Especially for the NHS dataset, since you will either see the patient in there or in a death index (unlike, say, US where they may have just gone to another hospital).

Also, the scope here is more like a longitudinal vaccine study than a clinical trial. 50M people will provide a lot of robustness that you wouldn't see in a 1000 patient trial.

Or you won't see any new information for the patient, because the patient is lost to followup. Or - worse - you'll see new information, but it'll be invisibly erroneous, because random GPs don't work to the standard that physicians administering examinations in clinical studies do.

> 50M people will provide a lot of robustness that you wouldn't see in a 1000 patient trial

Not if the 1000-patient trial is well designed, and the data of those 50M people is totally uncontrolled and unverified. This isn't warfare - Stalin's dictum has no place here. You can't overcome the flaws of a dirty dataset by adding more dirty data to it, especially when you literally cannot know either the magnitude or the nature of the inaccuracy, or even tell what's accurate from what's not.

So you will take this into account and emphasize the more reliable types of tests, like blood tests. Or you'll find ways of learning which doctors who do the tests more accurately(or train doctors to do so) and which people are more consistent/reliable with their relationship with their docs. Or maybe you'll get a few hypotheses which are relatively likely and that would incentivize the researchers/google to do small clinical trials on them.

It's worth a try at least.