direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Page Content

PhD Theses

Machine learning techniques for neurotechnology with applications for healthy users and patients
Citation key Hoehne2014
Author Höhne, Johannes
Year 2014
School Technische Universität Berlin
Abstract Advances in Neurotechnology are based on the recording and analysis of brain activity. Brain- Computer Interfaces (BCIs) constitute a very active research area within Neurotechnology. BCIs make it possible for brain activity to be directly translated into control commands and thus enable a communication channel that is independent of muscle control. One major goal of this research is to help people who cannot communicate independently, due to neural diseases such as stroke or Amyotrophic Lateral Sclerosis (ALS). BCIs may help these patients with advanced paralysis regain their communication abilities by using their minds to interact with their surroundings. This thesis contributes to the developments of BCIs in three ways. Firstly, novel auditory BCI paradigms – named PASS2D and CharStreamer – are described and evaluated in online studies with healthy users. Both paradigms are based on Event Related Potentials (ERPs) and provide an intuitive and fast communication with BCI for users with impaired vision. While prior auditory BCI paradigms are rather complicated to use, the CharStreamer can be operated with instructions as simple as "please attend to the letter that you want to spell". Additionally, two offline studies investigate the impact of stimulus properties on the ERPs, and the performance and usability of a BCI system. However, the above mentioned studies also indicate that the state-of-the-art analysis pipeline for ERP-based BCI paradigms might be suboptimal, as ERPs exhibit additional label information which is not exploited in a Linear Discriminant Analysis (LDA). Therefore, the second main contribution of this thesis deals with methodological improvements which yield more accurate data analysis than state-of-the-art methods. It is shown that neuroimaging data – in particular EEG data arising from BCI paradigms – exhibit intrinsic subclass structure, which can be exploited in a meaningful way. A novel Machine Learning method – called Relevance Subclass LDA (RSLDA) – is developed and tested on multiple EEG and fMRI data sets. It is shown that RSLDA yields increased classification accuracy, as well as a better interpretation of the underlying structure in the data. Both aspects are highly favorable, suggesting that RSLDA is suitable for various classification problems within neuroimaging and beyond. Thirdly, a BCI study is conducted with severely motor-impaired individuals. It is shown that the application of modern Machine Learning methods allows to set up a highly flexible BCI system for patients with severe paralysis. This enables to achieve significant BCI control within a very small number of sessions. Moreover, this study shows that communication via BCI can be faster and more robust than communication with other assisted technology which is based on muscle activity. This shows for the first time that the neuronal signals of an attempted motor execution can be detected prior to the muscular movement of a patient.
Link to publication Download Bibtex entry

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe