Sparse Bayesian Learning for EEG Source Recovery
Authors
Kragh, Trine Nyholm ; Mogensen, Laura Nyrup
Term
4. semester
Education
Publication year
2020
Submitted on
2020-06-03
Pages
119
Abstract
Specialet undersøger, hvordan man kan genskabe hjernens oprindelige kildesignaler ud fra lavdensitets-EEG, hvor målinger fra få elektroder på hovedbunden gør problemet særligt vanskeligt. Målet var at reproducere førende resultater med en algoritme, der kombinerer to metoder: covariance-domæne dictionary learning (Cov-DL), som forsøger at lære mønstre ud fra sammenhænge mellem kanaler, og multiple sparse Bayesian learning (M-SBL), en bayesisk metode til at udtrække få, aktive kilder blandt mange mulige. Den planlagte anvendelse af Cov-DL virkede ikke i denne sammenhæng, så der blev foreslået en alternativ løsning. Den endelige algoritme blev testet på EEG-data og sammenlignet med løsninger opnået ved uafhængig komponentanalyse (ICA) på høj-densitets-EEG. Derudover blev frekvensindholdet i rå EEG og de genskabte kilder sammenlignet. Af hensyn til praktisk anvendelse blev der også foreslået en metode til at anslå det ukendte antal aktive kilder. Konklusionen er, at den foreslåede samlede algoritme ikke kunne reproducere de bedste publicerede resultater. Til gengæld viste M-SBL alene sig at fungere godt, og metoden rummer potentiale til at estimere, hvor mange kilder der er aktive.
This thesis examines how to recover the brain’s original source signals from low-density EEG, where only a small number of scalp electrodes are available and the problem is especially challenging. The aim was to reproduce leading published results with an algorithm that combines two methods: covariance-domain dictionary learning (Cov-DL), which tries to learn patterns from relationships between channels, and multiple sparse Bayesian learning (M-SBL), a Bayesian approach that extracts a few active sources from many candidates. The proposed use of Cov-DL did not succeed in this setting, so an alternative solution was put forward. The final algorithm was tested on EEG data and compared with solutions obtained by independent component analysis (ICA) on high-density EEG. A frequency (spectral) analysis also compared the raw EEG with the recovered sources. For practical use, a way to estimate the unknown number of active sources was proposed. The study concludes that the combined algorithm could not reproduce state-of-the-art results. However, M-SBL on its own performed well and shows potential for estimating how many sources are active.
[This abstract was generated with the help of AI]
Keywords
Documents