[diffeomorphism]

No, that does not seem to be what they are saying.

> We evaluated the diagnostic power of the device in a cohort of 45 LC patients and 14 healthy pediatric donors. We estimated a 94% accuracy for the microclot count using the devices, significantly higher than the traditional counting of microclots on slides (66% accuracy).

They are comparing the predictive power and using accuracy (instead of sensitivity, recall, F1, etc.). For their method "using the devices", they compute an accuracy of the predictive power, not of the count, of 94%. For the previous method they say the accuracy is 66%.

Basic questions: Is accuracy even a good metric for this? Is 94% a good value or just the difference between bad and very bad?

It might very well be that their improvement is from bad to really good, but the point is that a raw stat of "94% accuracy" is useless without context and so is the headline.

[resoluteteeth]

OK, I looked at the actual paper, and what 94% actually is is the 0.94 area under the curve for the receiver-operating characteristic curve (the plot of the true positive rate (TPR) against the false positive rate (FPR) at each threshold setting) not the accuracy for a specific binary result (e.g. at a specific arbitrary threshold).

See https://www.sciencedirect.com/science/article/pii/S155608641...

> In general, an AUC of 0.5 suggests no discrimination (i.e., ability to diagnose patients with and without the disease or condition based on the test), 0.7 to 0.8 is considered acceptable, 0.8 to 0.9 is considered excellent, and more than 0.9 is considered outstanding

So .94 is actually extremely good.