Files in this item

FilesDescriptionFormat

application/pdf

application/pdfNORTON-THESIS-2019.pdf (1MB)Restricted Access
(no description provided)PDF

Description

Title:Applications of sequential hypothesis testing to the development of non-invasive brain-computer interfaces
Author(s):Norton, James J. S.
Advisor(s):Bretl, Timothy W.
Department / Program:Electrical and Computer Engineering
Discipline:Electrical and Computer Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Degree:M.S.
Genre:Thesis
Subject(s):brain-computer interfaces
SSVEP
N400
Abstract:Sequential hypothesis testing is applied in this thesis to make two contributions to the classification of electroencephalography (EEG) data for use in non-invasive brain computer interfaces (BCIs). The first contribution is a variable window-length classification method for use in a steady-state visual evoked potential (SSVEP)-based BCI. Instead of relying on a fixed window-length strategy—where a pre-specified amount of data is collected before classification is attempted—a sequential probability ratio test is used and data is collected until a confidence threshold is met. This variable window-length strategy was tested using a simple experiment where one of five visual stimuli were presented one at a time to three participants. An analysis of the data collected during this simple experiment show that the information transfer rate was improved by 43% when using the variable window-length strategy as compared to the fixed window-length strategy. The second contribution is an analysis showing that it is possible to classify expected versus unexpected endings to strongly constrained sentences at better than chance accuracies using single trials of EEG. This better than chance classification accuracy is demonstrated for features based on two different event-related potentials (ERPs) elicited in response to the neural processing of meaning-related information—the N400 and the frontal positivity (FP). Using an existing dataset, classification accuracies were computed for features based on each of the two brain signals for three different classifiers (Naïve Bayes [NB]; linear discriminant analysis [LDA]; and support vector machines [SVM]), three different electrode groupings, and three different ways of analyzing the individual trials (single trial classification; classification after averaging multiple trials; and the sequential classification of trials using a sequential probability ratio test [SPRT]). Using single trials of EEG, features based on the N400, and all 26 EEG electrodes, classification accuracy with an LDA classifier was 59.96%. In analyses with features based on both the N400 and FP, classification accuracies were higher (59.25%) when three trials were averaged together before classification than they were with single trials. The initial tests with the SPRT classifier were mixed. Classification accuracies were higher for SPRT (when using the same features) than for NB (but not for LDA or SVM) when single trials or the average of multiple trials were used for classification. The analyses of classification accuracies using features based on the N400 and FP, development of the ERP Classification GUI, and the SPRT classifier represent significant steps toward the development of a new BCI paradigm based on the processing of meaning-related information.
Issue Date:2019-01-29
Type:Text
URI:http://hdl.handle.net/2142/105128
Rights Information:Copyright 2019 James J. S. Norton
Date Available in IDEALS:2019-08-23
Date Deposited:2019-05


This item appears in the following Collection(s)

Item Statistics