UV Detector Non-Linearity

[Omar]

We are running a method linearity of a particular pharmaceutical compound (losartan) using reversed phase HPLC conditions and a UV detector. The analytical method is a gradient elution programme. The strange thing is that whn the response (area) is plotted versus the analyte concentration we are observing a significant y-intercept contribution. When analysing the data it was found that the data fits perfectly a curve rather than the linear regression. The experiment has been repeated and similar results have been obtained. See plot below. I will provide further details as you ask.

[Peter Apps]

How does it look if you delete the highest point ?

Peter

[Omar]

It actually improves a bit with each point you remove showing that the relation ship is really a curve. I've never encountered something like this so that is why I am asking here.

[Peter Apps]

Two possibilities then; the whole curve is close to the upper end of the linear range of the detector - that molecule has lots of UV absorbing structure to it - what is the LOD compared to your lowest point ?. Or (and possibly as well as) it is fluorescent and your monitoring wavelength overlaps its emission wavelengths..

Peter

[Consumer Products Guy]

That's a lot of points. In practicality, I'd likely use the midpoint area of the curve, like 100 to 400; unlikely that you need to use such a wide range.

For example, we document linearity as per USP, and concentrate on the accuracy +/- 25% of the target level. If out of that range, the results are OOS and not of that much interest "how much".

[Omar]

The absorbances are reasonable ranging from 0.1AU to 1.4AU absolute. We extended range to try to understand actual relationship. We are now looking into the possibility of fluorescence as an interference. Thanks all.

[tom jupille]

Without more data (specifically on the variability at each level), it's hard to tell whether you might be chasing a mirage. There are formal "lack of fit" tests that you can do on the linear and the quadratic fit to see if there is, in fact, a significant difference. Check these three articles by Vanetta and Coleman from their "Statistics in Analytical Chemistry" series for details:

http://www.americanlaboratory.com/913-T ... agnostics/

http://www.americanlaboratory.com/913-T ... continued/

http://www.americanlaboratory.com/913-T ... concluded/

[Sallybeetle]

Are those peak areas versus concentration? What are the units of the abscissa and ordinate?

For a UV detector, the peak areas seem enormous. Most UV detectors have a 1-absorbance unit/full scale maximum. Are you above or near this limit? I suspect you are way above the detector linearity. Have you tried a different HPLC system and had another analyst check the actual instrument set up, solution preparations, and calculations?

I understand your methodology constraints working with a USP method. In the past, I found missing dilutions, wrong wavelengths, wrong diluents, and much much more in USP methods. You might contact USP directly to see if corrections are slated for the next PF update. The monograph expert may take some time to respond, but they are usually helpful.

Also, when I had similar USP method problems in the past, I web-searched for other laboratories that actually run that test. I talked to the lab personnel to determine if they have or had similar problems. Many times they are more than willing to share their expertise.

Hope this helps. Keep us informed of your progress and any outcomes.

[Klaus I.]

It actually improves a bit with each point you remove showing that the relation ship is really a curve.

I'm not sure about this, especially the point for concentration 400 is a little bit to high for a perfect fit for the proposed polynom equation.
I have tried to extract the raw data from your plot resulting in a quite similar equation for the linear regression.

Code: Select all

10	79928,9520426
30	215808,170515
50	343694,493783
100	687388,987567
150	1015097,69094
200	1342806,39432
250	1662522,20249
300	1966252,22025
400	2629662,5222
600	3732682,06039

Code: Select all

Von x = 1,000000000000000e+01 bis x = 6,000000000000000e+02
B (y-intercept) = 5,818973365175724e+04 +/- 2,347552364405168e+04
A (slope) = 6,265046104296665e+03 +/- 8,580634161502927e+01
--------------------------------------------------------------------------------------
Chi^2/doF = 2,294888771781772e+09
R^2 = 0,998501595806494

When I remove the data for the highest concentration (with 1.4 AU peak-height?) I see a significant better R2:

Code: Select all

Von x = 1,000000000000000e+01 bis x = 4,000000000000000e+02
B (y-intercept) = 2,471072059394117e+04 +/- 6,284867809998377e+03
A (slope) = 6,524003455545055e+03 +/- 3,024975703242817e+01
--------------------------------------------------------------------------------------
Chi^2/doF = 1,297741125094361e+08
R^2 = 0,999849530629616

Without the datapoint for conc. 600 your data fits good for linear regression and y-intercept is lower.
R² seems also be good when we force the regression through zero using a function without y-intercept

Code: Select all

Von x = 1,000000000e+01 bis x = 4,000000000e+02
A = 6,618775589e+03 +/- 3,062300097e+01
--------------------------------------------------------------------------------------
Chi^2/doF = 3,643229407e+08
R^2 = 0,999517232

As Tom recommended above I would also suggest you a closer look on the variability within the respective level. Maybe it will be also a good idea to see the correlation for peak-heights instead of area.

Regards
Klaus

[lmh]

You've commented that the absorbances are 0.1 to 1.4AU; do check this, because it looks to me on your plot that they span a range much larger than 14-fold. Many UV detectors are not terribly linear up at 1.4 units of absorbance. If you want a really good linear response, you may need to stick to much lower absorbances.

[Klaus I.]

Just in the case that e.g. a micro-flowcell was used: Using a bigger flow-cell may also improve linearity.