Bayesian estimation of the performance of FDT, HRT & GDx-VCC for glaucoma screening (Dr Gisele Li)

posted 30May2011 from 26Jun2010 1455hrs presented at the COS Annual Meeting in Quebec City by Dr Gisele Li.

The full topic was the following: Bayesian estimation of the performance of frequency doubling perimetry, confocal scanning laser ophthalmoscopy and GDx variable corneal compensation scanning laser polarimetry for glaucoma screening in the absence of a gold standard for glaucoma diagnosis.

What to do if you don’t have a gold standard for diagnostic tests; we really don’t have a good standard - Bayesian statistics can help us. Much of the talk emphasized the methods rather than the results, in order to try to explain Bayesian statistics to the rest of us.

Data was analyzed from 500 patients already seen by optometrists and classified as high risk, mostly because of high cup to disc ratio. Normally we do a 2 x 2 table then true and false positives and true and false negatives are derived based on this. In other words, those who tested positive and have glaucoma are true positives, while those who test negative and don’t have glaucoma are true negatives. The other values are those who tested positive but do not have glaucoma (false positives) and those who test negative but do have glaucoma (false negatives.) These values are used to tell us the sensitivity and specificity of any diagnostic test, relying on the true positives and true negatives.

But, we aren’t really sure in many cases as to whether a patient has glaucoma, so how do we fit these positive and negative test results into our 2x2 table if we don’t know if the patient truly has glaucoma or not? This is where Bayesian theorem helps as tt is likelihood function of the data that we are looking at.

Looking at FDT 24-2, HRT III, GDx-VCC helps using Bayesian analysis eg if can do all three tests on each patient, you can analyze them altogether.

This led to a number of questions from the audience -

Q: Bayesian - can you explain some of the disadvantages of it

A: You need large numbers eg may need 1000 patients to show a 10% confidence but not terribly different from frequency stats. Also, as a clinician, we want to have a result to peg a diagnosis on.

Q: Can you use this for a specific patient or only to generalize about the tests

A: No, it is for populations

Q: Do we need prior data on the patients?

A; Using non-informative priors in the analysis, punching in random numbers as the prior, so don’t need to base this on previous data