Pattern Recognition in Electric Brain Signals - mind reading in the sleeping brain

Christian Vad Karsten

AbstractA machine learning framework for analyzing experimental EEG data is presented. The question of whether the human brain is capable of more abstract processing during sleep is partly answered by analyzing data from 18 sleeping subjects tested at a semantic level using two different classes of auditory input. Using a pattern recognition algorithm it is possible to localize significant discriminating activity in 12 subjects during sleep. To validate the method, it is applied to data from the same experiment obtained during wakefulness. Here it produces significant results for 16 subjects.
The purpose of the presented pattern-based analysis is twofold. The first objective is to consider whether classification is possible with the underlying presumption that if a classifier can label new examples with a better accuracy than chance, then the two conditions are indeed differently represented in the brain. The second is to make claims about information representation in the brain. Both objectives are fulfilled. Regardless of differences in latency and morphology at a single-subject level, patterns similar to results from the relevant literature concerning wakefulness do arise. This can be an indication of cognitive processing during sleep all the way up to motor planning.
The presented results are obtained using a novel method for image based analysis of EEG spectrograms at the sensor level. A non-linear support vector machine is trained directly on spectrograms and combined with an embedded feature selection scheme to overcome the challenges posed by low sample size high dimensional data. Opposite to classical analysis of EEG data, this method allows analysis at an individual subject level, where results are normally obtained at a group level. In addition to answering the question of whether there is information of interest (pattern discrimination), the method also to some degree answer the questions of where and how the information is encoded (pattern localization and pattern characterization).
TypeMaster's thesis [Academic thesis]
Year2012
PublisherTechnical University of Denmark, DTU Informatics, E-mail: reception@imm.dtu.dk
AddressAsmussens Alle, Building 305, DK-2800 Kgs. Lyngby, Denmark
SeriesIMM-M.Sc.-2012-91
NoteSupervised by Lars Kai Hansen, lkh@imm.dtu.dk, DTU Informatics, Carsten Stahlhut, and Sid Kouider
Electronic version(s)[pdf]
Publication linkhttp://www.imm.dtu.dk/English.aspx
BibTeX data [bibtex]
IMM Group(s)Intelligent Signal Processing


Back  ::  IMM Publications