Recommended Criteria for Accuracy

[jayroseville58]

As per ICH, Accuracy is being demonstrated by performing 3 replicates each of three separate sample concentrations.
Question:

Let say, I have an Assay Specification of 80% -120%, What would be the recommended criteria for its percent recovery (on Accuracy test). Does ICH has clear guideline on this?
Can I set it as 95% - 105% percent recovery?

Thanks.

[gtma]

That seems reasonable to me given your spec is relatively wide. The tightest I would set the accuracy limit is 97.0% to 103.0% for such a wide specification. Please made sure you take into consideration the potency of the drug and other variabilities e.g. other analytical method variabilities, potency loss on stability, and manufacturing variability.

[Blazer]

ICH does not have a recommendation for accuracy recovery. I would let your prevalidation data drive your acceptance criterion for your protocol.

[lmh]

The two values interact. Logically, your product specification needs to be aware of your accuracy of measurement. The thinking should be like this:

(1) For the intended application of this product, a variation more than (for example) +/- 20% would be bad.
(2) Include a general safety-factor if desired. Maybe if 20% deviation has seriously bad consequences, we would prefer to reject product that deviates by more than 10%.
(3) If we think the analytical precision and accuracy allow a 5% error, then if we specify +/- 10%, a measurement of 10% could actually be as bad as 15%, so instead we should insist on a smaller analytical error-window such that analysis still guarantees that the product is within the safety-window we specified. Therefore if we want to guarantee +/- 10%, we can specify our product as +/- 10% if our analysis is perfect, or perhaps +/- 5% if it's not...

What accuracy and precision you get is going to depend on the instrument, the analyte, the level at which you're analysing etc. etc.; it has to be good enough for purpose, but for once I agree with ICH in not specifying a rigid number.