How to calculate noise and detection limits

[joelyue]

I am using a dionex IC system and would like to know how to i determine the noise of the system and its detection limits.

[grzesiek]

It's probably a question about software options

I have never used dionex equipment so I cant help, but please write the name of the software for other users of this forum, and check wheter there's something in the user manual or software help about it

[unmgvar]

is it chromeleon or peak net?

[lmh]

How do you do detection limits anyway? Obviously there are various approaches. For example, you can do it on the height of peak relative to height of noise within a certain length of baseline either side of the peak, or you can do it by calculation from the standard deviation of the calibration curve. Both approaches are described in ICH Q2(R1).

The nice thing about doing it with the s.d. of the calibration curve is that if you don't feel confident about extracting the relevant numbers from your instrument software, you can export the list of integrated peaks from a calibration curve into Excel and do the regression there (with the linest function) and calculate a LOD. This approach should also work on almost any instrument.

[tom jupille]

And using the calibration curve avoids the whole issue of where and how long to measure the noise!

[orcicdejan]

If using the calibration curve approach, do not forget to check for residuals first! People often tend to accept the curve if correlation coefficient is high enough, but points with lower C may still have very high error, resulting in higher standard error and thus higher LoQ/LoD. If residuals tend to rise with concentration decrease, use weighted curve (1/C or 1/C2 should do the job, check again the residual curve)! [/i]

[grzesiek]

"If residuals tend to rise with concentration decrease" - that means using too wide concentration span

[lmh]

To clarify: there are two sorts of calibration curve here, the sort you'd use to quantify, and the sort you use to determine your limit of detection.

My take on it (and I am not a statistician, so beware!) is that when preparing a calibration curve that spans a wide range, for quantification, it makes sense to use weighting.

However, when determining a LOD, what you really want to know is the error at the LOD, and you're absolutly not interested in the error higher up. The whole concept of a standard deviation of the curve gets a bit hard to grasp if the errors are very different at different points along the curve. So for LOD, you should choose a curve with a small range close to the LOD and LOQ, over which the errors should be reasonably constant.

But I'd value input from someone who actually knows something about regression.

[grzesiek]

"when preparing a calibration curve that spans a wide range, for quantification, it makes sense to use weighting." - weighting makes sense when there is difference in variance at different concentrations, which is usually the case when a using wide range, but what is wide? here's when you use stat test to determine if variance is constant or not

[Alfred88]

Dear joelyue:
You have 2 questions. 1/ Calculate noise of system. 2/ Estimate detection limit.
With Chromeleon, you can determine noise/drift of system by using the sequence for PQ check (noise/drift sequence). The calculations are hidden in the OQ-PQ report, but you can view by exporting to Excel file. You only need to run the sequence, and the software will calculate for you.

For detection limit, it is more tricky, because you have to run samples at low levels, and check the signal/noise ratio. There are several papers on how to do it (sorry I can't list one).

Best wish.

[tom jupille]

With regard to LOD. The S/N ratio provides an estimate of LOD, but it is not the definition, and it seldom is the best way to measure LOD. Here is my preference, a CV vs A plot:

You first have to define LOD in terms of your desired reproducibility. I'm going to waaaay oversimplify the statistics involved and say that 30% (0.3) would be about right.

Then, you run sufficient replicates at various levels on your calibration plot (which you should do anyway!) to determine the CV at each level.

Now, plot CV as a function of the analyte concentration (or peak area, if you prefer). A log-log plot usually works best. Extrapolate that plot down until the y-value crosses your desired CV. That gives you your LOD.

LOD should always be verified by running replicates at that level to confirm that the reproducibility does, in fact, meet your target.

[lmh]

Seeing as the debate has got a bit deep, here's my two favourite references:
Épshtein, Pharmaceutical Chemistry Journal 38: 212-225
LC-GC Europe Feb 2009, 22:82-85

The Epshtein article describes your approach, Tom; I like it very much, but I've found it not entirely straightforward to use in practice. It needs llooooots of replicates to get a good estimate of the error at each concentration, otherwise the plot of CV versus area becomes wobbly, and the extrapolation can vary dramatically if you miss out a point (missing things out makes me unhappy, but I have a general principle that if data are good, missing out a single replicate at random should barely affect the result).

[tom jupille]

A lot of the issues around detection limits were discussed in three articles that appeared in American Laboratory in
November, 2008; March, 2009; and May, 2009

[allan101]

Hi guys,
This is Stephen, I found this forum. many information and its sollution are available in this forum. I like it. Eerybody Thanks for sharing knowledge with me.

[Newchromatographer]

Run replicate standard ( say from 0.01 ppm to 10 ppm )

calculate the standard error of the calibration curve

you have the slope then

LOD= 3*SD.ERROR/Slope
LOQ=10*SD.ERROR/SLOPE

Hope that help