Prediction models in healthcare: a playground for researchers
Clinical prediction models estimate an individual’s risk of a particular health outcome. Well-known examples include QRISK for estimating 10-year CVD risk and the Nottingham Prognostic Index for estimating 5-year survival probability. Thousands of prediction models are published each year in the healthcare literature, yet very few are reliable or fit for purpose. In this talk, I suggest that prediction model research has become the academic’s playground, and that machine learning and AI have only exacerbated this problem. In particular, I discuss the findings of our living review of COVID-19 prediction models, and highlight common pitfalls such as small sample sizes, model instability, and no assessment of model calibration or clinical utility. I then focus on methods and initiatives to improve current standards, include new sample size calculations for model development and validation, and the TRIPOD reporting guidelines.