Cardiovascular disease (CVD), diseases of the heart and circulation, are the leading reason for death globally and in the UK. However cases of CVD are preventable by addressing behavioural factors or starting the uptake of statins (cholesterol lowering medication). In the UK statins are recommended if a patient's 10 year risk of developing CVD is above 10%. Accurate risk prediction for all patients is therefore very important. Statistical models such as QRISK2/3 have been developed to predict CVD risk. Such models are commonly evaluated on their ability to predict the average risk across larger groups of patients. Despite this, these models are primarily used to make risk predictions for individual patients and to decide the treatment path for a patient. There may be more uncertainty around an individual risk prediction that needs to be explored and understood. If there is sufficient variation about a patient's risk prediction it may warrant the collection of further data before prescribing statins or recommending lifestyle changes. This project aims to identify sources of variation in patient risk that are not recognised in existing models.
We will assess the magnitude of the variation in patient risks due to sources not incorporated into the QRISK2/3 models. The overall objective is to assess what range of risks a patient may have given a predicted risk from the algorithm, which is based on the average for a group. We will run simulations on a patient level, generating the potential bias and excess variation in an individual patient's risk prediction. Hypothesised sources of uncertainty include: omitted covariates, failure to account for time trends, mean imputation of covariates, geographical variation and the interoperability of the model when applying in a different setting. In each stage we will get an estimate of bias as a relative rate and an estimate of the variance. This will be done by simulating potential patient risks based on information derived from the datasets (e.g. association of the time and incidence of CVD). The effect of each source will be calculated separately and then combined at the end to produce a new risk prediction for each patient. We can then look at the range of potential risks of patients with a similar risk core in the original model.
Coronary heart disease (angina and myocardial infarction); Stroke; Transient ischaemic attacks.
Tjeerd van Staa - Chief Investigator - University of Manchester
Alexander Pate - Corresponding Applicant - University of Manchester
Darren Ashcroft - Collaborator - University of Manchester
Richard Emsley - Collaborator - King's College London (KCL)
Taher Hamid - Collaborator - University of Manchester
HES Admitted Patient Care;ONS Death Registration Data;Patient Level Townsend Score