Point process survival models for epilepsy data

• Main Author: Lopez Kolkovska, Boryana Cristina
• Format: Dissertation
• Language: English
• Published: University of Warwick 2016
• Subjects: RC0321 Neuroscience. Biological psychiatry. Neuropsychiatry

The work carried out in this thesis is focused on the proposal, comparison and assessment of survival analysis models for epilepsy data. Although the Cox proportional hazards model provides a popular approach to medical recurrent events modelling, other accelerated life alternatives seem more appropriate when compared under goodness of t tests. In our research we apply the Cox proportional hazards model and two models consisting of a Poisson-Gamma mixture model that could assume the existence of a cure fraction , and which have been developed and proposed by B. Cowling[11] and J. Rogers[39] respectively. We applied these methods to the Multicentre study of early Epilepsy and Single Seizures (MESS) data set. In this epilepsy study, patients with different types of seizures were randomized to either immediate or delayed treatment, which consisted in being administered one of seven types of drugs. The aim of the study consisted in producing a prognosis with which the clinicians and patients could take an informed decision on whether or not it was preferable to take an anti-epileptic drug. We investigated the behaviour of the survival function for the Cox proportional hazards model, the joint model and the joint with cure fraction model under the epilepsy data set, under the framework of residual analysis studies, as well as empirical vs theoretical survival functions. As a final contribution of our work, we proposed modification of the accelerated life models. Since a patient cannot be diagnosed with epilepsy unless he or she presents at least two un-provoked seizures, we proposed a zero-truncated joint model, which considers the pre-randomization counts to be strictly positive. This model has been extended to consider a cure fraction of the population, but is still under development, since the corresponding parameter estimations become considerably more complicated.