Wallace, S;
(2016)
Modelling neonatal electroencephalogram time series.
Doctoral thesis , UCL (University College London).
Preview |
Text
Simon Wallace PhD Thesis April 2016.pdf Download (22MB) | Preview |
Abstract
Creating a model for brain activity is a highly complex task; this is especially true in modelling neonatal electroencephalogram (EEG) signals. Whereas previous work is motivated by improving seizure detection, this research focuses on describing the development of these complicated multivariate signals. Using data collected from inpatients at University College London Hospital at different degrees of prematurity, we propose a model for background and somatosensory response neonatal EEG sig- nals and subsequently make inferences about the observed EEG signals using this model. We construct a univariate model for neonatal EEG by analysing the second order prop- erties of these signals, taking into account time segments which have time-heterogeneous second order properties. To do so we utilise time, frequency and time-frequency domain methods. The presented univariate model is combined with a time domain correlation structure to generate a multivariate representation which is possible, in part, due to the resolution of the data. Furthermore, the parameters and signal components are best described by taking into account not only the age at which testing occurred, but also the age at which an infant was born. This research has attempted to create a model that is not only descriptive of somatosensory responses, but also applicable in other avenues of similar research. We propose to use generalised linear models to describe the age dependence of the ob- served time series, and use these models to simulate EEG observations. When modelling characteristics of the estimated parameters, all models require the age pairing - age at birth and age at test - as variables. Combined with an appropriate time domain corre- lation structure, this allows us to achieve suitable estimates of observed signal structure. The model class presented is a flexible and accurate representation of neonatal back- ground and somatosensory response electroencephalogram signals, and can be used to describe similar multivariate observations.
Archive Staff Only
View Item |