# Why we use precision/recall in binary classification but sensitivity(=recall)/specificity in medicine?

Sensitivity=recall is used in both fields, but the second metric is different. Why? Both tasks (classification and medicine) look same - data has two classes and we do some predictions on it and want to assess the accuracy.

• Can you point us to an example about specificity being used in medicine, to be completely in line with what you're asking? Commented Oct 19, 2020 at 8:39
• en.wikipedia.org/wiki/Sensitivity_and_specificity Commented Oct 19, 2020 at 8:56
• I have been trying to figure this out for the last decade. Hope the question gets answered. Commented Oct 24, 2020 at 21:11
• Of possible interest
– Dave
Commented Aug 4, 2021 at 21:05

Sensitivity and specificity have the attractive property that they do not depend on class prevalence - sensitivity is the accuracy among the real positives, while specificity is the accuracy among the real negatives. Since the real positives and real negatives are treated completely separately by these metrics, their relative proportions do not matter. In the medical field, sensitivity/specificity are characteristics of a particular test, completely independent of how many people have the condition being tested for. This makes the statistics of the test invariant in time and place - a test applied to a population with 90% incidence of the condition will have the same sensitivity and specificity when applied to a population with 10% incidence.

Precision, on the other hand, does depend on class prevalence - it is the accuracy among predicted positives, but how many people you predict positive will depend on the prevalence of the condition. If you apply the test to a population with 90% incidence of the condition, you will get one precision value. But when you apply that same test to a population with only 10% incidence, the much larger number of real negatives increases the number of false positives relative to the number of true positives, so the precision will be much lower. As the real positive population diminishes, so too does the probability that a positive test is correct (precision).

Specificity is a characteristic of a test independent of the population it's being applied to, while precision is a characteristic of the test in the specific population it's being applied to. Since condition prevalence can vary by geographic location, or subpopulation, or over time, specificity is usually the preferred means of describing a medical test. As a condition prevalence goes down, the precision of a fixed test goes down, but its specificity does not.

All of these measures are terribly defective in the context of medical research and medical decision making. See this which discusses severe issues with sensitivity and specificity, and how to get around these issues by using only forward-information-and-time probabilities. The latter concept is discussed here which details the severe problems caused by the use of backwards probability, i.e., probabilities that condition on unknowns to estimate the probability of events that have already happened.

I have recently published an article in the Journal of Classification that deals with exactly this question. Take a look at https://rdcu.be/dL1wK In short, it depends on the circumstances you are in.