| Author | Message | ||
     julen arruabarrena (julen) New member Username: julen Post Number: 4 Registered: 11-2010 |
Hi David, Thank you for your interest. The dosage form i'm analysing are tablets as you guessed. The tablets are made up of the API and 5 excipients. I made the doped samples with 2 of the excipients (i left the minor concentration ones out of the experimental design). Hence, the correlation coefficients among API and this two excipients have a value of 0,495 and 0,19. The SEC and RMSEP value of the non-OSC model are 0,021 and 0,066 (mass percentage units) for a 7 factor model in the 1100-1300nm spectral range. The OSC model uses the same spectral, 1 OSC factor and 4 PLS factors. The SEC value is 0,015 (mass percentage units as before). When prediction samples are included in the OSC model calculation, the RMSEP is 0,015 (SEC's same value), but when prediction samples' transformation is done with calibration samples' model, there is no RMSEP, the prediction value for all the samples is the same. I expect this long post could be understood, in spite of my English skill, Regards, Julen Arruabarrena | ||
| David W. Hopkins (dhopkins) Senior Member Username: dhopkins Post Number: 178 Registered: 10-2002 |
Hello Julen, I'd like to ask some more questions. What kind of dosage form are you using? Tablets? How many excipients are in the formula? How did you make the doped samples? If you mix x grams of API with sum-x g of the same excipient mixture, the experimental design may be so correlated that you cannot expect to use many factors in your PLS or PCR models. What were the SEC values for the calibration samples and RMSEP values for the validation samples for the PLS method and for your PLS with OSC pretreatment, and how many factors did you select for your models? Did you also try PCR, and how did it perform? Best wishes, Dave | ||
| julen arruabarrena (julen) New member Username: julen Post Number: 3 Registered: 11-2010 |
Hi, First, i want to say thank you for every answer, i didn't expect my question would be even answered! Answering to Katherine Bakeev's post, i have tried "classical" preatreatments, but I am trying to improve my results Answering to Johan Trygg's, i would try the O-PLS method, i have already installed the SIMCA-P 12 in my computer. Answering to Jerry Jin's questions, i don't have choose the calibration and validation sets randomly since each set must have doped and production samples, but when inversing calibration and validation sets the calibration is as good as before. I have applied OSC to the whole data set, but i think is not a good approach since Y-data is needed (validation samples' too) and this is the aim of the model: predict Y-data. As I said before: thanks to everyone. Regards, Julen Arruabarrena | ||
| Jerry Jin (jcg2000) Senior Member Username: jcg2000 Post Number: 37 Registered: 1-2009 |
Hi, Julen: If you can get a good model from your calibration data set following a OSC preprocessing, you should be able to apply it to a validation data set, as long as that the validation data are collected under the same condition as the calibration data set. Did you randomly choose your calibration data set and validation data set? Did you apply OSC to all the data you used before you build a model and validae it? Cheers! Jerry Jin | ||
| julen arruabarrena (julen) New member Username: julen Post Number: 2 Registered: 11-2010 |
Hi, First, thank you Gabi for answering my post. 1. the range is +/-25% over the nominal value 2. I think the results are acceptably good. I have a relative standard deviation of 0,5% on replicates of the sample (it's a HPLC method; replicate understood as the determination of different sample portions, making the sample preparation for each replicate). 3. I have 20 samples in the calibration set (12 doped samples to span the concentracion range and 8 production samples). The validation set is make up of 7 doped samples and 9 production samples. I expect this information is that you were asking for, Julen Arruabarrena | ||
| Johan Trygg (trygg) New member Username: trygg Post Number: 5 Registered: 1-2006 |
Hi Julen, Firstly, this is an interesting question since OSC preprocessing method is different in different software packages. Svante Wold published the original OSC paper back in 1998 and it has also been implemented in Umetrics software SIMCA-P. Since then, a plethora of preprocesing methods followed from 1999-2003, all different from the original Wold OSC method and some don�t even qualify as an OSC method. Unfortunately, when implemented in other software packages, these other methods were also named OSC, but having different algorithms and implementations compared to the Wold OSC method, this is confusing and also relates to your experience here (e.g. see Unscrambler, PLS toolbox). So try the Umetrics version, you can find a demo download at www.umetrics.com. Secondly, the prediction performance of an external testset after OSC preprocessing does not improve, this is also a misunderstanding since the "early days" in the development of these methods. Thirdly, actually I would advice against using these methods as they have an inherent overfitting problem, and require two separate modelling steps. Instead use the OPLS method (Trygg 2002). The OPLS method (also implemented in SIMCA-P 12) is a PLS method with an integral OSC filter. This means you fit one model (not two separate steps). With OPLS you get similar predictive capabilities as with PLS and in addition improved model interpretation and model diagnostics. There are many hundreds of publications on OPLS applications, but the methodology is described in below papers, Trygg J, Wold S. Orthogonal projections to latent structures, O-PLS. JOURNAL OF CHEMOMETRICS, 2002; 16: 119-128. Bylesj� B, Rantalainen M, Cloarec O, Nicholson JK, Holmes E, Trygg J, OPLS Discriminant Analysis, Combining the strengths of PLS-DA and SIMCA classification, JOURNAL OF CHEMOMETRICS, 2006, 20 (8-10), 341-351. regards, Johan Trygg Computational Life Science Cluster (CLiC) Department of Chemistry/Chemometrics Ume� University, Sweden | ||
| Katherine Bakeev (katherineb) New member Username: katherineb Post Number: 5 Registered: 12-2009 |
Julen, OSC may not be the best transformation for data, as it removes X data that is not correlated with Y in the calibration set, so can give an improved calibration model. It is important to use validation to ensure that the OSC has resulted in a better model. I have found that NIR calibrations for low levels of drug (I have one with 0-1 wt%) can work quite well when using an SNV or a derivative pretreatment. I have also seen cases where OSC works well - but one must be cautious that valuable information is not lost in applying any transformation. Regards, Katherine Bakeev | ||
| Gabi Levin (gabiruth) Senior Member Username: gabiruth Post Number: 46 Registered: 5-2009 |
Hi Julen, This is a case that calls for few questions: 1. What is the range over which you have samples? +/- 10% or +/-15% of the nominal 1.6? 2. What is the uncertainty in the reference values or as may be referred to the accuracy of the reference method - in other words, if you run same sample 10 times, what will be the spread of results. 3. How many samples in you calibration set? Based on that, it may be a little easier to try to understand and respond. Gabi | ||
| julen arruabarrena (julen) New member Username: julen Post Number: 1 Registered: 11-2010 |
Hello everyone, I'm trying to determine the API concentration of a pharmaceutical product at low concentration (1,6% w/w). Due to the low concentration, I thought OSC would be a good data pretreatment. I'm using Unscrambler X to do the OSC preatreatment. I create the OSC model with the calibration samples (the PLS model created with this samples is really good) and then I apply this model to validation samples, but this pretreatment makes all validation samples' spectra very similar, obtaining very similar prediction values (API concentration) for all validation samples (obviously far from reference value). Does anyone have any idea about what i'm doing wrong? Application of OSC itself? Could be a software problem? Thanks, Julen Arruabarrena |