Bioequivalence and Bioavailability Forum

Main page Policy/Terms of Use Abbreviations Latest Posts

 Log in |  Register |  Search

Back to the forum  2018-06-22 14:50 CEST (UTC+2h)
sagark
Junior

2010-07-29 02:26

Posting: # 5698
Views: 7,047
 

 Manual integration [Bioanalytics]

Dear members,

This is my first post to the forum. I would like to thank HS for running and all for contributing their views into this excellent forum. I have doubts for the concept of manual integration. In what cases it is done and what are the considerations for it? Is it accepted by USFDA? Also how does one know if the manual integration done is proper or not. Please guide

Sagar
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-07-29 15:30

@ sagark
Posting: # 5699
Views: 6,939
 

 Manual integration

Dear Sagar!

» This is my first post to the forum. I would like to thank HS for running and all for contributing their views into this excellent forum.

Welcome to the club and thanks for the :flower:

» I have doubts for the concept of manual integration. In what cases it is done and what are the considerations for it?

For some background have a look at these posts: 2261, 2728, 4447.

Let’s have a look how peak integration in chromatography is performed. The detector may deliver signals at high data rates. But if we would use this raw signal, we would see a lot of noise on top of even high peaks. Therefore the raw signal is bundled to slices (either – rarely – within the detector or by the CDS), based on an appropriate time constant. As a rule of thumb for the narrowest peak the width at half height should be divided by ≈10–20. For a 10s peak we would set the data aquisition rate to 0.5–1s (60–120Hz). For long runtimes it is advisable to increase the bundling rate for late eluting peaks. The data system detects the start end end of peak based on – at least – following parameters (terms may differ!):
  • Noise threshold: changes below this value are considered random noise and do not trigger peak detection.
  • Upward- / downward slope detection: the data system fits a couple of data points to a function (Polynomial, smoothing spline, Savitzky-Golay, …) and calculates the first derivative at each time point. If the derivative is positive and above the threshold = start of peak; if the slope is negative and below the threshold = end of peak. For a Gaussian peak upward- / downward thresholds would be the same, but in chromatography peaks are asymmetrical. Some data systems correct for that by using more slices if the slope is negative or even change to a different fitting algorithm.
  • Baseline drift: mainly important for gradient elution; nice question, next question.
  • Area threshold: values below this value are not followed.
Once you have set up integration parameters all these calculations are running behind the scenes. It’s always possible that the method’s integration parameters will fail. A tricky issue are peaks in the lower concentration range. But it is also possible that the algo fails for high peaks, mainly by problems with the downward-slope threshold. If you have random noise (±) which by chance is increasing the signal the data system will ‘see’ that as the end of the peak and the baseline is drawn too early. If on the other hand the random noise is lower than expected by the algo, the integration will continue – the baseline will be drawn to a too late time point.
It’s important to realize that there is no ‘right’ integration for any given peak. Even if you export the chroma­to­gram’s raw data (peak slices – detector’s response is not stored!) and try to obtain the same peak area with another data system it will be almost impossible because the internal algorithms are different (and not documented in the manual – just my 2¢).

» Is it accepted by USFDA?

Yes. You have to have an SOP in place and report which chromatograms were reintegrated (why, by whom, when: all the usual stuff needed for an audit trail). See also an article by John Dolan.1

» Also how does one know if the manual integration done is proper or not.

Our brain is a perfect pattern-recognition-system. IMHO any experienced chromatographer will be able to draw baselines better than the CDS’ ones. An excursion into history: before integrators were used in chromatography, strip chart recorders were used. The baseline was drawn by a ruler and the peak height measured. The first integrator was introduced in 1968 (Hewlett Packard 3370), but manual ‘integration’ was the rule until the mid 1970ies. If an inspector has some problems with manual integration remind him/her on drugs getting approval earlier. Would they question these results?

All chromatograms should be reviewed and the integration corrected if necessary. The only textbook on chromatographic integration methods2 states in the introduction:

No analytical report should be accepted unquestioningly.
An integrator might draw a baseline in the wrong place or separate two peaks by skimming a tangent where a perpendicular would better, or vice versa. It is important to inspect these features on the actual chromatogram and confirm their correctness […]

One thing is important: The review has to be done before () concentrations are calculated. Changing integration of a peak in order to bring a calibrator / QC to the expected value (e.g., make a batch valid which would be rejected otherwise) or a predose concentration <LLOQ would be clear evidence of fraud.

See also this presentation on manual integration by Bob Di Rienzo and Melinda Jacobson from The NELAC Institute's Forum for Laboratory Accreditation, Newport Beach, CA, January 2008.

