| Author | Message | ||
     [email protected] (Functionalgroup) |
to everyone who has given me their valuable input thank you. I will continue to work on the project and judge my results for myself. i will be more than happy to share the results when i am done with all interested parties. I apologize for not being able to be more clear about what exactly i am working with. I would also like to state that the feasibility is just that. From the results of my experiment, i will suggest to my supervisor which direction to go. You all have made yourselves overwhelmingly clear that I need reference testing and/or reputable consulting services. I agree. with that said. Thanks and Thanks. | ||
| Kathryn Lee |
Aquil, I would like to respond to your comment about consultants, "For example, say i hire some consultant, and that consultant tells me everything i've already learned for the past 2 years, for a total of 1500 per day? That would be even more wasteful than spending a couple of weeks exploring the capabilities of the machine and what exactly can and cannot be done using the resources available to me." First, hire a really good consultant that has been in the NIR business for over 20 years, not "some" consultant. That means the consultant has had 10 times the time to learn than you have. That also should mean that the consultant is a spectroscopist and has spent 4 years in graduate level studies, which may be 4 more than you have. Then there is the cost of $1500 per day that you are assuming. You took two years to learn all you know. So if you spent 10% of your time learning NIR, and make $50,000 to $100,000 per year, that means that the consultant costs about 0.75 to 1.5% of what it cost to have you learn. Or, if you want to spend 2 weeks on the project and make about $50,000 per year, then you have cost the company about $2000, which is more than your hypothetical $1500 consultant fee per day. A consultant may not be able to do in one day what you can do in 2 weeks, but the consultant probably can prevent you from wasting time and set you on the right track from the start, and save significant time and materials costs. But most importantly, we have all seen huge problems in NIR calibrations in labs that are not obvious to people relatively new to NIR. These run the gamut from taking a background only once and using it forever, to using a reflectance background on a transmission experiment, to libraries that can't distinguish sodium lauryl sulfate from magnesium stearate. Other problems we have seen are statisticians who know statistics, but not the specialized statistics needed for NIR. I hope your auditors have documented experience with NIR if you are relying on them to check your work. Otherwise, they may not be able to help you as much as you think. Yet your company has no problem paying for them. We all have heard of companies having great trouble with NIR after a consultant of questionable quality has worked with them, too. So hiring the right consultant will also be a challenge. In the end, a good consultant will be worth every penny, even if they come in and tell you that you have been doing everything perfectly, and there is no place whatsoever that you can improve. At least you will have confirmation and a really good "audit." I do realize the problem with having a consultant come in and possibly saying you could improve. It could make you look bad. However, I hope your boss has enough quality to know that by having a consultant come in, that you are actually helping the company deliver quality and that you will become more efficient in developing future NIR applications, which we all know can save companies hundreds of thousands of dollars. It will have high payback for your company. A phrase that I have heard is that 6 months of lab work will save you a good hour of research in the library. The same idea works for consultants. You have already saved your company big bucks by getting this excellent and free consulting, but I think we all realize there is still more that a specialist can do to help you with this and your other projects. We all love feasibility. Feasibility starts with scanning one sample to see if there is a NIR spectrum of ok signal to noise. It then moves to a second sample of same or different content to see if either that first spectrum is reproducible or the variation in the second sample can be detected. The next several samples to be analyzed are then pretty easy to pick what to vary, and don't cost too much to make, but then suddenly it all gets very complicated on its way to a viable calibration, and the experimentation itself gets complicated by cost/time/feasibility justifications. So by all means do feasibility, but at some point, step back and do the cost/time/feasibility/capabilities justifications to make sure you are on the right path. We all wish you the best. | ||
| hlmark |
Well, yes, there is one more thing, I think: if you're worrying about the effect up putting large pieces of sample material into a cup (or vial) then apparently nobody (including your external auditors) has introduced you to the idea of grinding the sample before measurement. Grinding is arguably the oldest sample prep method used in conjunction with NIR, and when done properly can give you reasonably consistent and reproducible particle size material for the NIR measurement, whether you use cups or vials. \o/ /_\ | ||
| [email protected] (Functionalgroup) |
I agree, This string on the message board had definitely been more than educational to me. All of the input i've received has been very insightful and given me alot to think about. I want to point out that my reluctance to divulge particulars about the project i am working on do not stem from any NDA's but rather something equally potentially devastating, alot of companies recently have been firing their employees for "BLOGGING" and that type of thing. I do not wish to incur the wrath of upper management by speaking openly about the details of their processes. I would also like to clear up the fact that I am not just using the NIR for this one project that I posed a query about. It is also simultaneously being used for quick ID and qualification of incoming raw materials of many different types. Not that i am trying to shy away from the popular idea of using sample cups, but the particle size and availability of many of the samples would really make using spring loaded cups just as effective as scintillation vials. For example, if i am working with some minerals which have large crystals in its natural form placing huge rocks(minerals) into a spring loaded sample cup would be just as effective as placing those same rocks into a scintillation vial. This same principal comes into play for me, when i am working with an exceptionally small number of the expensive raw materials. Even the samples we have in archive are only a few grams of (2-3 grams per lot). Please keep in mind, that the materials that are cost prohibitive to have the unlimited access to, unlike the other raw material, have traditional analytical methods in place to quantitate and perform other chemical batteries that would be of interest to any individuals auditing our manufacturing or testing. I would also like to take this time to say that i really appreciate all of the time you guys have taken to give me tips and advice concerning this project. A few of you have contacted me via email, i have received them and will respond when time permits. I would also like to clarify that although the NIR is used for Routine Analysis of incoming raw materials (ID and Qual), I am ONLY in the FEASIBILITY stages of developing the calibrations for quantitative analysis. Granted i haven't ascended as far up the ladder of NIR as you guys have, but at the same time i have a responsibility to explain to my superiors the need or the necessity for any additional expenses for the NIR to work. While i don't have my own personal corporate card to use for business related expenses as some people may have, I do have protocols that i can follow to obtain anything that may benefit the success of this machine my company has purchased. Please believe me when i say i appreciate every comment made in this string, but you must understand