Health economic evaluation: decision modelling under uncertainty.

Personalised health innovations can lead to improvements in health but consideration needs to be taken for the best use of the limited resources available in the healthcare systems when faced with uncertainty to ensure resources are shared more wisely and fairly for everyone whilst at the same time improving individual patients health outcomes and quality of life. In this thesis we develop, apply and validate mathematical, statistical and health economic methodologies that evaluate the health and economic consequences of new and existing health technologies where personalised care is feasible and desirable and where uncertainty exists in their clinical and cost-effectiveness. Health economic, mathematical and statistical methods are developed to inform parameter values in decision-analytic models which derive credible evidence on the clinical and cost-effectiveness of interventions under uncertainty.

To address the uncertainty in decision modelling in health economic evaluations, Bayesian network cost-effectiveness models are developed that synthesises the available evidence to inform parameter values in decision-analytic models comparing surgical interventions for abdominal aortic aneurysms (AAA). The results are validated against empirical evidence from the 15 year follow-up of the EAVAR-1 randomised controlled trial (RCT) which was published after the model was completed.

Randomised controlled trials are considered the gold standard [197] for statistical and health economic evaluation under uncertainty. A within trial statistical and cost-effectiveness analysis of the E-COMPARED RCT is performed to identify whether blended (face-to-face and internet based) cognitive behavioural therapy (bCBT) is non-inferior to treatment-as-usual (TAU) for major depressive disorder (MDD) and to determine the cost-effectiveness of both interventions.

For the longer term clinical and cost-effectiveness past the 12 months of the E-COMPARED RCT up to 5 years, a discrete event simulation model (DES) is developed for the cost-effectiveness analysis to address decision modelling under uncertainty. The DES model combines evidence from both the literature and individual patient data from the E-COMPARED RCT. The DES model incorporates memory into the model where individual patients are assigned attributes from the RCT and experience events and consume resources over time. An event can be the occurrence of clinical conditions such as the four R’s Remission, Recovery, Relapse and Recurrence, or progression of a disease to a new stage e.g. depression severity; resource use (e.g. admission to hospital); clinical decision (e.g., change in dose of anitdepressant); or even experiences outside of health care (e.g. failure to show up at work).

RCT results are incompletely disclosed unless both outcome rates and treatment effects across risk groups are described [210]. To address uncertainty in decision modelling in the context of personalised medicine and how a treatments effect varies across patients in the E-COMPARED RCT, machine learning methods are developed that takes account of all 278 patient characteristics measured in the E-COMPARED RCT. The thesis describes the development of a multi-class classification neural network which predicts the depressive severity (mild, moderate or severe depression or remission or recovery) of a unseen patient in the E-COMPARED RCT. The model was validated and shown to have high accuracy.