Below a nice example3 (LC/MS-MS, risperidone, protein precipitation, dilution factor 8, API 4000, software Analyst 1.4.1, n=10); 1 ng/ml and 0.1 ng/ml (LLOQ):
  • Automated (smoothing factor 1, bunching factor 2):
    CV 6.5% (1 ng/ml), 15.1% (0.1 ng/ml)
  • Manual correction (one analyst):
    CV 6.3% (1 ng/ml), 11.1% (0.1 ng/ml)
  • Manual correction (ten analysts):
    CV 5.2% [3.8%-6.8%] (1 ng/ml), 12.8% [6.9%-16.0%] (0.1 ng/ml)

If the employees were well trained the values in the region of the LOQ were rather better than with an automated integration.


Look at the CV ranges! Even ‘bad’ analysts got CVs close to the automated integration, but ‘good’ ones outmanouvered silicon-brains with great ease.


  1. Dolan JW. Integration Problems. LCGC North America, Oct 1, 2009.
    online
  2. Dyson N. Chromatographic Integration Methods. The Royal Society of Chemistry 1998 (2nd ed.):p10.
  3. H Kirchherr H. Data Evaluation in LC-MS. In: Kuss H-J, Kromidas S (eds.): Quantification in LC and GC. Wiley 2009:p243–59.

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-07-29 21:07

@ Helmut
Posting: # 5700
Views: 6,362
 

 Manual integration

» If an inspector has some problems with manual integration remind him/her on drugs getting approval earlier.

Ah, you mean something like "If we'd had properly working integration systems back in the 60'ies then you wouldn't have been born" or something along those lines?

EM.

if (3) 4

Best regards,
ElMaestro

"(...) targeted cancer therapies will benefit fewer than 2 percent of the cancer patients they’re aimed at. That reality is often lost on consumers, who are being fed a steady diet of winning anecdotes about miracle cures." New York Times (ed.), June 9, 2018.
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-07-30 01:30

@ ElMaestro
Posting: # 5701
Views: 6,421
 

 History

Mon capitaine!

» Ah, you mean something like "If we'd had properly working integration systems back in the 60'ies then you wouldn't have been born" or something along those lines?

Almost. :-D
But due to mercy of a (little) more recent birth I hadn’t to fiddle around with strip-chart recorders, scissors and weighing of clipped peaks. My first HPLC was a Spectra Physics SP8000 (nice: ternary gradient, column oven, and roughly twenty 12" circuit boards making up the data system). We had one of these HP integrators connected to a GC as well. Both printed ‘integration marks’ on the chromatogram – no baseline. Theoretically it was possible to change the settings and look at new marks, change again, and get stuck in something like:

10 REM some stupid BASIC code without a BREAK
20 GOTO 10
30 END

If we were not satisfied with the integration we used a ruler and a sharpened pencil (for GLP conformity I would recommend a 0.5 mm permanent marker instead). First line connecting the integrator’s marks, second the ‘eye-ball-baseline’. Both heights measured, and Anew = Aold × Hnew / Hold. Voilà!

I think the first systems where you could actually see the baseline on screen were LDC/Milton Roy’s CCM and the Altex Scientific 324 introduced in the early 1980s. Tape cassettes! Proprietary BASIC!! Moving around the chromatogram by means of arrow-keys (the mouse wasn’t invented yet…)!!!

[image]BTW, this was also the age of the infamous WISP 712 (Waters Intelligent Sample Processor) – which was renamed by my friends at the Sandoz Research Institute to WUSP (Waters Unintelligent Sample Processor) because it was notorious for trying to slam the needle through the bottom of sample vials or into the back of the analyst’s hand changing the 48 (!) sample tray.

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-07-30 23:46

@ Helmut
Posting: # 5704
Views: 6,303
 

 LSB - evil, terrible and annoying

Hi,

» Let’s have a look how peak integration in chromatography is performed.

...and how data are captured electronically:
Data acquisition is typically done on a voltage range (e.g. 0 - 5 mV); data acquisition hardware has a property related to resolution that is measured in bits.
The number of bits determines the minimum change in voltage that can be detected. For example, an 8-bit data acquisition system splits the capture range into 28 = 256 units each having the dimension of the smallest theoretically detecable amount of voltage change; if the capture range is 0-5 mV then the smalles change in voltage that can be detected is 5 mV/256 = 0.0195 mV (corresponding in theory1 to the least significant bit, LSB).