that i can't just procure $10,000/US for materials or resources that i don't have a definitive plan to be executed. I believe in experimentation, but not without lazer like focus. I know some people are incensed by this FEASIBILITY study and think that i am some type of bumbling slipshod lab-assistant, but please be informed that nothing can be further from the truth. Moreover, maybe some people who feel so strongly about this EXPERIMENT, should offer their advice to other purportedly "STELLAR" scientific organizations who the FDA has been kissing their but for ages but now, are in the midst of a legal imbroglio, of astronomical proportions. None of this is to say that at no point in the future, i won't be able to validate against a reference method, but i would be accused of lacking resourcefulness and foresight if a calibration could possibly be developed using less conventional methods. For example, say i hire some consultant, and that consultant tells me everything i've already learned for the past 2 years, for a total of 1500 per day? That would be even more wasteful than spending a couple of weeks exploring the capabilities of the machine and what exactly can and cannot be done using the resources available to me. And besides, we already have independent auditors giving advice to us. If anyone has anything additional they would like to add or would like to point out. Please feel free to do so. | ||
| Gabi Levin |
Hi everyone, I think this has probably been the longest string I can remember. I also think that all useful things have been said. I also think that in a situation where a company which is about to do business from materials worth thousands of dollars per Kg is reluctant to invest in a technology that will guarantee the quality of its products, I have a moral problem supporting this company. In my short time in the NIR business I have always seen the situation that people have all the wet/HPLC chemistry capability (or hire it) and then they debate whether to go NIR or not. This is to me a bizzare situation, they have NIR and they think they can avoid the reference. My suggestion remain - use the mixtures to prove feasibility, then sell to your management either buying the wet approach or hiring it. Don't fool yourself to believe that for any extent of time you will get away with mixtures, and that if at any time an FDA audit is done, that they will buy your approach, maybe you need to be shut down once to see that an HPLC is peanuts compared with three strikes with the FDA. I am out of this discussion from now on. Many thanks to all the good things I learned from all of you and I am looking for new topics. Gabi Levin Brimrose [email protected] | ||
| hlmark |
It seems like a good idea on the surface, but I think it doesn't hold up, Tony. Aquil, or someone else who might want to take advantage of such a "closed" discussion could never be sure that someone, even from the limited group, might not be one who they don't want to be allowed to access the secret information. To say nothing of the fact that their management, either Aquil's or that of some unknown future person wanting to pose a question, would want to protect their secrets by having everyone privy to it sign non-disclosure agreements. I've been a consultant long enough, and I'm sure you've been around long enough, Tony, to know that these things are not a matter of trust. They could trust all of us implicitely and still have to insist on NDAs. For those never involved with these legal matters, here's what happens: if a company should let any proprietary information out to someone whom they don't have a non-disclosure agreement with, and if in the future there should ever be a case involving patents or anything else that depends on the secrecy of the information, then the opposing lawyers could claim that the supposedly "secret" information was in fact in the public domain, and therefore any claims of secrecy are null and void, and the company would lose the case almost automatically on that basis. That's the way the world works these days. If you have a secret, it REALLY has to be kept secret, and told only to people who are legally obigated to keep the secret. So it's not Aquil's fault that he's not free to tell us more; he's also obligated to keep his company's secrets actually secret. \o/ /_\ | ||
| Tony Davies (Td) |
Hello again Aquil and Everyone, This topic (as Kathryn said) is irresistible but it has been going a long time and I suggest we need to go round it again from the beginning with ALL the information that is available to us. We now know; 1) The instrument is a FOSS 5000 with Vision 3.2 2) The material to be analysed is a dry(?) mixture of powders. Only a few batches have been made. They are prone to separation. We do not know how many ingredients are present or what range of active we are dealing (but it might be around 50%). We do not know if it is sold in powdered form. 3) For reasons we do not fully comprehend, Aquil is using scintillation vials rather than sample cups. 4) Reference chemistry is not available. I do think we actually know why; cost has been assumed. 5) The active is worth �thousands of dollars� per Kg. With this information I do not think we can give any more than the caution that laboratory mixing and production mixing are likely to be different so that a calibration based on laboratory mixing is not likely to give the true result on production samples. A calibration based on mixing will have to be validated against some reference chemistry on a number of production batches. To go any further we need to know that answers to several questions which Aquil is unlikely to feel able to give over the web. How many ingredients are in the mixture? Has there been a mixing study of the ingredients? Has there been a sampling study? How many batches per year are going to be produced? Has there been a cost/benefit analysis of the analytical approach? What is the problem with reference chemistry? I think this discussion has highlighted a problem of the open web discussion. When people pose a question they are naturally reluctant to give more than a minimum amount of information but gradually more is released. However, I think we are often still wandering around in the dark. What would people think of a small closed discussion group when it became obvious that in order to give good advice we need details that cannot be given over the open web? The composition of this group could be made up of the regular correspondents. They would agree that they would treat all information as confidential and declare an interest if one existed. This website does have the capability to run a password protected area so it would be easy to set up. What is your response Aquil to this idea? Best wishes, Tony | ||
| [email protected] (Functionalgroup) |
mark, I apologize if my previous statement was a little caustic. it was not aimed at anyway towards you personally. | ||
| W. Fred McClure (Mcclure) |
RE: Archiving Samples Bruker promotes, in conjunction with the MPA instrument, viles for archiving/measurement. Though the bottom is not "optical quality" - they claim that the variation imposed by the glass is compensated for in the calibration - which appears, from my experience thus far, appears to be true. If you turn the 5000 up on its back, you may be able to utilize such a vile - obtainable from any glass vile maker. Here again, you would need to validate the use of such viles. On the other hand, if you do not need to archive your calibration samples, you may want to do what Howard suggest - use the spring loaded cups, supplied by FOSS, over and over again. Fred | ||
| hlmark |
Hey, just trying to help. Don't know why you need 4,000 cups though, since they're reusable once you've recovered the material (little loss of the expensive stuff) and cleaned them (brushing them out is usually enough). Most places using NIR also run lots of samples (else they couldn't justify the NIR) and they get by with lots fewer cups than that. Also, for that volume of cups, should you actually need so many, it's likely that you could strike a deal with FOSS, since costs of production go down a lot with volume. They might even be willing to make a custom, low-cost cup for that number of cups. Like I said, just trying to help - don't take my head off for trying to come up with other ideas. \o/ /_\ | ||
| [email protected] (Functionalgroup) |
yes Mark, that is a good point. my challenge to you is to show me where i can get about 4,000 of these sample cups, at an affordable price. I know exactly what you are talking about. But the cost of the spring loaded sample cups to the volume of samples i have is cost-prohibitive at this point. moreover...my facility has a small range of operation hours for the chem lab. loading those spring loaded sample cups is even more time consuming than scooping raw material into the vials. While it would optimally be the best situation to use spring loaded cups. scintillation vials are cheap, easily replaceable and doesn't require your fingers to go through yoga class to use them. Notwithstanding that getting a strong signal out a small volume of a sample is alot easier in a scintillation vial than attempting to get a spectra of a small amount of sample (e.g. expense of sample) im talking about thousands of dollars per KG of raw material! | ||
| hlmark |
Aquil - I just realized that it's somewhat surprising that nobody brought this up before, but how important is it to measure the material in these vials? The reason I ask is because most manufacturers provide sample cups for measuring solid powders, and these cups have a spring-loaded backing plate of some sort, specifically designed to help ensure reproducble packing of the powder against the cup window. Certainly FOSS is included among those manufacturers. Since you're already aware of the need for getting reproducible spectra, you might want to consider this change to your procedures, so that you measure the material in one of these sample cups, for maximum measurement consistency. Howard \o/ /_\ | ||
| [email protected] (Functionalgroup) |
Ok, You guys brought up alot of good points. I agree that the possibility is strong that manufactured batches will pose a great challenge to the calibration i am developing. I've spoken with my supervisor and was told that reference testing will be done. I explained that any calibration methods developed on the NIR will need to be validated externally by some reference method. The situation is that most of our testing capabilities are being expended on more immediate needs at this point in time, (finished product and stability). Not to mention the costly standards and testing supplies that are necessary for accurate analysis. To answer Dave's question i am using VISION 3.2 by FOSS. Also a note on the grinding. I am not using a hand grinder i am using a coffee grinder. and Im grinding both the hand-mixed batches and the maunfactured samples, to ensure maximum homogeneity. As far as rotating the sample a few degrees during successive scan, that procedure has been firmly in place for quite some time. I learned early on that shaking the sample does not work too well with certain segments of the raw material population as they are bulky mixtures prone to separation even at the slightest hint of vibration. I find that gentle packing and sample rotation ensures i get consistent spectra. to comment on the point mentioned by Kathryn about the inter/intra vial lot variation. we use the same vials over and over. they are cleaned and dried, which isn't too hard because most of the samples are powders and granules. The same group of vials that were used to create the projects are the same samples used in routine analysis and in this current calibration project. And BTW the goal of this experiment is to get REAL numbers for percentages. | ||
| hlmark |
It occurs to me that if what Aquil really needs is a fast go/no-go test rather than actual quantitative values, then he might be better off using Mahalanobis Distances or Conformity Index instead of a quantitative algorithm. If he takes this approach then he not only won't have to worry about his inability to get lab values, but he could also base his training set on all production samples, and get away from the whole issue of relating lab samples to production samples. What he would need to do is to obtain many production samples and measure their NIR spectra, and then divide them into those that pass all final production tests, and those that don't, once the tests are done. Samples might not keep, but data will keep practically forever, if stored properly. I have to say, though, that the argument about mathematical or software complexity doesn't carry much weight with me. Nowadays everything is pre-programmed, and I don't see that which menu selection you click on makes a whole lot of difference. Validation has to be done in any case; we're all agreed on that, I think, but the choice of algorithm you use doesn't affect that. It's more important to choose a methodology that matches the task at hand and the constraints on it. In this case the constraints include the diffculties in getting reference lab values, and difficulties in making "hand" samples match the production samples, so if those can be avoided, then methods for doing so should be considered, I think. \o/ /_\ | ||
| W. Fred McClure (Mcclure) |
Maha. distance and conformity indices require rather elaborate software and some expertise in applying. I like the simple things - like r**square. If you compute the r**square between two spectra you can get a feel for similarity/dissimilarity very quickly - and a measure that is intuitive. If the new raw materials are dissimilar, then Aquil will have to beef up his calibration with new entries. Surely, none of us allow NIR calibrations to run forever without checks to determine their validity/accuracy. Sure, Gabi, he may have to continually build his calibration until it becomes "global." With the 5000, he can run 100+ samples in 8 hours. That means lots of samples/spectra; but with NIR lots of scans is no problem. Aquil has told us that checks on the final product are in place (and working) to assure product quality and reproducibility. Apparently, he cannot tell us what these tests are; and I am not asking him to tell us. Nonetheless, if the final product is tested by "wet chemistry", then he can begin to build his "wet chemistry" calibration (WCC) from these samples/tests - extending his understanding of the mixture calibration. At some point, he may want to switch to the WCC - but my guess, and it is purely a guess based on experience with mixture calibrations, he will stick with the mixture calibrations. For me, Aquil's project is very exciting - much like the old continued movies. I hope he will not leave us hanging - without knowing the outcome of his work. I would encourage him to "write it up" for publication in NIR news. Good luck, Aquil! You now have the blessings of the entire NIR community. Keep us posted. | ||
| hlmark |
Gabi - well, there are ways to ascertain whether a spectrum is consistent with the data in a calibration set or not; both Mahalanobis Distance and a new approach, Conformity Index (starting to take hold in the pharm industry), are ways to test that. Those are both based on properties of the spectrum alone, indpendently of the quantitative aspects. The problem Aquil will run into there is that most likely, as you intimated, real production samples will likely NOT look like the hand-ground ones, and thus mostly fail any of these types of tests. \o/ /_\ | ||
| Gabi Levin |
Hi Everyone, I am sure Aquil will make progress, and he will be able to show the applicability of the mixtures approach to a certain extent - but consider now the following: 1. He has a working calibration that works well with "validation" samples he prepared by hand mixing of materials from X number of batches of raw materials. 2. Now he gets raw materials batch numer X+1, how does he validate his calibration on this material? What if there is variation which is unaccounted for in the calibration? 3.Suppose he updates the calibration with mixtures made from the new raw materials, does he have to repat that with every new batch? 4. How does he validate his calibration on some regular regime of periodical proceudre? Does he validate his calibration on new mixtures or does he validate against some primary chemical method. I support him using the mixture approach in order to develop the confidence that further investment in reference metod, either by self owned equipment, or by purchased services is a useful investment, and that money will not be thrown away, I do object to carry it to the extent that the management will start to believe that they can get away without it. When they start beleiving they can get without it, that's where the first problem with the "easy" way will send every one in management lashing at Aquil. Well, Aquil, looks like your fate is in your hands now, do your best, but if this approach fails at some time, don't feel bad, it is not your fault, there are sometimes limitations even to "perfectly" executed calibration work. Thanks, Gabi Levin Brimrose | ||
