Dr Matthew Sperrin

Research interests

Matthew develops methods for clinical prediction modelling, with a focus on the following areas:

Prediction under hypothetical intervention. Prediction models are ill-suited to be used for decision support directly, because they do not inform about the consequences of taking actions. This research stream addresses this through the incorporation of causal inference methods in prediction modelling.

Missing data and informative observation. Prediction models call for special consideration of missing data - in particular because missing data (and the observation process) can be informative for prediction, but it is challenging to exploit this given that missing mechanisms are likely to change when models are deployed.

Adaptive sampling. Making predictions is not a one time activity, and prediction models do not tell us what should be measured, and when. This work aims to optimise when predictions are made, and using what information.

Dynamic modelling and updating. Prediction models should be updated over time and space as their deployment environment changes, but this raises questions about how and when models should be updated.