Clinical prediction models (CPMs) predict future outcomes for individuals based on their demographics and clinical risk factors [1]. Traditionally, developing a CPM is a one-time activity: it is developed once on a fixed dataset, and the model then validated and ‘deployed’, without further modification through time. However, health and care are constantly changing, and therefore predictions from such a static model can degrade quickly over time [2]. For instance, during the pandemic many models for predicting outcomes from Covid-19 became as quickly outdated as they emerged.
There is an opportunity to address this problem because infrastructure now exists to collect and analyse healthcare data in real-time [3]. Therefore, with each new patient, and new outcome, that is observed, we can retune our CPM to take account of recent changes. Furthermore, ideas from causal inference literature regarding generalisability and transportability can be used to build models that are also able to extrapolate to future times.
The methodological question is how to do this in the ‘best’ way: and this is what this PhD will tackle. The supervisory team has driven recent discussion in this area, and showcased the use of Bayesian techniques to ‘downweight’ historical data in an appropriate way [4-5]. However, further methodological work is required to optimise this ‘forgetting’ element of dynamic prediction. Some of the key areas of development are: 1) some parts of the model should adapt to recent data more quickly: e.g. the intercept may need to adapt quicker than predictor-outcome relationships (e.g., to ensure that the overall event rate is estimated correctly). 2) the ‘forgetting’ needs to take ensure adequate sample size to fit the model at any point in time – i.e. it needs to take account of recently derived sample size formulae for CPM development. 3) the ‘forgetting’ may need to be itself dynamic – e.g. forgetting historical data more quickly when the healthcare system is changing rapidly, such as in a pandemic. 4) models need to extrapolate, at least for a short period of time, into the future, as that is when they will be used. 5) we need to develop methods and standards for validating dynamic prediction models.
The work in this PhD has the potential to drive forward the methods used to develop CPMs, thereby enable the more widespread use of dynamic prediction modelling, which in turn would lead to CPMs that better support the decision-making process they aim to underpin.
Entry Requirements
Applicants must have obtained, or be about to obtain, at least an upper second class honours degree (or equivalent) in a relevant subject.
How to Apply
To be considered for this project you MUST submit a formal online application form. Please select EPSRC PhD Programme on the online application form. For information on how to apply for this project, please visit the Faculty of Biology, Medicine and Health Doctoral Academy website (https://www.bmh.manchester.ac.uk/study/research/apply/).
Applications must be submitted by the deadline, as late applications will not be considered. Incomplete applications will not be considered. Please ensure your application is complete and includes all required documentation before submission.
Applicants interested in this project should make direct contact with the Primary Supervisor to arrange to discuss the project further as soon as possible.
Equality, Diversity and Inclusion
Equality, diversity and inclusion is fundamental to the success of The University of Manchester, and is at the heart of all of our activities. The full Equality, diversity and inclusion statement can be found on the website https://www.bmh.manchester.ac.uk/study/research/apply/equality-diversity-inclusion/