https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html https://www.euro-online.org/enog/inoc2007/Papers/m https://www.euro-online.org/enog/inoc2007/Papers/mac-slots.html

SECV of animal meal model

Yen's picture
Forums: 

Dear all,

Did you build NIR model for animal meal (Meat Meal, Meat and Bone Meal, Blood Meal, Poultry Meal, Feather Meal, Fish Meal...)? How about the best SECV you can reach?

I had some models for animal meal and SECV of Protein is around 0.9-1.5%, even when I had a big data for development. I know that SECV are affected by uncertainty of wet-lab method and to be honestly, animal meal is a complex sample. 

This subject just to know about animal meal models you can have, for my reference.

Thank in advance for your information.

Best regards,

Yen 

miguelG's picture

Hello Yen,

I have no experience with animal meals in NIR, but i might give you something to think about.

Just to check, are you comparing your NIR method with the reference method using just the value of SECV? That seems not enough.

 Because firstly you should compare your results with your reference method, then you can compare results from other people.

If you know for certain that your results deviate too much from your reference method, you should reconsider the approach (preprocessing, sample preparation ...); if the results from NIR method don't deviate too much, you can be "satisfied" because you are reproducing your reference method (with the associated error). If you still want to compare your results with other people (honestly that is a good practice if done correctly), you should provide more data (range of analysis, uncertainty associated, reference method error, among others) because just the SECV is not enough.

 

DSanbornTec5's picture

Hello Yen,

Some thoughts:

(1) The standard error (as expressed by SECV, or preferrably SEP) of your model should agree with the error of your laboratory reference method. A signficantly better standard error for the NIR predicted samples verus laboratory results is cause for concern, and could indicate overfitting. This is a more likely occurance if you are working with a PC-based method like PLS, and you are retaining too many factors.
(2) Are you able to inquire about the error of the reference method, or can you furnish a sample set so that it may be determined (i.e., different aliquots of the same material submitted as different samples in a blind study)?
(3) Animal meal should be a relatively well established/profiled application. Have you searching journal articles, or contacted the vendor of your instrument to determine if they have some records that established what others have obtained for results for similar material blends/types?

Regards,

Daniel