S/N Determination using Empower Software

[Wendy38]

Hello,

We are trying to determine the signal:noise ratio using Empower software. The software allows us to choose noise intervals (closest to the peak of interest) under the suitability tab and then we use custom fields to get the signal to noise. This seems to work fine when we are looking at isocratic separations. However, we are now trying to do it with a gradient separation. The peaks of interest elute within the gradient portion of the method and hence are sitting on a fairly significant slope. When we choose noise intervals close to the peak of interest and on the slope of the baseline, we get results for signal:noise that do not seem accurate. i.e. S/N of 5 but the peak is very clear and easy to integrate. If we change the noise intervals to the beginning or end of the chromatogram where the baseline is flat but nowhere near the peak of interest, the S/N increases to about 30 for the same chromatogram. We are reluctant to change the noise interval to the flat part of the curve as it is not indicative of the noise at the peak of interest.

I would like to know if anyone else using the Empower software determination has run into the same thing with gradients and what other labs have done to determine S/N for gradients.

[aschultze]

Sorry, I can'y give you an answer to your question, but I have a question about the S/N calculation itself

The software allows us to choose noise intervals (closest to the peak of interest) under the suitability tab and then we use custom fields to get the signal to noise.

Could you elaborate on the custom field you created to calculate S/N?

[Wendy38]

Hello,

We take the height of the peak of interest/1000 divided by the noise in the region of the peak as determined by the software using the suitability tab in the processing method.

[aceto_81]

The noise calculated by empower is the maximum absorbance in the specified region - the minimum absorbance in the specified region.
So if you have a drifting baseline, you have a problem.

If we do a method validation we have to determine the LOD (S/N > 3), but with gradients, we inject 5 or more references near LOD and use the standard error of the calibration line instead of the noise as this gives us a more accurate idea of LOD as the baseline noise.

If you have to calculate S/N for routine assays with a gradient, the only way around is to set your interval at an isocratic part I guess.

Ace

[Wendy38]

Hi Ace,

Could you please expand on what you are doing with the calibration line to determine the noise of the baseline and how you are getting the LOD from this. This sounds like a better way of doing things and I'd like to have more information. I think I've seen something like what you are describing in a guidance but confess I didn't get the gist of it. So if you could be detailed that would be better.

We could set at isocratic in order to estimate what concentrations to use around LOD and LOQ.

Thanks for your reply.

[JGK]

Hi Ace,

Could you please expand on what you are doing with the calibration line to determine the noise of the baseline and how you are getting the LOD from this. This sounds like a better way of doing things and I'd like to have more information. I think I've seen something like what you are describing in a guidance but confess I didn't get the gist of it. So if you could be detailed that would be better.

We could set at isocratic in order to estimate what concentrations to use around LOD and LOQ.

Thanks for your reply.

You can find an explanation of Ace's technique in the ICH Q2(R1) document downloadable at
http://www.ich.org/cache/compo/276-254-1.html


I use the same technique, calculating a theoretical LOQ/LOD from the cal curve and then testing with actual solutions

[aceto_81]

Hi Wendy,

Here is an example with fictive data:

The data:

Code: Select all

conc (µg/ml)	area
	1		1002.132351
	2		1999.460303
	3		3000.248222
	4		3998.520506
	5		5000.376108

I will give the example in MS Excel:
Paste the above values in a new datasheet on position A1.
Now make a matrix formula (see excel help about how to do this)
=LINEST(B2:D6;A2:A6;TRUE;TRUE)

This gives you:

Code: Select all

999.5547716	1.483183363
0.41388839	1.372712496
0.999999486	1.308830011
5832391.979	3
9991097.414	5.139107994

Limit of detection = 1.372712469 (standard error of intercept) / 999.5547716 (slope) * 3.3 = 0.004532 µg/ml

LOQ = 3*LOD = 0.013595907



Good luck

Ace

[Rob Burgess]

Hi Wendy,

Welcome to the message board. What version of Empower are you using? If it is Empower 2 are you also using the noise and drift entries tab?

