Signal to noise ratio

[anrao_2]

One of the product developer with whom I am working is having difficulty in grasping the concept of the LOD and LOQ. He wants to know why estimations are based on area measurements whereas the signal to noise ratio is dependent on height only. I have explained the concept given in ICH guidelines also. However it is still not clear to him. Does anybody know a good reference where kind of question is addressed. Kindly advise.

[lmh]

My two standard references are ICH Q2(R1) which you are already using, and Epshtein, Pharmaceutical Chemistry Journal 38:212-225.

But I must warn you that Epshtein (a very thorough article) raises exactly the issue that is bothering your product developer. Epshtein also suggests an alternative approach in the catogory of basing LOQ on the calibration curve, using a different method that ICH haven't listed. It makes perfect sense, and may appeal to your product developer, so you may be opening a can of worms you'd prefer to keep shut.

Frankly, the ICH guidelines are a bit short in places, and some paragraphs (e.g. determination of the quantitation limit "Based on Visual Evaluation") are very ambiguous.

[Fiz]

AFAIK, a rough discussion regarding LOD and LOQ in HPLC was published in a scientific article , which simply became over the years common standard. Therefore, there is no any clear concept behind it.

[tom jupille]

Actually, the *concept* is quite clear, it's the practical implementation that has huge shortcomings. As a matter of practice ICH and USP give recipes for calculating things called "LOD" and "LOQ". You don't have to worry about what they mean, you just have to measure/calculate according to one of the approved recipes.

When you try to *understand* them, you realize that both LOD and LOQ are incomplete values, because they don't include a specification for confidence levels.

Let's take "LOD" (Limit of Detection) as an example. The concept is the amount of analyte below which you cannot be certain of detection (or conversely, above which you can be certain that some analyte is there). How certain? 99%, 95%, 90%, 50% ?? LOD can be defined in meaningful statistical terms (see, for example, the may issue of American Laboratory: Am. Lab, 41(6) 50-52 (2009) ) but the regulators prefer to avoid the issue by letting us use "shortcuts" like 3 times the noise or 3.3 times the standard error of the calibration line slope.

The same general argument applies to "LOQ" (Limit of Quantitation). The concept is the amount above which you can reliably quantitate. But, how reliably? What confidence interval do you want to use (again, 99%, 95%, . . .), how many replicates are you running? and how wide a confidence interval will you tolerate (1%, 2%, 5%, 10% . . . ). Again, this is amenable to a meaningful statistical treatment, but the regulators dodge the issue (and avoid the work!) by letting us use shortcuts.

In both cases, the regulatory recipes are almost devoid of meaningful significance; they are arbitrary: not unreasonable, but incomplete.

[HW Mueller]

No matter what you do here, it is arbitrary. Personal preference is that someone states that he sees xg of analyte at a S/N of 3, or whatever, and that he doesn´t want to give values below this. Now that is a 100% confidence that one has something there if he doesn´t have an interference in some samples. Now if one determines a confidence level % it also holds only if one doesn´t have an inadvertend interference in real samples. (Both S/N and confidence level will have been determined with some "artificial" sample). If one has a sample with no analyte but an interference at the same retention time one will be a 100% off wether he gave S/N or something other.

[lmh]

Tom (and HW Mueller), thanks for bringing up the important issue of what determines the LOQ etc. (i.e. % error you can tolerate).

For what it's worth, in case anyone is interested, working in an academic environment (luckily) I use the following approach, and encourage it amongst colleagues:

We regard LOQ's as (largely) irrelevant, because in typical studies we're interested in the population mean, not the value of a particular individual. We are therefore always working with replicate samples, and the variability we see within these (which contains sample variability and analytical error) is measured as the standard error of the mean, and used in any statistical analysis. The tolerable error depends on the size of any effect. Small effects can only be seen when errors are small, so more replicates are necessary.

LOQ's are obviously very important for those looking at single analyses, where you need to know that the likely error on this single measurement is less than X%, and you don't have any way to estimate the error from the sample. In theory, Tom's right, the LOQ should depend on why you are measuring, for exactly the same reasons as we need different numbers of replicates in different studies. If a 5% increase in some value indicates a problem, you need better measurements than you would if you only worry about 50% increases.

LOD's are important to us, because if we find nothing, the scientist carrying out the study obviously needs to know what "nothing" means. We estimate these by typical ICH methods, but do a quick believability check on our estimate. It doesn't matter if we can actually see less than we claim.

[bisnettrj2]

Does anyone here have any experience in defining LOD and LOQ in terms of the Department of Defense Quality Systems Manual version 4.1? I've gone back and forth with my QAUD regarding how to interpret the requirements in this manual, as we have an important audit coming up soon.

http://www.navylabs.navy.mil/QSM%20Version%204.1.pdf

[grzesiek]

I think you should ask more specific questions - something along this - "I have read this and think I should do that, do you think it should be done so or not?"
It will be easier to answer
There are so many kind people here but asking good questions is important if you want a good answer