> Driggs’s group trained its own model using a data set that contained a mix of scans taken when patients were lying down and standing up. Because patients scanned while lying down were more likely to be seriously ill, the AI learned wrongly to predict serious covid risk from a person’s position.
> Errors like these seem obvious in hindsight. They can also be fixed by adjusting the models, if researchers are aware of them.
I wouldn't trust any method of fixing errors in data collection. You need to redo data collection and perform a proper case-control study. You need to select your cases and controls such that the demographics look similar.
In general, your goal is not to pick cases out against a background of the general population. It's to pick cases out from an at risk group.
And if there's any difference in procedure used for data collection between cases and controls, your data is compromised. Researchers often suggest correcting the dataset (cropping out artifacts etc.). In my opinion this doesn't really work... artifacts (for example in image data) can often cause subtile global differences in illumination which ML will pick up on but not be obvious to someone inspecting the images.
The only solution when you have such artifacts (and only one that would be acceptable to me if I was doing DD, or evaluating the research) is to redo data collection in the context of a more reasonable case-control study.