Forecast Modelling For The Systematic Design Of Medical Technological Infrastructure

Abstract

Forecast models represent our understanding about how decisions and actions deliver impact in health systems. In the opening section of, forecast modelling concepts are described, framing the challenge of interest: the forecasting of outcomes with respect to some proposed action. An important step in building forecasts is determining which outcomes should be modelled. An approach for systematically eliciting sets of outcomes is described (see section 2.2). This is a qualitative outcome forecast, precursing the more data-focused methods described subsequently. This approach is combined with an approach for prioritising design and forecasting opportunities in medical technological infrastructure, adopted from the previous work of the author. This resulting focus: forecasting for proactive maintenance and service arrangement negotiation. Following a literature review into forecast modelling in these areas (chapter 4), key technical challenges are elicited. Chief among these are: - Data/Model Stratification - Current methods fail to satisfactorily account for various categorical classifications that affect device modes of operation and failure behaviour. - Data Scarcity - Current modelling methods arguably do not handle data scarcity well and many of these fail to represent the resulting uncertainty. In order to address these, a novel forecasting approach is developed and tested. It works by reconciling and combining data and models within a hierarchical taxonomy1. It extends the reliability analysis of medical devices using survival methods, within a Bayesian framework wherein the relative probabilities of potential failure characteristics are evaluated. The Bayesian approach crucially enables uncertainty to be quantified. Unlike other hierarchically-structured data analyses, this method does not rely on presupposition of model hyperparameters - meaning that it evaluates the meaningfulness of hierarchical classification structures and adapts the analysis appropriately. Furthermore, this recursive nature of the method gives it the potential to be applied more flexibly, and in multidimensional hierarchies - e.g. It could jointly consider both hierarchies of device type and use-context, enabling more personalised or context-specific forecasts. In a simulation-based study, this approach is shown to enhance the analysis of sparse data. It has also been shown to be successful in predicting the failure of high cost x-ray tubes (thus informing service arrangement negotiation). While this method shows promise, it is vulnerable is in its scalability. Its computational efficiency is largely contingent on the implementation of sophisticated particle filtering, that could limit its scope of application. Nonetheless this research project has provided insight and progress to this line of enquiry.

Date of Award	1 Aug 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Azzam Taktak (Supervisor)