Originally published at: https://tidbits.com/2019/03/19/apple-and-stanford-announce-results-of-heart-study/

Can wearables like the Apple Watch help advance medical research? The results of the massive Apple Heart Study would suggest so.

# Apple and Stanford Announce Results of Heart Study

**ace**(Adam Engst) #1

**duanewilliams**(Duane Williams) #2

The study says the algorithm has a 71% positive predictive value, while 84% of those who triggered a notification actually had aFib. Am I the only one confused by that? What does â€śpositive predictive valueâ€ť mean if not having the condition when the algorithm says you do?

**ace**(Adam Engst) #3

A good question, and one I had too. Maybe @ron has some insight into the medical-speak.

**dwtapley**(David) #4

I canâ€™t locate anything beyond the report linked to (itâ€™s not published yet in a peer-reviewed journal, I assume), so Iâ€™m just speculating here. Sometimes PPVs are corrected/normalized if the population that is sampled differs from the larger population. So the 84% would represent the â€śraw dataâ€ť from this study, but the 71% would represent a conservative estimate of the value for a more representative sample.

I have no idea if this is why the numbers differ in this instance â€“ Iâ€™m just speculating from general statistical knowledge without seeing the underlying data. It does seem a rather large correction, if this is why they differ.

**cwilcox**(Curtis Wilcox) #5

I assume â€śpositive predictive valueâ€ť would encompass identifying people who have the condition but also identifying those who do not.

**dwtapley**(David) #6

PPV is the probability that a person with a positive test result DOES actually have the condition. The negative test results are ignored in this calculation.

NPV is the probability that a person with a negative test result does NOT actually have the condition.

I am not sure why the PPV in the Apple study is 71% and not 84%, but if the investigators had reason to believe that the sample in the Apple heart study was not reflective of the population as a whole, they could have applied a correction to normalize their results to the broader population.

It could also be poor wording in the summary. Weâ€™d need to see a more detailed explanation to really know.

**jbr**(Jeremy Roussak) #7

The positive predictive value (PPV) of a test is the probability that someone who returns a positive value for the test actually has the condition.

Example: assume a condition with a prevalence in the sample population of 1 in 200. Apply a test which has a sensitivity of 99% (that is, if someone has the condition, thereâ€™s a 99% chance that the result will be positive) and a specificity of 98% (that is, a 98% chance that if someone doesnâ€™t have the condition, the test will be negative).

The result is positive. What is the chance that the person has the condition?

Answer: 20%, near as dammit.

Why? Assume a representative sample of 200 people.

One has the condition. Given a 99% sensitivity, itâ€™s close to certain that she will return a positive result.

199 donâ€™t have the condition. 2% (100-98) will return false positive results; thatâ€™s 4.

So there will be 5 positive results, of which 1 will be a true positive. The PPV of that test, *applied to that population* is 20%.

The key point is that while sensitivity and specificity are characteristics of the test, PPV (and its converse, negative predictive value) are functions of the test and the population to which itâ€™s applied. If the conditionâ€™s prevalence in the sample population were 1:1000, the PPV of the test falls to less than 5% (21 positive tests, one true).

This is why specificity is at least as important as sensitivity for a screening test.

Jeremy