| Epilepsy was defined conceptually in 2005 as a disorder of the brain characterized by an enduring predisposition to generate epileptic seizures.According to the report of World Health Organization(WHO),approximately 50 million people worldwide have epilepsy,making it one of the most common neurological diseases globally.Electroencephalography(EEG)visually inspected by the board-certified neurologist is usually treated as the clinical gold standard.EEG is the simplest,cheapest,and probably most effective modality for providing information related to epilepsy diagnosis.Many computer-aided methods have been developed to assist neurolo-gists in analyzing epileptic EEG in the past decade.Combining signal processing and machine learning,most approaches model the clinical tasks into classification problems,such as epilep-tic vs.normal for epilepsy diagnosis,ictal(seizure onset)vs.interictal(between seizures)for seizure onset detection,interictal vs.preictal for seizure prediction,etc.In recent years,Wavelet Transform(WT)has shown promising results in epilepsy-related EEG signal classification,in-cluding seizure detection,and seizure diagnosis.Human brain is a high complex biological system.EEG is nonstationary signal which measures voltage fluctuations resulting from ionic current within the neurons from the deep of the brain.EEG recording usually cost hours to days requiring tedious visual inspection performed by highly qualified neurophysiologists.WT is quite suitable for the analysis of EEG according to the nonstationary and the time-consuming inspection.We creatively performed the following investigations:1.Using wavelet transform in epileptic focus localization.Epileptic focus localization is a challenging but important task in epilepsy diagnosis and treatment.It is usually a pre-surgical procedure to estimate the epileptogenic zone(EZ).Neurosurgeons have been using EEG as a standard clinical practice to localize the EZ since 1972.Since traditional EEG-based methods for epileptic focus localization require tedious visual inspection per-formed by highly qualified neurophysiologists,there is a growing need for automating it.Considering this,an epileptic focus localization method based on Discrete Wavelet Trans-form(DWT)was constructed here.This algorithm aims at constructing a generalizable computer-aided signal analysis method to achieve the optimal epileptic focus localization accuracy with low computational cost.2.Systematically investigating the effects of various factors in wavelet-based epileptic EEG analysis.The previous work using wavelet for epileptic EEG analysis are partial and pilot.We considered all the four factors affecting the accuracy and the computational cost of any DWT-based approach:the mother wavelet,the level of decomposition,the frequency band and the features extracted from DWT coefficients.Computational re-sults indicate that the priorities of the fours factors varies in different datasets.In clinical practice,the algorithm should be modified to fit specific situation.3.Constructing a two-block feature selection method based on wavelet and machine learning to improve the accuracy in epileptic focus localization and seizure detec-tion.Considering the four factors affecting the algorithm performance,we construct a two-block selection based on wavelet and machine learning to efficiently improve the ac-curacy(>80%for epileptic focus localization,>90%for seizure detection)and reduce the computational cost in epileptic focus localization and seizure detection.The approach has the theoretical consequences and is clinically worth-trying.The main contents in this work are as follows:1.Epileptic focus localization.Epileptic focus localization is formulated into a binary clas-sification problem,where signals from multiple EEG channels belong to two categories,"focal" or "non-focal".The algorithm contains two selection blocks,including,Wavelet-Level Selection(Block 1)and Band-Feature Selection(Block 2).In Block 1,the EEG signals were decomposed to the maximum decomposition level.Each EEG segment was transformed into a feature vector which contained features extracted from all detail bands and the last approximation band.The best wavelet in each wavelet family and the cor-responding decomposition level were retained as the input parameters in Block 2.EEG is typically described in terms of rhythm.Different frequency bands in DWT correspond to various EEG rhythms.Given an EEG dataset,focal and non-focal EEG segments may exhibit significant difference in specific rhythmic activity.DWT coefficient features have similar characteristics as frequency bands.In a specific EEG dataset,some features may help distinguish focal and non-focal EEG segments while other features produces feature redundancy only.Therefore,BandFeature Selection is introduced to search for the bands and features that lead to the most accurate classification and the least feature redundancy.We find that the choice of wavelet family is not strongly related to the performance if given a sufficient level of decomposition,and that there is a high redundancy in DWT bands and coefficient features.The algorithm established in this work shows strong ro-bust output in epileptic focus localization using DWT,with over 80%accuracy on two de facto iEEG datasets.The algorithm should be modified to fit different situations in clinical practices.2.Seizure detection.Similar to epileptic focus localization,seizure detection,a common clinical problem,could also be transformed into a binary classification problem,i.e.,ictal vs.inter-ictal.The DWT-based algorithm in epileptic focus localization was modified to fit the requirement in seizure detection.When a series of preprocessing steps(filtering,cutting,grouping,etc)were finished,Block 1 searched the best wavelet and correspond-ing decomposition level in each wavelet family.The output in Block 1 was the input in Block 2.The frequency band(s)and feature(s)leading to the optimal seizure detection accuracy with the lowest computational cost made up the output in Block 2.Experiment results show that the accuracy is very sensitive to decomposition level regardless of the mother wavelet when classifying some complex EEG signals.Otherwise,the accuracy is sensitive to neither of them.Due to the structure difference between EEG datasets,var-ious features and frequency bands have different significance in seizure detection.The Band-Feature Selection can abandon redundant features and frequency bands to improve detection accuracy and computation efficiency greatly.The algorithm should be modified to fit different situations in clinical practices.Overall,properly setting the parameters in the DWT-based seizure detection algorithm can achieve an accuracy over 90%.3.Generalized feature extraction based on EEG.The performance of wavelet in epilep-tic focus localization and seizure detection pointed out that feature extraction should be treated as the key in epileptic EEG analysis.The feature could be extracted according to statistical characteristics,patient pathology,or other "black boxes"."Emotion",which might be used as a "black box" in epileptic EEG analysis was introduced here.An exper-iment of using EEG to track the real-time emotion changes when someone playing video game was carried out.If video game events can produce trackable emotion changes,re-searchers might used a similar method for epileptic EEG analysis.Using "black box" like emotion in epileptic EEG analysis is worth-trying since this method is easy-achieving and completely noninvasive. |
PDF Full Text Request