Laure Wynants

So that means that other researchers have no way of testing in their data whether that model works.

Or if you want to apply it in hospital software, there's no way of doing it because the model is not reported anywhere.

So that's also one of the key findings.

Reporting was not good enough.

Most of the models we saw were built on data from China.

And that's a complicated factor because their admission and testing criteria and their discharge criteria are completely different than in our healthcare systems.

And in general, if you transport models from one healthcare system to the next, we know from prediction research that often won't work.

you get miscalibrated predictions.

So you overestimate the event of the probability of mortality, for example, or you underestimate it.

So

Always in such a case, you need independent validation on data from that healthcare setting and often also updating it, adjusting it to that healthcare setting.

And especially for these studies where the risk of bias were so high, that would be crucial.

I don't think any of these models can just be directly applied into practice as they are right now.