| Kathryn Lee |
Aquil, You said that you were using scintillation vials. Several people have suggested rotating your sample, but they may be thinking that you are using a different cell where the material is pressed into the cell and the window area is about 2 inches across. I think your sample may be actually loosely packed in your vial, which may be less than an inch in diameter on the surface that is being sampled. In this case you may want to shake up the sample instead of rotating it to effect sample variation. If you are pressing your sample into a cell, you may want to repack the sample to see if that causes significant spectral differences. You can check to see if rotation or shaking or repacking affects your spectra. If you still have different particle sizes of the components you may want to be careful not to tap the vial because that can cause deblending. I have also seen cases where static on a cell window caused deblending. Vials may also have some variation intra- or inter-lot which will need to be taken into consideration in your calibration development. | ||
| David W. Hopkins (Dhopkins) |
Aquil, You should consider making a calibration set of 20 samples (at least) and 20 samples for a test set. You need to have sufficient numbers of samples in both sets of samples so that your statistics are meaningful. Don't trust cross validation as a validation of the calibration; it is useful for guidance in selecting the number of factors to use in your PLS models. You need to have the calibration and test sets be representative of the product, and covering the same range, but the calibration and test sets must be independent. Then you should read at least some of the samples, better all the samples, with the replications on rotation that Fred suggests. This will tell you how reproducible the readings are, and whether replicates would be useful in routine measurements. With hand-grinding, a big source of variability will be particle size of the samples. This will have a big effect on the baselines of the spectra, and also affect the light path within the samples. The baseline effects can be removed by pretreatments such as derivatives and/or multiplicative scatter correction (MSC) or Standard Normal Variate (SNV) transformations. I have found that 2nd der followed by MSC (or SNV) is often effective. What chemometrics/calibration software are you using? Best wishes, Dave | ||
| [email protected] (Functionalgroup) |
Thanks Fred for the encouragement, Perhaps i should consider making more batches to increase the size of my "matrix". i only have about maybe 20 samples mixed by hand. I think it would be interesting to see just how the calibration will improve as i add more entries to the calibration. Do you think the averaging the collected spectra into a new set is redundant? being that the calibration methods innately average out any noise? | ||
| W. Fred McClure (Mcclure) |
Aquil, You have my sincere compliments on your decision to move forward with your ideas. There are a couple of suggestions: 1. Make up as many mixtures as possible (in my work with tobacco mixtures, I had 200). The more you have the better the calibration will be. 2. The consistency of your calibration samples should be equivalent to your final product - following the same preparation and mixing procedures used for producing the final product. 3. If your 5000 does not have a rotating cup, rotate the cup manually 90 deg each time for four spectra per sample. You can use these spectra to study mixing variations. No doubt, you will get other suggestions. Best wishes for success, Fred | ||
| W. Fred McClure (Mcclure) |
I would hope that Aquil is not confused by all this debate. He (I apologize if I should be using "she") is to be commended for arriving at the mixture experiment on his own. Well done, Aquil! This is a valid approach. Indicative of our debate is our frustration over the fact that we cannot use mixture calibrations universally. BUT, there are many examples out there where it does work. Another example, was Shootout-2000 (the alcohol dataset) which had as its objective to calibrate on mixtures of ethanol and water and predict hard spirits (drinking alcohol). It was clearly shown that the alcohol content of many hard spirits could be predicted from the mixture calibration. Where it failed was in predicting liqueurs and other dinks with strange additives. I hasten to add, if we had known what the additives were we could have predicted the liqueurs too - from a mixture calibration that included the additive. The only place we have been unfair to Aquil is when we tell him to "do nothing until you get support for a full blown program." It is unfair to him to encourage him to sit on his thumbs until management comes up with the bucks. If I were his manager and I thought he was doing nothing until he had the money for wet chemistry, I would be very disappointed in his foresight. I would be even more concerned that this new toy (the 5000) was idle. It is also unfair, to indicate that he needs a consultant or intensive training in order to "do NIR correctly." There never has been a correct way to do NIR. That is why NIR is honored as a full member of the spectroscopy community. We got here by pushing the envelope. Doing things that others (particularly, the classical spectroscopist) said we could not do. Aquil is much further along with his understanding of NIR than some give him credit. He understands the calibration process. He has a method for calibrating his instrument - he should not be discouraged from continuing with his original thesis - "mixture calibrations will work for this application." And I agree with him. Sure, it may not be the ultimate, most accurate - BUT, none of us know that at this point. Do the experiment, Aquil. Make up 100 mixtures that emulate your final product. Calibrate! Do not be discouraged by all the gloom and doom sayers. Keep studying, keep learning (like I do, and I have been in NIR for 46 years; still do not know everything), keep trying YOUR IDEAS. NIR spectroscopy is still an infant compared to other IR methods. Help us to make it stronger by proving your ideas work. | ||
| [email protected] (Functionalgroup) |
mark and fred, you guys both bring up some strong points. I definitely want to maximize the probability of this feasibility study working out while doing minimum damage to my scientific career. One thing i have learned in my short few years as a scientist is that even the resemblance of incompetence, doesn't look too good in the eyes of corporate superiors not inured to the subtleties of the scientific realm. As for my plan with the project, I have decided to go forward with the plan. I will detail it here, while trying to be as general as possible: I start by weighing out several trials of the mixture in a multivariate way. such as 3 samples of 50%,56%,59% and other concentrations on the perhiphery of the target concentration range. These mini batches were mixed by hand to simulate the manufacturing process, for about 10-15 minutes over several days each. After the mixing process they were placed into scintillation vials (quartz) and spectrum were measured. 5 scans per trial. a calibration was then attempted with this data, cross validation shows errors of about +.3 percentage points. Then the collected spectra was averaged over the 3 trials. and a calibration was then attempted with this data. I'll get back to you on the cross validation details but there was a slight improvement. Next i took the original samples and ground them to a fine powder. to match the consistency of the finished product i am trying to simulate. Scans were taken, spectra averaged, im still working a few trials so when i get the data i'll be sure to let you guys know. I want to avoid any dehydration or other chemical processes that will introduce chemical changesin the sample not present in the manufacturing. Time is not a problem, my lab is not a high work-volume laboratory, at least not with my department. If anything it would be an interesting experiment in NIR capabilities. | ||
| hlmark |
Fred - I think that's a little unfair - I'm not assuming he'll fail, I'm trying to help think of ways to maximize his chances for success. That's usually especially important for a first project, no matter what the technology; once you've shown that a particular technology works, it's easier to keep the support resources flowing. Also, statisticians never said that MLR wouldn't work; after all, they invented the technique (remember Gauss!). But think of poor Aquil: our internal debates must be confusing him no end!! In the end he will have to decide which way to proceed. To some extent, as I said at the very beginning of this thread, he seems like he's in a situation where he's going to have to try out the mixture approach, because his alternative is to abandon trying NIR at all, and THAT would probably be a mistake. But while he's at it, he should also strngly attempt to follow the other pieces of advice that have come up: use external laboratories to get analytical results at least for vaidation, etc. \o/ /_\ | ||
| W. Fred McClure (Mcclure) |
There you go again, Howard. Assuming Aquil will fail. Chances are, Aquil will succeed. How do yo know the mixture-project will fail. You got some crystal ball? I'll bet you, Howard, you have never done a mixture experiment in your life. Credibility? You know as well as I, there is seldom a "complete failure." Only degrees. That means, if something is a partial failure - it was a partial success. NIR technology is built on "stabs-in-the-dark." All along it has done far more than the "pureists" said it would. Remember when the statisticians said MLR would not work - but it did. Remember when the classical chemists said we couldn't measure anything less than 5% - spectra too complicated. Well, we did. Remember when the pharmaceutical industry said NIR was too much of a black box - well, now they are embracing the black box they once denied. I say, give it a shot, Aquil. You've got nothing to lose and a lot to gain. Do it at night if you have to - but do it! You will be glad you did. Fred | ||
| hlmark |
Fred - you've got it: it was Marbach. Aquil does have something to lose, even more than one thing: time and credibility. Time is probably not too big a concern, since an NIR development project will take a long time anyway, and a few days extra will not matter too much by the end of the project. The real problem is credibility: if he spends all that time and then fails, he will damage the reputation of NIR for that company, possibly to the point where they will not allow NIR in the door again; more than one company feels that way already. So especially for his first NIR project, he needs to maximize the probability that it will succeed, for the sakes of both his career and the reputation of NIR. \o/ /_\ | ||
| W. Fred McClure (Mcclure) |
Howard, You must be talking about Ralf Marbach - He has an interesting approach but not well documented and certainly not verified. I am still convinced that Aquil has nothing to lose by investigating the mixture approach. And -may find it quite adequate. It has been shown to work quite nicely. McClure, W. F., H. Maeda, J. Dong, Y. Liu and Y. Ozaki. 1996. Two-dimensional correlation of FTNIR and FT-Raman spectra I: Mixtures of sugar and protein. J. Appl. Spectrosc. 50 (4): 14-22. Fred | ||
| hlmark |
First of all, Don: I think you should go to work for the FDA; what you've described is essentially their whole PAT initative, in a nutshell. That's exactly what they want the pharmaceutical industry to start doing: study their processes thoroughly and learn how they behave under different conditions, develop theories to describe the behavior and scientifically relate the theories to the actual quantitative behavior. It reminds me of one description I read of what medical students go through in medical school: before they can learn to treat disease, first they have to learn how the healthy body is put together and works, so that they can tell when, and what, disease is present. Purely empirical studies of the form we are used to doing, aren't going to cut it in the regulatory world. And the neutraceutical industry, as well as any others that aren't currently doing this, could do a lot worse than to take that advice to heart. Secondly, there exists (I think) a "magic algorithm" that might be most of the answer to Aquil's problem, looking at it from the empirical point of view. I don't know how well that algorithm works because as far as I know, nobody has yet used it except it's developer. Unfortunately I also don't recall the full name of the fellow who developed it, I only recall that his first name was "Rolf". Basically, what he showed was that it is possible to decouple the need to find corrections to account for the uncontrollable variations in a data set (e.g., moisture, other constituents, particle size, etc.) from the quantitation of the analyte. This being the case, you don't need a lot of analyzed samples in order to create a calibration, you only need a lot of samples that include the expected variations in composition, much as we ordinarily recommend for quantitative work; but those samples don't need to be analyzed. In addition, you do need some analyzed samples, but you only need a very few of them (in the most favorable case, only one of these). He then showed how to combine the information to create a complete calibration. The concepts he presented can be applied to any of our common calibration methods: PCR, PLS, MLR etc. The major problem, as I said, was that utility and effectiveness of this procedure has not yet verified by anyone else, which is the most basic test of any scientific proposal. Howard \o/ /_\ | ||
| djdahm |
I know that today's managers don't want to hear that quick and dirty doesn't work, but Fred is right. Spending 5 days doing background analyses is a good step. That isn't "quick and dirty". That's doing your homework. This will probably be of no practical help to our young friend, but it will make me feel better. Yes, if you want to do the NIR spectroscopy the way it is currently usually practiced, you will have to be ready to invest considerable resources. Exactly what form that takes: reference analyses, a consultant, or an employee with relevant experience; I don't know. But I know the initial steps you outlined are good things to be thinking about, not bad. First, of all, NIR is not inherently a secondary technique; we just choose to use it that way most of the time. One also has the right to choose to be an analytical chemist using NIR as a tool, and do all the things that we do to prove the validity of our results in any analytical procedure. Now several people have made very important points in this chain. I, however, what to challenge our "NIR arrogance": The idea that because we have developed a way using reference data, sound statistical sampling, and powerful "Chemometric" techniques, that is the only way NIR should be done. When I managed an analytical group at a major chemical company, if a product group needed our group to produce the monitoring technique; a person might be assigned to look at the feasibility of NIR as the process/product-monitoring tool. I would have been livid if they only looked at reference analyses on finished product and developed a calibration. I would have expected them to get spectra of all the raw materials, look at mixtures in pairs, look at the reaction mix before reaction, looked at the material during reaction, and the lab and pilot plant (if it existed) product. In short, I would have expected them to learn as much as they could about the process (including whether any components reacted with water) before they even thought about a