Automatic Detection Of High-frequency Oscillations For Accurate Localization Of Epileptic Seizure Onset Zones

High-frequency oscillations(HFOs),as a new biomarker of epileptic seizure onset zones(SOZs),are becoming a golden standardfor localizing epileptic SOZs.Conventionally,HFOs are detected by manually marking an EEG.However,this method is very time-consuming and subjective.So,the automatic detection of HFOs is crucial for thesystematic study of HFOs and for eventual clinical applications.In this thesis,the time-frequency-analysis-based and the clustering-analysis-based algorithmsforautomaticallydetecting HFOs are developed.The concentrations of detected HFOs,regarded as a localization index,are used to localize epileptic SOZs.On the one hand,in consideration of the advantages and disadvantages of the Shannon-entropy-based complex Morlet wavelet transform(SE-CMWT)algorithm and the adaptive-genetic-algorithm-based matching pursuit(AGA-MP)algorithm,this thesis presents a new time-frequency-analysis-based algorithm wherethe Hilbert transformmethod is used to identify events of interest.The SE-CMWT-based power spectral densitymethod and the Turkey’s upper fence are applied to acquire channels of interests(CoIs).By using the Morlet wavelets as thedictionary elements,the AGA-MP method combined with Morlet wavelets is developed to detectHFOs on CoIs with faster speed and higher accuracy than theAGA-MP method.On the other hand,a clustering-analysis-based algorithmis developed to solve the uncertainty of the number of clustersand to improve the detection efficiency.Four features,thatis,fuzzy entropy,short time energy,power ratio,and spectral centroid,of candidate HFOs are extracted.Then,the feature vector is regarded as the input of an expectationmaximization Gaussian mixture model-based clustering algorithm.This algorithm uses parameters calculated by a fuzzy c-means algorithm as the initial parameters,which yields a high computational speed.And the number of clusters in the data is not presupposed,but is determined as an optimal value by the quantization errormodeling.Then,a new characteristic index is defined to solve the problem that the conventional indexes have failed to express the characteristics of HFOs.The percent deviations ofthe number of HFOs,and their duration are chosen as two parameters indicating the concentration of HFOs.The concentrations of HFOs prove to be an efficient index for localizing epileptic SOZs.Moreover,the mean values of the concentration of HFOs are larger in SOZs than innon-SOZs.It is known that,the bigger the concentrations of HFOs are,the larger the possibilityof an epileptic SOZ is.Then a threshold is selected to yield the epileptic SOZs.The results show that the two algorithms in this thesis have higher localization performance than the conventional algorithms.And the time-frequency-analysis-based algorithm might be suitable for situations thatrequire a high specificity,and the clustering-analysis-based algorithm might begood for applications that require a high sensitivity.