Developing Clinical Prediction Models: when and how to utilise existing research

Abstract:

Clinical prediction models (CPMs) aim to predict the presence (diagnostic prediction) or future occurrence (prognostic prediction) of an event of interest, and are predominately derived in a single dataset by estimating the associations between the outcome of interest and multiple risk factors (covariates). Such CPMs are increasingly developed to support local healthcare decision-making.

Although relevant CPMs may already exist, each derived for similar outcomes but in distinct populations, their predictive performance frequently drops when they are applied in a new population to that in which the model was developed. The mainstream strategy to handle this situation is to reject the existing CPMs and re-develop a model de novo. However, this approach neglects the existing research, leads to many CPMs for the same prediction task, and is susceptible to over-fitting.

“In this talk, I will give an overview of methodological work from my PhD, which aimed to improve the efficiency and workflow of developing CPMs. By aggregating multiple (relevant) CPMs that already exist, I aim to highlight how we can move away from the current practice of iteratively re-developing a new model each time new data becomes available.”