| Author | Message | ||
     Stanislav Nikolov (stanislavnikolov) New member Username: stanislavnikolov Post Number: 3 Registered: 4-2013 |
Hi David. Thanks for the help and for the warm welcoming. In the software after internal validation of model I can choose the value of the threshold without any problems.The default value which the software suggest is 0.95.Ergo it should not be problem to change it to higher threshold value and then the false positive validation of Indometacin with Acemetacin will give me the much desired Unidentified result. And for last, but not the least-yes, I'm aware that result 0.96 is far much worse than result 0.999, but thanks again for the explanation. I'm really glad for finding that forum. Ii is very useful. Thanks again and best regards to all of you. | ||
| David Semmes (dsemmes) Member Username: dsemmes Post Number: 15 Registered: 6-2010 |
Hi Stanislav, welcome to the group. I�m not familiar with the details of the Microphazir implementation of the kNN algorithm, but I think I can explain your results. If the correlation coefficient of the Acemetacin validation sample is 0.96, and the Indometacin library samples gave 0.999, then the Acemetacin would be correctly rejected with any Pass/Fail threshold between 0.96 and 0.999. 0.96 may seem too close to 0.999, but the experimental variability of matching samples is likely far smaller, on the order of 0.001, not 0.040. Also, since Acemetacin and Indometacin have similar structures, you would expect that sample to be relatively close in match score. I suggest you find how the threshold is set in the software, and whether you can change it. Sometimes it is set to 0.95 and cannot be changed, which in effect means the selectivity of the ID test will not depend on your requirements or the measurement or lot variability. Instead, the selectivity will likely be limited to materials that are more different than Acemetacin and Indometacin. In essence, your validation experiment would be illustrating the selectivity of the ID test using that software. Good luck! David | ||
| Ian Michael (admin) Board Administrator Username: admin Post Number: 40 Registered: 1-2006 |
"NIR qualitative model validation" can now be found at http://www.impublications.com/cgi-bin/discus/show.cgi?tpc=5&post=6150#POST6150 | ||
| Stanislav Nikolov (stanislavnikolov) New member Username: stanislavnikolov Post Number: 1 Registered: 4-2013 |
Hi to all.Let me first introduce myself. I'm from Bulgaria, and currently I'm working as QC chemist at Sopharma. Here at our company we have one NIR apparatus, Microphazir RX, with optical range from 1600 to 2400nm.At this link all of you may see it: http://www.ahurascientific.com/material-verification/products/microphazirrx/index.php The main purpose of this NIR is to use it for qualification of each drum with substance before manufacturing process. So in that order I try do develop a working qualitative model for recognition of some substances. In the applied software for modeling with the device (Polychromix Method Generator) I have options to use kNN or PCA algorithm. I manage to use kNN, since that algorithm is simple, and in return I receive Correlation Coefficient to the builded library. The library itself contains for the time being, five kind of substances (EMEA guidelines). These five substances are validated at CC=0.999 and for now everything is ok. Today I received an ordered substance for false positive recognition about Indometacin. I have ordered Acemetacin, since his chemical structure is close to the Indomeacin. And here comes the problem. When I'm scanning the probe the result is Indometacin with CC=0.96, which bothers me very much. Why? Because according to EMEA guideline the false positive probe should give Unidentified result. I will be glad if someone helps me someway by any kind of suggestion, how to make it work and not recognizing it. The parameters that I use for modeling are: Normalize Range-1600/2223nm S.Golay-1600/2223nm Derivative:1 Points:13(Maximum allowed by the software) Degree:2 kNN:3 Thanks in advance for all of you. | ||
| Stanislav Nikolov (stanislavnikolov) New member Username: stanislavnikolov Post Number: 2 Registered: 4-2013 |
I'm really sorry about this topic which is in the wrong category.Please, someone of the moderators to move it to the appropriate section. Thanks in advance. |