Practical proof of the presence of an LSB:
Take any captured chromatogram and enlarge any portion of it with your software. A piece of baseline is fine for this purpose. Superficially it looks like a flat line with noise. Enlarge even more and sooner or later a distinct saw-tooth like pattern becomes evident. The values that are captured jump between discrete levels - these levels represent the LSB.

The consequence of it all is that for small peaks, the LSB can in some cases be of "signficiant" magnitude. Conversely, as peaks get larger, the digital signal can for all practical purposes be considered continuous. Most software peak detection Al Gore Rhytms are based on such an assumption, mainly because it is terribly difficult to write anything meaningful that is based on a signal that makes quantum leaps, and if such software would be functional noone would be able to tell how small a peak should be before algo1 should be used in stead of algo2 in a given run. And people would prolly start wars when discussing the same issue between runs. And some intelligence agency would prolly want that algo and prevent it from being published if it became a reality.

As computers and data acquisition boards get better (more bits!) the significance (no pun intended) of the LSB is in practice slowly disappearing, fortunately. But it is still in 2010 A.D. showing its ugly face with a few drugs out there.

I agree very much with the essense of HS' post: Common sense and human eyeballing is better by a large factor than any detection algo.

Best regards
EM



1: Real life is cruel. The LSB is never as "good" as in this theory due to factors like electronic noise, voltage drift and more. A 16 bit system is often only 14-15 bit in practice.

if (3) 4

Best regards,
ElMaestro

"(...) targeted cancer therapies will benefit fewer than 2 percent of the cancer patients they’re aimed at. That reality is often lost on consumers, who are being fed a steady diet of winning anecdotes about miracle cures." New York Times (ed.), June 9, 2018.
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-07-30 20:24

@ sagark
Posting: # 5703
Views: 6,417
 

 Bad integration: Example

[image]Dear Sagar,

a picture tells more than a thousand words.

Following my previous post – which was a little bit theoretical – let’s look at an example (not from the last century; I received it a couple of days ago):

This chromatogram shows the low CC standard of a chiral separation of drug X; first peak is the (inactive) d-enantiomer and second the (active) l-enantiomer. IMHO the analyst made a couple of mistakes…
  • He set the integration window in such a way that the first enantiomer was not integrated at all. Debatable from a pharmacokinetic point of view (what if a regulator asks you later on to evaluate the other enantiomer as well?) – but also a killer for the integration algo (see below).

  • He relied on automatic integration (black area), which ended too early due to the small shoulder at ≈9 min. If you compare the noise before and after the peaks it’s clear that this shoulder is most likely not an interference, but random noise.

  • Even if the second peak would have been correctly integrated by his method (▬▬), the area would be underestimated.

  • Only if the first peak is also integrated (▬▬), the peak areas would reflect the enantiomeric ratio of 50:50 of the racemate (which was used in calibration).

  • We are currently investigating whether the CRO did not review chromatograms (bad), or had an SOP in place which did not allow manual reintegration (even worse).

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
sagark
Junior

2010-07-31 12:57

@ Helmut
Posting: # 5705
Views: 6,226
 

 Bad integration: Example

Dear members,

Thank you very much for your detailed explanations. Aspects for integration are much clear for me.

Dear HS, in the example you mentioned, can we say that the method was not properly developed. Ideally one would have expected a method which separates the RT of the two enantiomers by a consideration time to minimise overlapping. You said the first peak was not integrated at all. Does that mean the separation was achieved but not quantification of the first peak isomer was achieved? Please guide

Sagar
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-08-01 02:09

@ sagark
Posting: # 5706
Views: 6,248
 

 Bad integration: Example

[image]Dear Sagar!

» […] in the example you mentioned, can we say that the method was not properly developed. Ideally one would have expected a method which separates the RT of the two enantiomers by aconsideration time to minimise overlapping.

Especially for a chiral method I would say that peaks were sufficiently separated. Pharmacopoeias (like USP) call for a chromatographic resolution Rs of ≥1.5, whilst FDA (Reviewer Guidance - Validation of Chromatographic Methods, Nov 1994) recommends an Rs >2 (see also this post). At the right is the chromatogram of a high aqueous standard, showing a resolution of ≈2.5.

» You said the first peak was not integrated at all.

Yes; IMHO that’s a flaw in the integration method.

» Does that mean the separation was achieved but not quantification of the first peak isomer was achieved?

Exactly. He excluded the first peak. Therefore even for the aqueous standard with a perfect baseline the area of the second enantiomer is systematically too low. That’s why I stated in my previous post:
» » Even if the second peak would have been correctly
» »   integrated by his method (▬▬▬), the area would be underestimated.