calibration. When they developed a calibration, I would have expected them to keep the method as simple as would do the job. Throwing in extraneous sources of variation was a "bad thing" in those days (rather than assume, "What the hell, it can only make the calibration more robust). Using only relevant wavelengths was a good thing, not old-fashioned. If they were going to rely on reference analyses, I would have insisted that those methods were sound and asked to see the statistics. If they were using standard mixes as the calibrations, I would have insisted on seeing spikes, splits, and replicates. I may have had to call the product manager and explained the facts of life to him, but they weren't going to get a half-assed job. Proud as we are of all that we collectively have accomplished in the field we call NIR Spectroscopy, we are still largely a satellite group to analytical chemistry, rather than a part of it. We bemoan the fact that our "discipline" is not taught in colleges, but we as a "discipline" do very few things that show what is going on. Rather we highlight building bigger, better smoke screens from which results can mysteriously be pulled into the light. Now I not blaming those "Chemometric" folk. In some ways, they have saved our collective butts and helped us have enough success that it is even worth having this discussion. And I don't blame the FDA. They have plenty of reason to be suspect of a renegade technique that isn't working it's tail off to become mainstream by generating the understanding that supports its conclusions. I'm blameing us, those who make a livign practicing the technique. It really upsets me that when I read: " I still say - try it with the mixtures. What have you got to lose - 5 days work. I've wasted more time than that fishing." and find that this is treated as a controversial proposal. I realize that our technique was born through its success at handling tough problems. But even though the pioneers wound up accepting empirical evidence (like a validation set) that what they were doing a good analysis, it didn't keep them from doing sound science at every step they could. I applaud every time Karl Norris shows that MLR gives better results than PLS for a particular problem. It's a reminder that our statistical black boxes don't work as well as they should. | ||
| [email protected] (Functionalgroup) |
Well as much as it frightens me to say it, you guys are absolutely correct! raw material quality can and does vary from lot to lot. But the company that i work for does extensive testing and quality control to ensure the maximum homogeneity for our raw materials. Mosture content is one of the things we are able to accurately quantitate, among other qualities. Moreover, although this quantitative analysis conundrum is slowing down my progress on the machine, the NIR performs brilliantly qualitatively allowing us to quickly uncover any anomalies in incoming raw material. Although the FDA doesn't regulate the neutraceutical industry as tightly as the pharm industry, we have our own safeguards in place to insure maximum product reproducabilty. And the FDA does periodically come to check out our lab, and so far no complaints. In fact my lab has even been said to exceed standards set in place by the FDA. And as we have all seen in the news, FDA scrutiny doesn't necessarily make a material safe or effective for use, honesty and thorough testing does. Not to mention how the FDA and it's lobbyists frequently lambaste the neutraceutical industry and homeopathy. But I digress I get the idea, handmixing batches for nir calibration is not a good idea for a great deal of reasons as you guys so politely pointed out. I guess i should just tell my supervisor that i strongly reccommend that we find an affordable reference method for the materials we wish to quantitate with the NIR. I am aware of the chemistry of most micronutrients. Granted degradation is accelerated in a few products, but we do stability testing. My lab is not out of the stone age. Neither does my lab release finished products without thorough testing via other more reliable methods. BTW scary is some of the studies done by many supposedly scientific corporations. Again, i digress. I don't have any illusions about NIR, in fact i hold probably an unhealthy amount of skcepticism about the quantitative work i am able to achieve via hand mixing batches for my calibration set. Which is why i asked you guys for your input. I would love to work with some of the data we have accumilated via wet chemistry and other analytical methods we are capable of here. This question basically arose out of a feasibility study im doing on a material my company has only produced once, MAYBE twice. And to get a feel for the limits of the machine im working with. This isn't going to be integrated into manufacturing within microseconds. So you guys think i should either hire a consultant or go for training? | ||
| Gabi Levin |
Kathryn is absolutely right, now that we know that the issue is nutraceuticals I am in a position to be even more skeptical of the usefulness of the quick and dirty approach. Manufacturers of raw materials for the pharma industry at least try to maintain small variability batch to batch,the FDA is on their backs. Suppliers to the nutraceuticals are not even close to be reproducible, anyhting can happen, the worst of all, but not the only is the % purity of the ingredients, and the consistency of the composition of the "estraneous" compounds that are present. I worked once with a nutraceutical company on Glucoseamide and chondroitin - beleive me you don't want to know the variability batch to batch. These people at least were honest - they focused on one supplier with reasonable quality and dropped the others. Gabi Levin Brimrose [email protected] | ||
| Kathryn Lee |
This is an irresistable topic! First, if moisture is not accounted for, the simple proportionality of the samples will not be very accurate, and hence the calibration will not be very accurate. When you weigh a 100 g sample with 10 % moisture, you have only 90 g of what you thought you were weighing, and 10 grams of water. That would be ok if you always had 10% moisture, but it probably varies a few percent. Second, all of you please open your chemistry texts to discover that vitamins and minerals are reactive, even if they are not called "actives." The least of the chemistry they can do is form different hydrates, then there are varying degrees of oxidation, degradation, polymerization, etc. Are you planning on selling these neutraceuticals with label claims with a certain amount of each material? That's a bit scary... How many tablets were you planning on checking by NIR? It may be interesting to check the beginning, middle, and ends of the run of production samples just to see the spectral variation and compare it to the spectral variation in the calibration set that covers the complete ranges of the individual components. That might give you an idea of how much variation you have. There are some other factors to consider, such as sample presentation to the instrument, whether you are using transmission or reflection, and what spectral regions you are using. Some spectral regions may be too absorbing to use. Then there is design of experiment to make sure that all the components vary independently, and that they cover a wide range of raw material batches. Then there is always deblending to consider if the particle sizes and shapes of the materials are different, and I know that that is true for vitamins and minerals. Plus, some vitamins come pre-formulated, and are not pure, so you have to worry about the proportionality of components in your raw materials. If management wants to get results from a very complex spectroscopic technique, they should be willing to spend the money to hire a spectroscopist or to send you to a spectroscopy training course for several weeks. You are getting long distance advice from experts who have tens of years of experience, and who would still have to work hard to solve your problem. Let me pose this question: Has anyone ever seen a non-spectroscopist build a calibration that works properly for the long run? (I chose my words well. I am not asking if a