From what I have gethered Empower 1 was difficult for calculating reliable S/N values for grdaient separtions. Empower 2 gives you much greater freedom and can even average noise and calaculate different values e.g. EP S/N.

[Wendy38]

Hello,

Yes! We are using Empower 2 software. The noise and drift that we are calculating is under the suitability tab.

Wendy

[Wendy38]

Hello Ace,

I tried the Linest feature with some of our data but Excel is only returning a single number instead of the expected array of numbers.

Any ideas?


Wendy

[Rob Burgess]

Can you post some screen shots of what parameters you are using for your noise as well as chromatograms obtained with s/n values.

[JS]

Hi Wendy,

We had a similar problem recently with calculating S/N ratio on a gradient using Empower 2 and contacted Waters about it. First, Empower calculates baseline noise in the following way;



As stated, Empower is monitoring the maximum voltage change from the detector - as this increases with the increasing organic content in the gradient, Empower interprets it as noise with obvious consequences for your S/N ratio. You can minimise the effect by choosing a small window to calculate your noise value from.

Second, as stated by Rob Burgess upthread, Empower 2 has more noise calculation algorithms to choose from than its predecessor, some of which are more suited to calculating noise on a drifting baseline;



Please try some of these alternative methods of noise calculation and let us know how this affects your results.

Apologies to all for the large image size - the resolution is sufficiently poor that I couldn't compress it any more without the text becoming illegible.

Edit - Corrected typo.

[sdegrace]

I wanted to do similar to you and measure noise with a method using a gradient in Empower 2, so I called up Waters and got them to explain what the various noise determining algorithms actually do and which to pick.

The Baseline Noise algorithms is what is exaplined above and is only suitable for isocratic methods.

For gradient methods, the right ones to pick are Peak to Peak Noise or Average Peak to Peak Noise. You set the interval to read noise (and the interval over which to average if using averaged PTP) on the Noise and Drift tab in the processing method. Peak to Peak noise as I was given to understand finds the best linear fit through the region where noise is calculated and gives the root-mean-square of the residuals through that line. If using averaged noise, the time where noise will be monitored will be divided into chunks of the specified size, the noise measured for each, and the average noise determination will be used. I've run this by several different Waters people to try and corroborate it and so far it has checked out with everyone I've talked to.

I pick the averaged noise if the baseline has a tendency to "wander."

If you have the System Suitability package you can then get your s/n for any peak by just including the s/n field in your table or report. If you have the System Suitability tab in your Processing Method, I think that means you probably must have the System Suitability Empower package installed, and I would definitely recommend simply taking the s/n field it provides instead of making a custom field.

Detector noise is a measure of electronic noise from the detector (I understand this one less well); it always comes up with much lower noise values that don't jive with what you get just my eyeballing the chromatogram or printing it off and measuring, so I don't touch this on.

Anyway, using Peak to Peak Noise has worked very well for us.

Stephen

[aceto_81]

Hello Ace,

I tried the Linest feature with some of our data but Excel is only returning a single number instead of the expected array of numbers.

Any ideas?


Wendy

Seems that I missed a question, I'm so sorry Wendy.

But the reason why you get a single number is that you should make a matrix formula in excel or set statistics true:
select 2 columns and 5 rows and type in the first cell:
=LINEST(y-range;x-range; TRUE; TRUE)
and then, instead of pushing ENTER, push CONTROL+SHIFT+ENTER

This will give you the matrix.

Ace

[jskjei]

PF 34(5) includes the following definition of signal-to-noise ratio:

S/N = 2h/hn, where h is "the height of the peak corresponding to the component concerned" and hn is "the difference between the largest and smallest noise values observed over a distance equal to at least five times the width at the half-height of the peak and, if possible, situated equally around the peak of interest."

Does anyone plan on implementing this method of calculating S/N if the proposed change is adopted?