Clinical prediction under informative presence: Exploring what the patterns and timing of repeated observations in routinely collected data can tell us about disease risk.

In an era of increasingly large data, the use of routinely collected healthcare data for medical research is becoming more common, particularly in the area of risk prediction/prognosis research. Examples include developing risk prediction tools for cardiovascular disease using primary care data(1), and the validation of risk prediction tools for chronic kidney disease using electronic health record data from Salford(2). When using routinely collected data, we should consider the reasons why the observed data were collected in the first place. Unhealthier people are more likely to have more frequent measurements taken of disease risk factors, such as estimated glomerular filtration rate (eGFR) for chronic kidney disease and blood pressure and cholesterol for cardiovascular disease. More generally, the pattern and timing of observed measurements may contain information about the underlying health status of the individual. In the context of prediction modelling, we may be able to use this information to predict disease risk more accurately.

The aim of this project is to explore methods for incorporating information about patterns of observations into prediction modelling, utilising simulation studies and real-world data. One method is joint modelling of the observation process with the outcomes of interest, but this may not be feasible to implement in large datasets. Other methods which may be more scalable include methods for clustering patients according to their pattern of informative observation such as latent class modelling, profile regression, and more general mixture modelling and machine learning methods. The methods may be extended to include modelling multiple longitudinal risk factors and within-individual variability which differs between individuals.

The project will be supervised by Dr Jessica Barrett and Dr Brian Tom from the MRC Biostatistics Unit, in collaboration with Prof. Niels Peek, Dr Glen Martin and Dr Matthew Sperrin from the University of Manchester.