Classification of focal and non-focal epileptic EEG signals inEMD-wavelet domain

Abstract

In this thesis, a comprehensive analysis of focal and non-focal electroencephalography is carried out in the empirical mode decomposition (EMD) and discrete wavelet transform (DWT) domains. First, the analysis is carried out in the EMD domain and its variants, for example, in ensemble empirical mode decomposition (EEMD) and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) domains. A number of spectral entropy-based features such as the Shannon entropy, log-energy entropy and Renyi entropy are calculated in EMD, EEMD and CEEMDAN domains. In lieu of using the direct signals from the EEG channels, the differences between two adjacent EEG channels are used due to its robustness to noise and interference. The EEG signals are obtained from a publicly available electroencephalography database that consists of 7500 signal pairs which contain over 80 hours of electroencephalogram data collected from five epilepsy patients. Then, the ability of the entropy-based features in separating the focal and non-focal EEG signals is explored utilizing the one-way ANOVA analysis and the box-whisker plots. After that, well-known classifiers like support vector machine (SVM) and k-nearest neighbor (KNN) have been utilized to classify focal and non-focal EEG signals. Next, similar analysis is carried out in discrete wavelet transform domains and the efficacy in discriminating the focal and non-focal EEG signals is investigated. It is observed that the entropy-based features perform better in DWT domain to classify the EEG signals than in EMD domain. In this regard, it is interesting to investigate the capability of the same features to discriminate the EEG data in the combined EMD-DWT domain. It is shown that in the log-energy entropy, when calculated in the combined EMD-DWT domain, gives a better discrimination of these signals as compared to that of the other entropy measures as well as to that obtained in EMD or DWT domain, and utilizing a KNN classifier, it provides 89.4% accuracy (with 90.7% sensitivity), which is higher than that of the state-of-the-art methods. Overall, the proposed classification method reports a significant improvement in terms of sensitivity, specificity and accuracy in comparison to the existing techniques. Besides, for being computationally fast, the proposed method has the potential for identifying the epileptogenic zones, which is an important step prior to resective surgery usually performed on patients with low responsiveness to anti-epileptic medications. The analysis may encourage the researchers to develop improved algorithms to classify these signals.