Hacker News Reader

Dumb question: does "90% accuracy" mean - "of the cancers it detected, it was 90% correct"?

Or does it mean, "of all the patients who developed cancer, it detected 90% of them"?

Where the later is referring to the entire population, and the former is only about those detected.

I really like the table of confusion as a way to look at this stuff. I find the terminology somewhat confusing and it's much easier for me to just look up the formulas.

That means 90% accuracy should mean (true_positive + true_negative)/(whole_population) == 90%.

https://en.wikipedia.org/wiki/Confusion_matrix

So a trivial test that always says "No" will have better accuracy?

https://www.cancer.org/cancer/types/prostate-cancer/about/ke...

> About 1 in 8 men will get diagnosed with that cancer, through their whole lifetime.

So true_negative/whole_population>=87.5% at least (life time vs currently has cancer).

Or am I misunderstanding the meaning?

This is why accuracy is useless for reasoning about detection of rare things.

Beside the already linked-to article on the confusion matrix, you can also have a look at the one on sensitivity and specificity:

https://en.m.wikipedia.org/wiki/Sensitivity_and_specificity

Reply on news.ycombinator.com