Please help: HPLC method validation - Limit of detection

[ultima_86]

Hi everyone, I am making method validation and having some troubles with limit of detection. I am using Shimadzu LC2030 HPLC and following guideline of ICH. Here is the quotation of the guideline:
"6.2 Based on Signal-to-Noise
This approach can only be applied to analytical procedures which exhibit baseline noise.Determination of the signal-to-noise ratio is performed by comparing measured signals from samples with known low concentrations of analyte with those of blank samples and establishing the minimum concentration at which the analyte can be reliably detected. A signal-to-noise ratio between 3 or 2:1 is generally considered acceptable for estimating the detection limit."
Please read the bold marked ones, they are related to my questions:
1. Samples with known low concentrations of analyte: how do I determine the concentration? Is it based on the concentration giving S/N = 3?
2. A signal-to-noise ratio between 3 or 2:1: The value is not confusing since that is the value I collect when I divide the Heigh of analyte to the Noise, what does 2:1 mean? Is it 3:1 or 2:1, isn't it?
3. From my understanding, I need to prepare a solution with a certain concentration which gives me the S/N ratio about 3:1. I then use the equation of LOD: Thrsehold * COncetration/ Signal to Noise,
where: thrsehold is 2 or 3
4. When I set the range for Noise calculation, someone said, the range must be either the baseline on the left or right of the main peak but must not include the main peak and the range should be about 1 minute (For example: I have the retention time of my main peak from 7.1 to 8.1 minute, and the run time is 12 minute, I then can set the range either from 6.0 to 7.0 or 8.2 to 9.2). Should I choose the area with the worst baseline or with best baseline?
Thanks in advance!

[tom jupille]

You have essentially cited all the reasons I dislike S/N for determination of LOD. I much prefer to use 3.3 times the standard error of the Y-intercept.

[per_oxid]

You wont know the noise beneth the peak. It may be different to the one beside it. When noise is low you can get huge varability in s/n between runs, measuring close to zero is often inaccurate. Agilent posted a note that they recommended rsd% of replicates as a better estimste as it will include random noise within the peak. There are differend defintions of lod so it is will deoend on what definition will suit your application. I find lod quite confusing myself and prefer loq with defined limits based on how the data will be used.

[Consumer Products Guy]

I'm retired.

I didn't like doing LOD and LOQ parts; we likely didn't do this absolutely correctly, but our QA department was fine with it. We took lab-made placebo product and spiked that with very low levels of analyte. When we got down to a level that had a peak of 3:1 (usually based upon printing the chromatogram and measuring mm height of the analyte peak compared to the noise, that was taken as the LOD.

We cGMP validated a GC procedure for ethanol in hand sanitizer product at a level about 66%, and management still had us determine LOD/LOQ there, even though LOD/LOQ was a gazillion times lower than the working concentration, and not easy to make a placebo product leaving out 66% of the formula.....

[lmh]

I get quite irritated with LODs and the ICH advice.
(1) I agree with Tom that S/N is really messy for estimation of LODs. Some instruments will have no noise anyway (accurate mass instruments may give no peep of a signal at all at the required mass; in another run they might have a random spike at some random point; if you're unlucky, the random spike might be within the window of expected retention times. What do you do with a S/N ratio that is, most days, infinite?). In any case, the S/N ratio depends on who estimates it; in Thermo-world, the Genesis and ICIS integrators give S/N values that differ by an order or magnitude or more for the same data.
(2) But even estimating from the s.d. of a calibration curve at low concentrations (low meaning around the region where you expect the limit of detection to be) isn't perfect. This approach makes far more sense for the limit of quantification, because that's the point at which the error becomes too great a proportion of the measurement, so the LOQ depends directly on the thing you're measuring, the s.d. of the curve.
(3) But with a LOD, what you're saying is "at this point, I can detect it; if I detect it, it was there, and if it was there, I'd detect it.". The s.d. approach is good if, and only if, your integrator can happily integrate a still smaller peak. Because it's based on the idea that you could measure all the way down to zero, but in order to be 99% sure the peak is not a false positive, you have to be more than X s.d.'s above zero, and yet if you injected this level multiple times, half the time you'd be below the level you just thought of, so in order to be 99% sure that given a sample at the LOD you really would detect it, you have to be 2X errors above zero.
So whatever approach you adopt, since you're promising to detect this much analyte, you really must check that your integrator will detect it.

I favour Tom's approach of basing it on the s.d. of the curve, but tempered by the characteristics of the integrator, which is after all doing the detecting.

[Fernando]

Hi ultima_86

How about the usp approach in chapter 1210? I used the S/N approach before but I turned for this method and I nave no problems.

Fernando