Prognostic factors for epilepsy

Abstract

Introduction and Aims: Epilepsy is a neurological disorder and is a heterogeneous condition both in terms of cause and prognosis. Prognostic factors identify patients at varying degrees of risk for specific outcomes which facilitates treatment choice and aids patient counselling. Few prognostic models based on prospective cohorts or randomised controlled trial data have been published in epilepsy. Patients with epilepsy can be loosely categorised as having had a first seizure, being newly diagnosed with epilepsy, having established epilepsy or frequent unremitting seizures despite optimum treatment. This thesis concerns modelling prognostic factors for these patient groups, for outcomes including seizure recurrence, seizure remission and treatment failure. Methods: Methods for modelling prognostic factors are discussed and applied to several examples including eligibility to drive following a first seizure and following withdrawal of treatment after a period of remission from seizures. Internal and external model validation techniques are reviewed. The latter is investigated further in a simulation study, the results of which are demonstrated in a motivating example. Mixture modelling is introduced and assessed to better predict whether a patient would achieve remission from seizures immediately, at a later time point, or whether they may never achieve remission. Results: Multivariable models identified a number of significant factors. Future risk of a seizure was therefore obtained for various patient subgroups. The models identified that the chance of a second seizure was below the risk threshold for driving, set by the DVLA, after six months, and the risk of a seizure following treatment withdrawal after a period of remission from seizures was below the risk threshold after three months. Selected models were found to be internally valid and the simulation study indicated that concordance and a variety of imputation methods for handling covariates missing from the validation dataset were useful approaches for external validation of prognostic models. Assessing these methods for a selected model indicated that the model was valid in independent datasets. Mixture modelling techniques begin to show an improved prognostic model for the frequently reported outcome time to 12-month remission. Conclusions: The models described within this thesis can be used to predict outcome for patients with first seizures or epilepsy aiding individual patient risk stratification and the design and analysis of future epilepsy trials. Prognostic models are not commonly externally validated. A method of external validation in the presence of a missing covariate has been proposed and may facilitate validation of prognostic models making the evidence base more transparent and reliable and instil confidence in any significant findings.