Dissertation/Thesis Abstract

Towards Efficient Auditory BCI Through Optimized Paradigms and Methods
by Schreuder, Evert-Jan Martijn, Dr.Nat., Technische Universitaet Berlin (Germany), 2014, 176; 10697509
Abstract (Summary)

To date, the brain-computer interface (BCI) based on visual stimulation is by far the most investigated. This makes a lot of sense given the excellent visual capabilities of humans, and it has been a most successful approach. Nevertheless, a non-negligible part of the BCI end-user population with advanced paralysis is incapable of directing their eye gaze or of seeing at all. Traditional visual BCIs will rarely be a solution for these end-users and alternative paradigms are required that rely on covert attention. Furthermore, though BCIs already benefit greatly from statistical methods coming from the field of machine learning, a faster and robust performance is required for adoption by the clinical community. This thesis contributes in two key facets in an attempt to take a step towards full inclusion of all end-users. First, an auditory alternative is proposed to address the intrinsic problems that exist with BCIs based on the visual event-related potential (ERP). It is named AMUSE (short for Auditory MUlticlass Spatial Event-related potential). Though it is not the first auditory BCI to be proposed, it differs from the mostly binary contemporary solutions in a crucial aspect. Instead of relying on pitch or other physical features of the stimuli to define the class membership, AMUSE targets the human capability of spatial localization of sound sources. With AMUSE, stimuli can contain physical differences but the main defining feature is the sound source or, more practically, the loudspeaker position. Using this feature, the stimuli can be short, stimulation speed can be high, and the number of classes can easily be increased by adding sources. The fundamentals of AMUSE as a BCI paradigm are carefully validated, and an AMUSE-based BCI speller is proposed with minimal reliance on the visual domain. The vast majority of the tested healthy subjects is able to write a full sentence using this speller. The average performance is high compared to contemporary covert attention-based BCIs, both in terms of the information transfer rate (ITR) and the amount of written characters per minute (char/min). Most BCI research is directed towards the final goal of successful end-user application. Accordingly, the AMUSE-based speller is tested with five end-users with advanced paralysis. Up to four online sessions are performed for each end-user, and performance is above chance on all but two sessions. Nevertheless, the performance does not suffice for the end-users to gain meaningful control over the AMUSE-based speller. This result is discussed in the context of the available attention resources, which was pathologically low for at least one end-user. The second important contribution of this thesis is a novel algorithm for performance optimization in ERP-based BCI in general, called rank diff. ERP-based BCIs have an intrinsic trade-off between the length of a trial and the accuracy of the resulting decision; longer trials typically result in more accurate decisions. In literature, this trade-off is mostly ignored and the length of the trial is simply fixed, leading to suboptimal performance. Rank diff uses the available calibration data to estimate a threshold on the required evidence necessary for a correct decision. Online, a trial is stopped as soon as this threshold is reached, leading to trials of varying length. By including rank diff in the AMUSE-based speller, the performance increases by about 40% for healthy subjects. As a result, the speller is competitive with covert attention-based visual spellers. Rank diff is further validated on benchmark data along with several other methods, showing the beneficial effect not only of rank diff, but of evidence accumulation methods in general. As average performance increases as high as 78% are found, these methods can no longer be ignored.

Indexing (document details)
Advisor: Müller, Klaus-Robert
School: Technische Universitaet Berlin (Germany)
School Location: Germany
Source: DAI-C 81/1(E), Dissertation Abstracts International
Subjects: Computer science
Keywords: Brain-computer interface
Publication Number: 10697509
ISBN: 9781392571743
Copyright © 2020 ProQuest LLC. All rights reserved. Terms and Conditions Privacy Policy Cookie Policy