that doesn't bother me but what is actually suspicious is persistently only mentioning "accuracy" but not sensitivity / specificity / precision etc.

Basically, almost everybody doesn't have Alzheimers. Sampling from the general population you can get better than 94.5% accuracy just by returning negative on every test. You have to know sensitivity and specificity separately to make any informed judgement ... which they try extremely hard not to tell you.

> Compared with the final diagnosis, the pre-biomarker diagnosis was maintained in 71/200 cases (75.5%) (Kappa = 0.576), while the post-biomarker diagnosis was maintained in 189/200 cases (94.5%) (Kappa = 0.906).

Even without this method, the doctors have been able to give diagnosis with 75.5% accuracy (according to the paper's claim).

It says "accuracy". For the overall assessment.

I.e. it needs the original 75% accuracy or so and boosts it another 20%.

The problem is that the assessment itself is slow, expensive and requires skill.

What we really want from a test is high specificity (a positive test means you have accuracy) and high sensitivity (if you get a negative test you don't have it).

This is how we can offer screening.

> 94.5% in itself is suspicious

No it's not, that's a reported mean, presumably with the right number of significant digits.

If you want to criticize the variance/stddev, do so, but you picked the wrong metric if that's what you wanted to complain about.