Analytical Complexity Required for Diagnosing Bipolar Disorder, Schizophrenia, and Dementia

Author: Tyler Grummett

Grummett, Tyler, 2018 Analytical Complexity Required for Diagnosing Bipolar Disorder, Schizophrenia, and Dementia, Flinders University, College of Science and Engineering

Terms of Use: This electronic version is (or will be) made publicly available by Flinders University in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. You may use this material for uses permitted under the Copyright Act 1968. If you are the owner of any included third party copyright material and/or you believe that any material has been made available without permission of the copyright owner please contact copyright@flinders.edu.au with the details.

Abstract

Diagnostic tools, currently used in the clinical setting to aid in the diagnosis of various

neurological diseases, are often limited by subjective interpretation and the absence

of continuing neurological symptoms (e.g. seizures) required to make a diagnosis.

Consequently, these limitations may reduce the sensitivity of the investigation, the

specificity, or both. Electroencephalography researchers have attempted to reduce

this subjectivity and the reliance on neurological symptoms by utilising computer-assisted

classification algorithms that use scalp electroencephalogram (EEG) data

to make a diagnostic decision that is not reliant on neurological symptoms being

present. Measures of EEG characteristics are then obtained and compared. Most

commonly, EEG power in specific frequency bands is calculated (spectral analysis)

and diagnosis attempted. In recent years, adjacency matrices based on some form of

connectivity between EEG sensors or sources are constructed which comprise a number

of nodes (regions of the brain) and edges (connection strengths between these

nodes). Adjacency matrices may also be used to calculate graph theory measures,

such as the clustering coefficient and the shortest path length, which de ne how

all the nodes interact with each other and are able to determine whether there are

clusters of brain regions that are highly interconnected. Several recent studies have

utilised the adjacency matrices and graph theory measures in classifying neurological

diseases such as epilepsy and Alzheimer's disease and have shown them to be successful

(Magnin et al., 2009; Zhang et al., 2012; Challis et al., 2015; Khazaee et al.,

2015; Wang et al., 2015b; Hassan et al., 2015). However, there are few studies that

test whether these measures are superior to less complex analytical techniques such as spectral analysis. This distinction is important as diagnosing a disease may not

require complex analytical techniques even though such techniques might be essential

in understanding the disease. Moreover, it is beneficial to the clinical setting to

determine whether there are any cheaper alternatives to the current EEG systems,

or systems that are mobile, as a number of patients may be house-bound. In the

present study, I will be testing whether the simplest analytical methods for EEG are

on-par with some of the more complex analytical methods that have been utilised

in previous studies, and to determine whether there is a simple method that can

be implemented in the clinical setting. I will then determine if the signal quality

from cheaper/mobile EEG systems are on-par with the current top-of-the-line EEG

systems to increase diagnostic accessibility and to reduce costs in the clinical setting.

High classification accuracy was consistently found with bipolar disorder, schizophrenia,

and dementia, but not with the other diseases studied. It was also found

that, in most cases, the simpler analytical techniques yielded a higher proportion

of informative channels. This finding strengthens the possibility of utilising simpler

methods in diagnostic studies with a large number of participants, that is achieving

the goal of eventually building a diagnostic tool to be used widely in the clinical

setting. It was also found that cheaper/mobile EEG systems recorded data that was

on-par with highly sophisticated EEG systems, in terms of data quality. Together,

these findings support the prospect of having disease classification programs using

relatively simple diagnostic features utilised in mobile EEG systems that can be

brought to a location more convenient for patients.

Keywords: EEG, electroencephalography, SVM, support vector machine, connectivity, coherence, transfer entropy, graph theory, classify, classification, diagnose, bipolar disorder, schizophrenia, dementia, Alzheimer's disease

Subject: Engineering thesis

Thesis type: Doctor of Philosophy
Completed: 2018
School: College of Science and Engineering
Supervisor: Kenneth Pope