calibration was used for years, but whether it provided accurate answers.) Ok, how about on a system more complex than hydroxyl number? From what I have seen, you can spend a lot of time trying to get the NIR to work well without the proper training and knowledge, and have it give you improper answers and not realize it, or realize you cannot get it to work, and then have it collect dust, or you can just go directly to collecting dust. Every project needs cost justification. How much profit will this new product yield? How much loss if selling bad stuff closes down the plant? That might change the perspective on the NIR calibration vs. chemical testing, and getting some good spectroscopic advice via hiring or consulting. What I have heard, and please let me know if anyone has done otherwise, if you develop a calibration on lab samples you have a calibration for lab samples, not production samples. You will at some time have to add production samples to your calibration, and without concetration values for those samples, you will not be able to add them. I think that someone already alluded to the fact that you can use the spectrometer to help you pick which production samples have the broadest spectral differences and therefore would be best to perform primary analysis on. That might help reduce costs. If you can run primary analysis on all of the other components, you might be able to know how much is active, but I think the chemistry involved may be a bit out of your lab capabilities. Good luck! | ||
| mpdc |
do you have a drying oven? I guess that if you want success you will have to take your hand-made batches and slowly grind them down to smaller particle sizes, with at every step taking several spectra at different moisture values. As your samples will not be independent, you will need more than the 40 - 60 samples usually recommended. | ||
| [email protected] (Functionalgroup) |
Thanks to everyone who is giving me their opinion on this matter. I really appreciate it! To respond to Gabi's comments, and to clarify, I am not working with highly-reactive, FDA regulated "ACTIVES" in the sense you are talking about. The materials i work with are fairly inactive and easily accessible. I am working primarily with vitamins, herbals, minerals and things of that nature. Secondly, I have established that i want my results to be within +/- 5% of the input in my hand-mixed batches. to be perfectly honest, I haven't even really considered taking moisture variation into account,even though i realize it plays a crucial role in the spectrum of the final product. One thing i've come up with is to take some of the few batches that have already been manufactured and grind them up to a fine consistency, THEN take the hand-mixed batches and grind them up until the consistency matches the consistency of the manufactured sample, previously ground up. Upon setting up some calibrations with the hand-batches i've already created, I get good linearity and fairly small errors when cross validation is done, granted. But when i try to predict the concentration of a sample the error goes off the chart. I've found the KRAMER text that was reccommended by someone on this message board to be extremely helpful as far as understanding the optimization of calibration and training sets. So thanks for that piece of help. as for fishing, never been. But salmon is pretty tasty. | ||
| Nuno Matos (Nmatos) |
I've been following this topic with great interest and I don't I've got something to add. I like the small victories strategy. By having small victories one can justify an investment in the technology. For this particular case I would suggest, due to the lack of financing, the use of NIR for itself. This is my suggestion: - You take spectra from the greater number of a raw-material batches possible; - In a lab-scale (the most similar with the manuf-sclae as possible) use those same batches individually to get the final product - Use the batches that have originated good final products as a calibration set - Use a clustering method (for instance PCA) toconstruct a model with desired quality: "qualificative" model This will only pay-off if obtaining the final product is not cost-demanding AND the quality testing on the final product is not, also, cost-demanding. Best regards Nuno | ||
| Gabi Levin |
Fred, If I don't change the rules now, and if he is in somewhere involved with pharmaceuticals, as may be iferred from the use of the word active, the FDA will. He just needs to know that. Instead of doing analysis you propose that he will create calibration mixtures every time he gets new batches, fine, I will buy that, but what will he do about changing moisture if the mixtures are being processed under different conditions, and from experience, even fluid bed drying under same machine parameters does not yield same moisture run after run. And on top of it, we don't even know what range of concentrations and accuracy he needs to achieve, which is also a factor to be considered. +/- 2% is different at 20% level from +/-0.2% at 2% level. Well, I still think that his management needs to understand that you can't, for a long time, use secondary methods like NIR without good control of a primary method. Unfortunately NIR is less easy than MID-IR or UV or visible for that matter. About fishing - I don't do much of it, but it is a nicer way to spend your time. Have fun- life is too short. Gabi Levin Brimrose Corp | ||
| Bruce H. Campbell (Campclan) |
There is a possible other approach and that is to use "classical" spectroscopy. That is, take spectra of the laboratory mixutures, both before and after processing, see if any band of interest is proportional in height to concenration. I know the band(s) may not be well separated. If this is so, one can take derivatives to better resolve the band(s) of interest. I have done this to follow a reaction. The use of the fourth derivative was satisfactory because the band had a fairly high intensity. The baseline problem disappeared also with the fourth derivative. I do not advocate using the fourth derivative with chemometrics, however. All this could also show if using chemometrics is consistent with the results from peak height measurement, giving a little more reassurance. The process should be repeated periodically, as Fred and others have pointed out. Bruce | ||
| hlmark |
Fred said, "What have you got to lose - 5 days work. I've wasted more time than that fishing." Ah, but, Fred, which was more worthwhile??? \o/ /_\ | ||
| W. Fred McClure (Mcclure) |
Gabi, Now you are changing the rules. Blending tobacco was once a common problem. Working with tobacco companies in the 70s, we demonstrated that we could calibrate on mixtures and control mixing on-line. It sounded to me like this researchers problem is much simpler - he is blending rather "inert things" with an active. If that be the case, then it should work as well as a chemical analyis. Having said that, I agree that if you continue to change the character of your raw materials, then you are going to have to update your calibration. You are not going to have a long lasting calibration. Nonetheless, updating your calibration by mixtures would be much easier than doing chemical analyses - and far less costly. Who knows, he may get enough mixtures into his calibraion to produce a global calibration that works for his process. You know better than I, it is hazardous to leave a calibraiton unattended. The safe thing to do is to check any calibration periodically - some require more checks than others. What is this guy going to do, shut down his NIR operation until management comes up with the big bucks. No! He ought to go into the "Boss" as say, "I'm gonna try doing this with mixtures. BUT, if you want have a backup, then come up with some bucks for chemical analyses. The bucks (analyses) will tell us how well the mixture calibration is working." Thus, as he checks his mixture calibration he can begin to build, if necessary, a better calibraiton - which in the end - will have to be monitored like any other calibration. Howard, this is not purely a non-technical solution. If he had data showing that the mixture calibration does not work, then it would be non-technical. Quit changing the rules of the discussion. I still say - try it with the mixtures. What have you got to lose - 5 days work. I've wasted more time than that fishing. | ||
