Statistical Learning and Adaptive Observation in Clinical Prediction: Methodology and Applications

Project Description: (maximum of 4,000 characters)

Clinical prediction models (CPMs), which predict the occurrence of an event of interest given what is known about an individual, could enable a personalized and preventative approach to healthcare. However, existing CPMs are usually based on cross-sectional data, thereby ignoring the rich longitudinal information in medical records. As such, exploring ways of incorporating longitudinal data into CPMs is an active area of research.

One problem is that the location/setting of where longitudinal observations are recorded (e.g. primary care or secondary care) can influence the recorded observations, which can introduce bias into the CPMs if not properly accounted for. Additionally, limited resources restrict how frequently healthcare professionals can observe longitudinal information about patients; in principle, CPMs could be used to triage adaptive periods of high- and low-frequency observation for patients, but methodological research is required before this can be done in practice.

Therefore, this PhD project will develop statistical methods to allow the derivation of CPMs that "learn" from patient-level behaviours and observation patterns. The project has initially two avenues of investigation. Firstly, we will explore if knowing how and where a risk factor has been measured leads to improved prediction of clinical outcomes (measurement error perspective). Secondly, we will look to develop probabilistic modelling frameworks to evaluate whether there are heterogeneous groups of patients who need stratified/personalised interventions. Additional areas of investigation will be data sample size, prediction paradox, informed presence.