Title: Statistical methods for dynamic cardiovascular risk prediction
Authors: Angela Wood - University of Cambridge (United Kingdom)
Jessica Barrett - MRC Biostatistics Unit (United Kingdom) [presenting]
Michael Sweeting - University of Leicester (United Kingdom)
Ellie Paige - Australian National University (Australia)
David Stevens - University of Cambridge (United Kingdom)
Abstract: A risk prediction model aims to accurately predict the probability of some event occurring within a pre-specified time window for a new individual. A dynamic risk prediction model allows risk predictions to be updated over time in response to new information becoming available. Methods for dynamic risk prediction include (i) using the last-observation-carried forward (LOCF) of each risk factor as a time-varying covariate in a time-to-event model, (ii) landmarking, where a discrete set of landmark times is specified at which risk predictions are to be made and survival is modelled from the landmark time only for individuals still at risk, and (iii) joint modelling, where repeated risk factor measurements and the time to event are modelled simultaneously, e.g., for cardiovascular disease (CVD), the 10-year risk of a CVD event is typically used to make clinical decisions about whether to prescribe lipid-lowering medication. Time-varying CVD risk factors, such as blood pressure, cholesterol and smoking status, may be monitored over time and used to dynamically update CVD risk predictions. Dynamic risk prediction models for CVD will be compared in different data scenarios, including a single cohort study, an individual participant data meta-analysis and UK electronic health records.