| Gabi Levin |
Hi Fred, I can't agree with you that he can get a reliable, long lasting calibration that will be robust and withstand the variabilities in: 1. Particle size that will come along number of batches of different raw materials and procesing conditions. 2. Variabilities in small contaminants that always come along with raw materials from even the same supplier, let alone different suppliers. 3. Differences in residual moisture resultig from previous processing. To account for all these variabilities (and I haven't even covered them all) he will have to develop his calibration with samples taken from real process batches, and he will have to get them from at least 10 different batches, assuming they will represent different batches of raw materials, different suppliers. I am in final stages of implementing the determination of two contaminants (0 to 3% each) in distillate after receipt of FDA approval for NIR based release of the distillate, and in order to account for variability in residual contaminants that are not measured, we needed over 150 samples from different days, batches, etc. and this did not include particle size because we talk about transparent liquid. Thanks again, Gabi Levin Brimrose [email protected] | ||
| hlmark |
Fred - the major issue is whether (or perhaps: how well) Aquil can mimic the process condition in his laboratory. I think he's hit the nail on the head, though, his major problem is the non-technical one, to convince his management that they'll have to give him the resource he needs if they want this project to succeed. Emil has written a very nice description of the problem, but few will see it because he wrote it in a pharmaceutical magazine. In it he takes the role of a manager, who says to the newbie chemist/NIR spectroscopist something like: "I don't understand why you can't learn to work the NIR just like you learned to work the HPLC, and get us some results in a month or so." Well, we all know why, but the fact remains that the managers don't and it's the guys like Aquil who get caught in the middle. \o/ /_\ | ||
| W. Fred McClure (Mcclure) |
Frankly, I can't see how doing chemical analysis is going give greater accuracy than the mixing data. After all, chemical analysis of the final mixture is going to be determined by all the mixing factors - mentioned by Gabi. If you make an adequate number of mixtures, the regression on the mixtures will be as good as the regression with chemical analyses. (Assuming there are no serious chemical reactions in the mixtures.) If you don't get a good mix, you are going to get erroneous chemical data too. If I were this person, I would try it, even as an effort is put forth to get support for chemical analyses. A mixing trial would be very interesting and could prove quite adequate. I have seen it done successfully with other mixtures. Why not here too. | ||
| [email protected] (Functionalgroup) |
Ok, I see that work-arounds should really be avoided with NIR. To answer some questions, the types of materials i am working with are powders and they include minerals, herbal extracts, vitamins and the like (neutraceutical). I know that reference chemistry is crucial to the reliablilty of NIR predictions, however my superior only wants me to focus on using NIR to quantitate actives that we currently do not have the ability to analyze with "classical" chemical methods. I suppose that a more apropos question at this point is, how do I a virtual corporate-scientific tyro, convince my superiors of the support required to really make this NIR machine come alive for them. I really don't want to jeopardize my own career by working on some project that turns up to be one slipshod mess of analytical chemistry. | ||
| Tony Davies (Td) |
Hello Aquil, I would like to emphasise what Gabi said. Before you use NIR spectroscopy it must be validated against reference chemistry, so why try to calibrate it on the cheap? If your analysis is important then there has to be funds available to do the job. NIR analysis is rapid and low cost in operation but it MUST be set up correctly. Most "cheap and cheerful" approaches end in failure and give NIR analysis a bad name with the bosses. So please get properly funded or do not do it! Best wishes, Tony | ||
| Gabi Levin |
Hi every one, I am still around, guys. I agree with what was said, and not knowing exactly what type of materials you are talking about - you mention active and you mention active in raw materials, that is confusing a little bit. So, if by any chance you are referring to powders that get mixed, or liquids for that matter, then you can gain one thing from preparing "aritifical" samples - if prepared carefully, and if you can be sure that the desired species are unifromly distributed within the mix, and if you collect spectra from sufficient volume of sample to get good representation of the nominal value - you can get a good feeling for the feasibility - based on the regression and if the regression is truly using the infromation from the spectrum of the desired species. Once you have more confidnce in the feasibiltiy you will at least be in a better position to decide to spend the money on outsourcing the reference analysis. Under any circumstances - even if you have a "working" calibration - how would you test it to validate it without reference testing? So, there are no shortcuts of any significance! Gabi Levin Brimrose [email protected] | ||
| Bruce H. Campbell (Campclan) |
I agree with Howard. One simple step you should consider is to visually examine the spectra of production samples with a few "artificial" samples you prepare. Even if the spectra are similar success is not guaranteed. Basically you are looking to see if there are drastic differences that would harm any calibration. Have you considered having an independent lab analyze the samples? Provided you have the funds and not tied to secrecy. Bruce | ||
| hlmark |
It's usually not recommended. In emergencies, you may have to do it anyway, because there's no other way to get any sort of calibration data. If you're in that position (as you seem to be) then what you have to do is to try to have your sample preparation mimic the processing that the actual samples will go through. This is the part that makes the procedure "not recommended": the difficulty in reproducing process conditions in the laboratory. Sometimes pilot plant samples have been used, that's usually better than laboratory-prepared samples because the pilot plant at least attempts to reproduce the process. And even then, the inability of pilot plants to do that very well is notorious. The bottom line is that your success will depend on how closely your synthetic samples will look and act (in respect to their near-infrared behavior) like the process samples do. If you can tell us more about your samples we may be able to come up with more specific suggestions, but that's always going to be with the goal of achieving the results described in the previous sentence. \o/ /_\ | ||
| [email protected] (Functionalgroup) |
Hello everyone, My name is Aquil and i am currently working on developing calibration methods for quantitative analysis using a FOSS NIRSystems NIR machine, model-5000. I understand the need for an independent reference method to quantitate the actual amount of active present in one's raw material, however, what if your lab is not equipped with the resources necessary to perform specialized tests on some active materials present in finished products...BUT you have access to all of the raw materials (actives and excipients) that make up the finished product? Could one get away with developing a calibration set based on carefully-monitored, manually-prepared, mini-batches of the finished product, for which the input of the actives in question was recorded and concentration values calculated after the mini-batches were prepared, without compromising the reliability of the calibration? any responses or ideas about the trustworthiness of this particular approach to building calibration sets for quantitative analysis would be greatly appreciated! -Aquil |