EEG artefact removal methods compared using semi-synthetic data for the analysis of ADHD EEG-Data.

Date
2021-12
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
ENGLISH ABSTRACT: The electroencephalogram (EEG) is a measure of the biological electrical sig- nals that reflect the brain’s functional state, allied to a person’s mental condi- tion and nervous system activity. EEG is, however, an extremely weak signal, thus easily contaminated by artefacts. Artefacts are activities that do not di- rectly originate from the brain but are still present in the EEG data. Artefacts significantly complicate, distort and obscure the analysis of the data originat- ing exclusively from the brain. This thesis aimed to identify, test, and automate robust methods for remov- ing EEG artefacts. This aim was achieved by developing a simulated dataset, based on real ‘clean’ EEG and artefacts, namely a semi-synthetic dataset with significant variation in types, forms, intensity and combinations of physiolog- ical artefacts. This dataset was used to test the effectiveness and efficiency of three blind source separation (BSS) techniques, namely Extended Infomax, second-order blind identification (SOBI), canonical correlation analysis (CCA) and a developed auto threshold method. Additionally, the BSS methods were fully automated, using a novel but simple approach, to prevent the preprocess- ing of EEG data from becoming a bottleneck for data analysis. The semi-synthetic dataset consisted of ‘clean’ EEG datasets, which was contaminated separately and together by electrocardiography (ECG), elec- trooculography (EOG) and electromyography (EMG) artefacts varied for each EEG dataset. This thesis compared the time-series, topography, amplitude spectra, and the signal-to-noise ratio (SNR) characteristics of real ‘clean’ EEG and artefacts, found in the relevant literature, to the characteristics of the semi-synthetic dataset developed, for the purpose of validation. BSS techniques were used because they were the most popular for artefact removal. Furthermore, independent component analysis (ICA) was identified as the most popular BSS subcategory. The two most popular ICA methods, namely Extended Infomax and SOBI, were tested along with CCA. In addi- tion, an auto threshold method, using the standard deviation of the data, was created and tested on the data. The effectiveness of the cleaning methods was determined and compared using the SNR increase of the contaminated data. The efficiency was deter- mined and compared using the average time each BSS method took to identify the components and for the auto threshold method took identify all the arte- fact ranges. The SNR and time results were further analysed using boxplots and t-tests. With the removal of EOG artefacts, CCA was the most effective. Extended Infomax was the most effective with the removal of EMG artefacts. With the removal of ECG artefacts, SOBI outperformed the other methods in terms of effectiveness. Furthermore, when combining all three artefacts, the effec- tiveness of the BSS methods was less distinguishable, having closer P values, with Extended Infomax being the most effective. The auto threshold method showed comparable effectiveness results to the BSS methods, but in terms of efficiency, it was about 10, 20 and 100 times faster than CCA, Extended Info- max and SOBI, respectively. Concerning the automation of the BSS methods, the fully automated and semi-automatic Extended Infomax methods showed no significant difference, based on a t-test, in the effectiveness of EOG removal.
AFRIKAANSE OPSOMMING: Die elektroencefalogram (EEG) is ’n maatstaf van die biologiese elektriese seine wat die brein se funksionele toestand weerspieël, wat verband hou met ’n persoon se geestestoestand en aktiwiteit van die senuweestelsel. EEG is egter ’n uiters swak sein, wat dus maklik deur artefakte gekontamineer kan word. Artefakte is aktiwiteite wat nie direk uit die brein afkomstig is nie, maar steeds in die EEG-data voorkom. Artefakte bemoeilik, verdraai en verduister die analise van die data wat uitsluitlik uit die brein afkomstig is. Hierdie tesis het ten doel gehad om robuust metodes vir die verwydering van EEG-artefakte te identifiseer, te toets en te outomatiseer. Hierdie doel is bereik deur ’n gesimuleerde datastel te ontwikkel, gebaseer op werklike ‘skoon’ EEG en artefakte, naamlik ’n semi-sintetiese datastel met ’n beduidende vari- asie in tipes, vorms, intensiteit en kombinasies van fisiologiese artefakte. Hier- die datastel is gebruik om die doeltreffendheid en tyddoeltreffendheid van drie Blinde bron skeiding (BBS) tegnieke te toets, naamlik Uitgebreide Infomax, tweede-orde blinde identifikasie (TOBI), kanonieke korrelasie-analise (KKA) en ’n ontwikkelde outomatiese drumpel metode. Die BBS-metodes is volledig geoutomatiseer, met ’n nuwe maar eenvoudige benadering, om te voorkom dat die voorafverwerking van EEG-data ’n ‘bottleneck’ word vir data-analise. Die semi-sintetiese datastel bestaan uit ‘skoon’ EEG-datastelle, wat afson- derlik en saam gekontamineer is deur elektrokardiografie (EKG), elektrookulo- grafie (EOG) en elektromyografie (EMG) artefakte wat vir elke EEG-datastel gevarieer is. Hierdie tesis vergelyk die tydreekse, topografie, amplitude-spektra n die sein-tot-geraas-verhouding (SGV) kenmerk van die ‘skoon’ EEG en ar- tefakte in die relevante literatuur met die kenmerke van die semi-sintetiese datastel wat ontwikkel is vir validasie doeleindes. BBS-tegnieke is gebruik omdat dit die gewildste is vir die verwydering van artefakte. Verder is onafhanklike komponent analise (OKA) geïdentifiseer as die gewildste BBS-subkategorie. Die twee gewildste OKA-metodes, naamlik Extended Infomax en TOBI, is saam met KKA getoets. Boonop is ’n outoma- tiese drumpel metode, met behulp van die standaardafwyking van die data, geskep en getoets op die data. Die doeltreffendheid van die skoonmaakmetodes is bepaal en vergelyk met behulp van SGV-toename van die gekontamineerde data. Die doeltreffend- heid is bepaal en vergelyk deur gebruik te maak van die gemiddelde tyd wat elke BBS-metode geneem het om die komponente te identifiseer, en vir die outomatiese drumpelmetode al die artefakreekse te identifiseer. Die SGV- en tydresultate is verder geanaliseer met behulp van boksplotte en t-toetse. Met die verwydering van EOG-artefakte was KKA die doeltreffendste. Uitgebreide Infomax was die doeltreffendste met die verwydering van EMG- artefakte. Met die verwydering van EKG-artefakte het TOBI beter gevaar as die ander metodes wat doeltreffendheid betref. Boonop was die doeltreffend- heid van die BBS-metodes by die kombinasie van al drie artefakte minder on- derskeibaar, met nader P-waardes, met Extended Infomax die doeltreffendste. Die outomatiese drumpel metode het vergelykbare doeltreffendheid resultate getoon met die BBS-metodes, maar wat tyddoeltreffendheid betref, is dit on- geveer 10, 20 en 100 keer vinniger as onderskeidelik KKA, Extended Infomax en TOBI. Met die BBS-metodes outomatiser, het die volledig outomatiese en semi-outomatiese uitgebreide Infomax-metodes geen beduidende verskil getoon op grond van ’n t-toets in die doeltreffendheid van EOG verwydering nie.
Description
Thesis (MEng)--Stellenbosch University, 2021.
Keywords
Electroencephalography, Electromyography, Central nervous system, Electrooculography, Semi-synthetic -- Dataset, UCTD
Citation