Assume a model for medical diagnosis uses ``accuracy" for evaluation. Given that only 5% (ground truth) of patients have a rare disease, this evaluation metric is considered problematic.
(a) Give a general explanation of why and
(b) a specific case for the medical diagnosis model that demonstrates this.
(a) The accuracy score is high as long as 95% people who have no disease are correctly diagnosed without this rare disease.
This cannot reflect how likely a patient who has this disease is correctly diagnosed.
(b) Consider an example with 100 subjects, with 5 that have this disease and 95 do not.
Suppose all 95 subjects who do not have this disease have correct test results.
But all other 5 subjects who have this disease are all tested with wrong results.
Therefore, the overall accuracy score is 0.95.
However, within the group of subjects who have this disease, the test accuracy is 0.