binary classifiers to improve the detection accuracy. The following stages have been incorporated to carry out the above work: (I) Framing of EEG database: each segment of EEG data sets is framed into 16 sub segments each having 256 samples. (ii) Wavelet analysis: three wavelet based features are considered to analyze the EEG signals; first, fApEn to measure the complexity of each sub-band; second, the largest Lyapunov exponent to quantify the chaoticity of each sub-band; and the last, standard deviation to quantify the sub-bands signal variance. (iii) Feature selection: statistical analysis of these features is performed by two sample t-test to find out the significant difference between the feature values of each sub-band of each data set. These differences are utilized to form a feature vector to detect the epileptic seizure using SVMs. (iv) Classifier design: the performance evaluation indices of classifier are compared for different experiments which are devised based on the combination of sub-bands features It is observed that the combination of D1-D5 sub-band features give the best classification results which are significantly higher than the reported results by other investigators. To design the computer based multiclass EEG classification system, the following tasks have been performed: Comparison of wavelet families for selection of the best wavelet filter, selection of features using genetic algorithm, feature space dimensionality reduction using PCA, and design of multiclass classifiers. Mutticlass classification system is designed to classify the normal, interictal, and ictal EEGs These different epileptic data EEGs are generally required for pre-surgical evaluation of epilepsy. Various multi-resolution features such as mean, standard deviation, energy and fApEn are computed using various compact support wavelet filters including Daubechies, Biorthogonal, Symlet and Coiflet. Characterizations of EEG signals in transform domain reveal that the selection of wavelet filter is highly essential because improper selection of wavelet filter will affect the characterization performance of the classifier system. The classifier characterization performances obtained from the 35 different wavelet filters (such as db2 to dblo, sym2 to sym8, coifi to coif5, bior 1.5 to bior 6.8) are compared to search the best wavelet filter. It is found that the Daubechies family is the best family over other wavelet families. The Daubechies wavelet filters of order seven (db7) give the best performance evaluation indices than the other wavelet filters for classification of EEG signals into three classes using LIBSVM. The k-fold cross-validation technique is used to obtain the best optimum training parameters of predictive model. It is observed that the combined feature vector of length 24 obtained using db7 gives the best classification accuracy. During feature selection, GA-SVM is being employed for obtaining the optimal feature sub set from the feature set. An adequate representation of data and appropriate fitness function are taken into account during the implementation of GA. The number of feature reduces upto 50% of original feature set using GA-SVM without much compromise in the classification results by removing the irrelevant features. Feature space dimension reduction is performed by PCA to get the optimum feature vector space for classification. Dimension of the feature vector is always taken care during designing of classifier system. It should always be smallest otherwise interference of extraneous feature can lead to reduced learning performance of the classifier. This will further increase the time taken to perform the classification and it may reduce the classification accuracy. The performances of proposed methods are evaluated on the EEG data sets which are obtained from the University of Bonn, Germany. These EEG signals are recorded from the normal and epileptic subjects by surface and intracranial electrodes. Results obtained from the present work are compared with the other existing methods. Finally, it is observed that the performance evaluation indices of proposed epileptic seizure detection methods are significantly improved over the methods proposed by other researchers. Overall accuracy for classification of different classes using LIBSVM is significantly improved and number of features also reduces appreciably to get the maximum classification accuracy over the methods proposed by other researchers.