| Author | Message | ||
     Marta Lichtig (marta) New member Username: marta Post Number: 4 Registered: 8-2009 |
Thanks for your insight Richard. I agree with you. I ordered already the ASTM you wrote about, and I hope it will help in performing the so-called "robustness" part of the validation. | ||
| Richard Kramer (kramer) Intermediate Member Username: kramer Post Number: 20 Registered: 1-2001 |
>>Typically, specificity and robustness would be the criteria needed.<< "Robustness" is a dangerous term. It is ill defined and often misused to justify applying a calibration to candidate unknown samples which are ineligible for estimation by the calibration because they were not adequately represented in the data set which was used to validate the purportedly "robust" calibration. It is all too common to assume a calibration must be robust because of the comprehensive nature of the training data set used to derive the calibration without regard to the adequacy or lack thereof of the validation data set used to validate the calibration. Another common error, even worse in some respects, is to use the "validation" data set to choose the "most robust" candidate calibration without subsequently validating that calibration with a completely independent validation data set which was NOT used to pick the winning calibration. Instead of "robustness, we should be talking about specific, proper validation (see, for example, ASTM E2617) rather than vague, ill-defined concepts such as robustness. The only meaningful way to understand and defend the "robustness" of a calibration is by adequate initial and ongoing validation. Accordingly, let's just directly reference the specifics of how a calibration is validated and how each candidate unknown sample is qualified for eligibility for estimation by the validated calibration. Those are precise and meaningful concepts whereas "robustness" is not. | ||
| Marta Lichtig (marta) New member Username: marta Post Number: 3 Registered: 8-2009 |
Thank you Sheelagh. You are perfectly right - being a qualitative method, we don't need anything but specificity and robustness. | ||
| Sheelagh Halsey (shalsey) Intermediate Member Username: shalsey Post Number: 19 Registered: 12-2006 |
Hi Marta If you are making a qualitative method, then you don't need to prove accuracy, precision and linearity. Typically, specificity and robustness would be the criteria needed. You will need to make sure that your whole spectrum/PCA final product model will pick up any out of specification product. If you are only following a standard deviation value, you need to prove that you have the right formulation mix. You will also need to validate that your good blend samples as shown by NIR also analyse for good content uniformity by your current methods. As you say, specificity is quite straightforward as you could make a placebo blend and prove that you obtained a poor match to your target spectrum. Robustness could be measured by following several blends with slightly varying component ratios (still within specification) to make sure that your qualitative model was still valid. Regards Sheelagh | ||
| Marta Lichtig (marta) New member Username: marta Post Number: 2 Registered: 8-2009 |
Thanks Sheelag for answering. My questions wasn't detailed enough to get the answer I was looking for. In fact, we have lots of blending data already, and a model for it. We have problems validating the model according to the USP/EP and other pharmaceutical requirements,i.e proving accuracy, precision and linearity of the method (specificity too, but this is the easy one). The second problem is that we don't intend to evaluate the blend by numerical results of the components.(we'll do it only as proof of concept), but as similarity of spectra. | ||
| Sheelagh Halsey (shalsey) Intermediate Member Username: shalsey Post Number: 18 Registered: 12-2006 |
Hi Marta There is a large amount of information published in the literature on the use of NIR for non-invasive measurement of pharmaceutical blends. There are a number of instruments available for this type of analysis, but most will take a measurement through a modified bin lid with a sapphire window added. They are battery powered and will take measurements triggered by the rotation of the blender when upside down. The spectrum is then sent by wireless to a remote PC. There are two schools of thought on how to measure the progress of a blend. The first is a qualitative approach where you determine a unique peak for the active and follow this by calculating a standard deviation after each rotation of the blender. Once the peak stops changing the blend should be homogeneous. This approach should be used in conjunction with a whole spectrum / principal component method to check that all components of the blend are there in the correct concentration. The second approach is to make a full quantitative calibration for the active and predict the actual concentration of the API at each rotation. This method is more onerous has you will have to make up calibration standards, usually between 80-120% strength, to make the calibration. I hope this will start you on your way, Regards Sheelagh | ||
| Marta Lichtig (marta) New member Username: marta Post Number: 1 Registered: 8-2009 |
I'm following for quite a while discussions in this forum, I got a few good ideas for our NIR work. But we face now a problem I couldn't find a good answer : validating the blending end-point by NIR. I have to add is on a currently manufactured product, so no sampling allowed during blending ! |