Consciousness is the result of an extremely complicated brain function. The exact functionality of the
brain resulting in consciousness remains unsolved. Combined forces from many different scientific
fields are working on this to get a better understanding on consciousness and its disorders. Medicine,
neuropsychology, mathematics and biology are only a few of those fields. Specifically, the medical
model can provide us with unique insights as to the functions of typical states of consciousness.
This thesis is focusing on patients with disorders of consciousness. This kind of patients are brainlesioned individuals which in numerous cases are incapable of responding to requests, despite the
fact that they might still have preserved conscious functions. Often, the remaining functionality of a
brain is sufficient for perceiving and decoding the surrounding environment or the position of patients
in it. Nowadays, we know that lack of responses do not necessarily indicate lack of consciousness.
Behavioural-assessment scales for the evaluation of consciousness often provide a vague diagnosis.
Mis-diagnosis of consciousness raises clinical as well as ethical issues.
Functional neuroimaging can be used to address this problem by providing an inner overview of the
brain functionality of patients with disorders of consciousness. Functional Magnetic Resonance Imaging and Positron Emission Tomography are two commonly used modalities of functional neuroimaging,
which are used in the projects of this thesis. They provide a quantification of different brain properties
in combination with an accurate spatial representation, which makes them a unique source of information. Machine Learning, being part of the wider Artificial Intelligence field, incorporates algorithms
that can efficiently handle high-dimensional data. Such algorithms can unveil patterns of data and undercover interactions of brain regions, using data-driven approaches. Additionally, they provide tools
that can ensure success in predicting unseen data. Therefore, they can constitute a necessary and
complementary tool to classical statistics for the analysis of Functional Neuroimaging data in Disorders
The combination of behavioural assessments and functional neuroimaging form an extremely important and unique source of information, for both clinical use and the scientific study of consciousness.
The former is showing the thin line between consciousness and un-consciousness and the latter provides the means to explore it.
This thesis aims at providing tools to assist the behavioral diagnosis of consciousness using Machine Learning in functional neuroimaging data from patients with disorders of consciousness. The
studies composing it focused mainly on the two groups that are considered to lie on the border line
of responsiveness: i) Minimally Conscious State, and ii) Unresponsive Wakefulness State. Two different modalities, which capture different properties of brain function, have been used. At first we used
functional Magnetic Resonance Imaging, from which we extracted brain connectivity features. To those
features we applied machine learning techniques to identify the contribution of brain networks to the
classification of patients. In the second project, we used the metabolic activity of the brain extracted
from Positron Emission Tomography, to classify patients with brain lesions and extract regional information. We applied certain practices, in order to overcome problems such us noisy images, redundant
features and limited samples.
Both projects are highlighting these brain regions with the maximum contribution to the classification
process, assuming that they are significant to higher order cognitive functions, therefore shedding light
on the mechanistic counterpart of the phenomenon of consciousness