Date of Award
Doctor of Philosophy (PhD)
G. Faye Boudreaux-Bartels
The analysis of electroencephalogram or EEG plays an important role in diagnosis and detection of brain related disorders like seizures. In this dissertation, we propose three new seizure detection algorithms that can classify seizure from nonseizure data with high accuracy. The first algorithm is based on time-domain features which are the approximate entropy (ApEn), the maximum singular value (MSV) and the median absolute deviation (MAD). These features were fed into the AdaBoost and the Support Vector Machine (SVM) algorithms, which were used to classify the signal as either seizure or non-seizure. The accuracy of these classifications was summarized and compared to different algorithms in the literature.
In the second algorithm, the Rényi entropy was extracted from different spectral components after the EEG signal was decomposed using either Empirical Mode Decomposition (EMD) or the Discrete Dyadic Wavelet Transform (DWT). The knearest neighbor (k-NN) classifier was use to classify the seizure segments based on the extracted features. In the third algorithm, we decompose the EEG signal into subcomponents occupying different spectral sub-bands using the EMD. A decomposition energy measure was used to discard those sub-components estimated to contain mostly noise. Different time-frequency representations (TFRs) were computed of the remaining sub-components. Local energy measures were estimated and fed into a linear classifier to determine whether or not the EEG signal contained a seizure. The three algorithms were tested on noisy EEG signals from roaming rats as well as the relatively noise free human seizure from a well-known public dataset provided on-line (Andrzejak et al., 2001). Using Metrics of total Sensitivity, Specificity and Accuracy, it was demonstrated that the proposed algorithms gave either equivalent or superior performance when compared against several other brain seizure algorithms previously reported in the literature.
Furthermore, we propose a new warping function to create a new class of warped Time-Frequency Representations (TFRs) that is a generalization of the previously proposed kth Power Class and Exponential Class TFRs. The new warping function is w^(t) – eatt1/k . We provide the formulas for the one-to-one derivative warping function and its inverse defined using the Lambert-W function. Examples are provided demonstrating how the new warping function can be successfully used on wide variety of non-linear FM chirp signals to linearize their support in the warped Time-Frequency plane.
An optimization scheme was proposed to find the optimal parameter, “a”, of the new warping function for a given non-linear FM chirp signal; algorithms have previously been proposed for finding the k. The performance of the optimization technique was compared to other warped Time-Frequency Representations; the new warped TFRs achieved better linearization in several cases. The new warping function was used to develop a new algorithm which iteratively isolates and separates nonlinear FM signal components in a multicomponent signal. The isolated components have negligible interference terms and have energy support concentrated along a curve close to the true instantaneous frequency.
Feltane, Amal, "Time-Frequency Based Methods for Non-Stationary Signal Analysis with Application To EEG Signals" (2016). Open Access Dissertations. Paper 445.