Uhhh, they trained their AI on only 18 women with diabetes? This can’t be done correctly.
Sure it can. “Do they sound fat and over 50? :If yes, answer diabetes. If no, answer no diabetes”
It’s bullshit. It’s the typical mixture of overly ambitious scientists and clickbait driven media.
Remember the 200 cures for cancer last year?
Yup, total bullshit. When I got to:
Kaufman hopes it will “transform how the medical community screens for diabetes”.
I started to lose faith that there was anything of interest there. For those who don’t know, “how the medical community screens for diabetes” currently is to…draw blood. Like, that’s literally it. You fast overnight, go to the doctor’s office, get blood taken, and the next day you learn if you’re diabetic. If your doctor is really fancy, they may do the thing where they take blood once, then ask you to drink some ungodly sickeningly sweet glucose potion and take blood a second time so they can see how your body responds. But that’s about the extent of it.
The authors are making it sound like you currently have to hike through the Himalayas to get a diagnosis now. No, you just take blood. It’s fast. It’s cheap. It’s easy. And it’s just about 100% accurate.
I can see that something like this could come up in some niche situations where someone’s very remote and it’s better than nothing, but “transform how the medical community screens for diabetes” overall is pretty laughable.
Just to play devil’s advocate here: what you’re describing is not a screening. A screening means, testing a large percentage of the population with a cheap and easy method, accepting a large amount of false positives. So _in principle _ this could be a screening test. But given the ease of the actual test, as you described, this point is kind of moot.
Sample size is relevant as a proportion of the difference you are looking for.
For example:
Sample A: 1.1, 1.3, 1.5, 1.2, 1.1
Sample B: 345.3, 323.4, 322.3, 355.2
Determining a statistical difference between these two groups where a meaningful difference is 20%, does not require more samples. The chance of error on making a claim that A is less than B will be quite low.
Not saying that N=18 in this case is sufficient, just stating that the number alone does not give you enough information to determine whether a claim has weight to it or not.
Poorly phrased in the article if I understand correctly. It means it gave a correct positive diagnosis in patients who had already been diagnosed, with those percentages.