Read on Twitter
Unpopular opinion: in any prediction model that has clinical use/improves patient outcomes, continuous updating will either be infeasible or unethical.
ok so a tweet is not a place for an elaborate discussion but what I meant to say is this. Yes, time trends can make a model invalid (eg predict IVF succes, IVF techniques have improved considerably last few decades). Yes, give me a dataset and I am able to re-estimate the
model so that risks are calibrated again. But if the model has a clear clinical benefit, it likely alters risks (treat high risk patients to lower risk), changes on whom outcomes are observed (only invasive/expensive diagnostics in high risk) or even alters the target population
(learning effects of the model.) It becomes a philosophical question (what do you want to estimate, in who, etc). Simply reesimating the model on available patients/outcomes usually won't do. You'd have to set up elaborate studies and rely on assumptions, etc. Techniques to deal
with these questions are in their infancy. I see the term continuous updating used as a magic solution for postmarket surveillance a lot, but I think we should look into this more before we decide this should be done routinely.
