Bayesian approaches to personalise prognostic modelling.

Prognostic models are tools to predict whether a person will have an event (e.g. a heart attack) at some point in the future based on what we know about them now. Building such models is an area of substantial interest throughout the healthcare system since these models have the potential to support both evidence-based decision making and the drive for precision medicine. There are, however, significant methodological challenges that require a statistically sophisticated approach: these models typically been based on simple modelling approaches, such as logistic regression or Cox regression, and use only a small number of risk factors measured at a single point in time. Additional contextual information, such as current (and past) treatments, risk factor trajectories (e.g. weight over time), and degree of contact with the health service, are not fully utilised. Using such information could help to produce more tailored predictions for each individual.
Some statistical methods to incorporate such data are available, for example through jointly modelling longitudinal and survival data in a classical statistical framework [3]. An alternative is the use of machine learning or computationally intensive Bayesian approaches [4]. These will be implemented and extended during the PhD, for example in developing models to handle the scenario where more is known about some patients than others (e.g. blood tests that are available only for a potentially biased subset of the population)
This project has two aims:
1) To review, and potentially extend, methodology for more advanced prognostic modelling that can be more personalised to the individual. This could include: (i) exploiting longitudinal data at the patient level using e.g. the methods of or Gaussian processes; (ii) dealing with the case where more data is available for some individuals than others, (iii) considering the availability of a very large number of potential predictors (technically speaking, p>n) using techniques such as
penalised regression.
2) To demonstrate the potential benefits of more sophisticated modelling techniques in increasing model performance and validity - through predictions that are more tailored to, and useful for, individuals.