[swsieber]

What do you mean by 90% accurate? Do you mean for every 1/10 people analyzed, a result is wrong? That would be a lot of false positives.

[kvejdani]

Somewhat correct. If you test the software with a 1000 MRI scans from people with known diagnoses, the software will diagnose ~900 of them correctly. Roughly 97% sensitivity (3% false negative) and 85% specificity (15% false positive). False positive here is not real false positive, as the software is detecting abnormality in people who are still cognitively normal.

[got2surf]

1/10 false positives doesn't sound too bad to me. Having seen older family members impacted by the disease with little to no warning, a 90% chance at early detection would be incredible. Even with a false positive, the changes you would make (lifestyle changes, brain stimulation, continuous monitoring) would only be beneficial, IMO.

[dragonwriter]

> 1/10 false positives doesn't sound too bad to me.

If the 15-year risk for the population taking the test is 10%, an unbiased 10% error rate would mean:

81% true negatives, 9% false positives, 9% true positives, 1% true negatives.

> Even with a false positive, the changes you would make (lifestyle changes, brain stimulation, continuous monitoring) would only be beneficial

Not if prioritizing them crowded out things you could be doing to address a real risk that is deprioritized because of the false result.

And, also, of course, an error rate doesn't have to be unbiased. With the same population, a test with these results would also be 90% accurate:

90% true negatives, 10% false negatives.