No big deal for this particular method (BE of one the second enantiomer to the racemate; no in vivo interconversion expected), since all chromatograms are evaluated in the same way. But think about another drug where enantiomers are interconverted. The ratio of the enantiomers will not be the same in subject’s samples compared to CCs and QCs. Furthermore the ratio will change within the time profile… Only the perpendicular drop integration (▬▬) in my first example will give correct results.

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-08-12 20:11

@ Helmut
Posting: # 5774
Views: 6,200
 

 Better integration? Example

Dear all,

I played around with the last example. Since I had no access to the peak slices, I fired up SigmaSCAN Pro 5 to extract the chromatogram from the report. Next I used PeakFIT 4.12 with the EMG (exponentially modified Gaussian) model to deal with tailing peaks:

[image]

Interesting results. The tangential integration method of the original chromatogram (red line) overestimates the tailing of the first peak. The system has no other choice, because the baseline is established with the end of the second peak. The enantiomeric ratio is given with 54.2%/45.8%. The perpendicular drop at the valley (blue line) would give a ratio of 51.7%/48.3%. If we calculate the areas of EMG-fitted peaks we get 52.0%/48.0%. Assuming a theoretical ratio of 50%/50% for the racemate, we get biases of 8.35%, 3.39%, and 4.01% for the three methods (since I don’t know the enantiomeric purity of the standard, this claim is rather specultive). But it’s clear, that the tangential integration performs worst and should be avoided.

Unfortunatelly no (!!) current commercial CDS allows peak fitting. Merck/Hitachi’s mid-1990s D-7000 HPLC system manager (HSM v4.1) allowed deconvolution of two merged peaks based on the EMG model. Hillebrand* showed for a large combination of parameters (resolution, tailing, theoretical ratios of merged peaks, noise) that both the HSM and PeakFIT gave unbiased results in all cases, whereas all conventional integration methods performed worse in all cases. Though SigmaFIT is able to import AIA chromatography files, it’s unclear whether it’s application would be acceptable in a regulated environment.

A close-up of the intersection is given below:

[image]


  • Hillebrand H. Deconvolution. In: Kuss H-J, Kromidas S (eds.): Quantification in LC and GC. Wiley 2009:p123–51.

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
ElMaestro
Hero

Denmark,
2010-08-12 23:14

@ Helmut
Posting: # 5775
Views: 6,059
 

 Better integration? Example

Hi HS,

interesting stuff - I'd love to be a programmer making new ways to deal with this.
I find your example quite informative, thanks for posting it.

Let me ask a broad question here, in relation to one of your points:

» But it’s clear, that the tangential integration performs worst and should be avoided.

It is clear that in this case the Tangential Method (TM) gives a relatively large error on a peak estimate. However, for BE purposes this might or might not be a worry since we are interested in T/R. T and R could in principle be "a lot wrong" as long as we have some trust in T/R itself; this is implicitly how it works for e.g. AUC with the noncomp. model, typically with over-estimated area before the top and under-estimated thereafter. We just pray that the error we introduce on AUC(T) is percentually the same as on AUC(R).
When the true T/R is not 1 one could argue that there could (I am not saying there is) be an issue, regardless of whether we talk chromatograms or AUCs.

I am not aware of this issue being discussed anywhere in the scientific literature, and the bioanalysts I have asked about it also don't have any information.

From your example, I would think that above all that the problem with TM is likely to be bigger if we use it in studies to characterise for example the half-life of a new chemical entity or similar.

Let me hear your or anyone elses thoughts, please.
Finally, I solemnly promise I will not make a lengthy post ever again until next time.

if (3) 4

Best regards,
ElMaestro

"(...) targeted cancer therapies will benefit fewer than 2 percent of the cancer patients they’re aimed at. That reality is often lost on consumers, who are being fed a steady diet of winning anecdotes about miracle cures." New York Times (ed.), June 9, 2018.
Helmut
Hero
avatar
Homepage
Vienna, Austria,
2010-08-13 13:43

@ ElMaestro
Posting: # 5780
Views: 6,156
 

 Better algorithms / more awareness of analysts

Hi ElMaestro!

» I'd love to be a programmer making new ways to deal with this.

Well, might be nice homework. But PeakFIT comes for just 600$...

» Let me ask a broad question here, in relation to one of your points:
»
» When the true T/R is not 1 one could argue that there could (I am not saying there is) be an issue, regardless of whether we talk chromatograms or AUCs.

[image]Or the estimation of λz, or AUCtlast (if tlast,T tlast,R = the apples-and-oranges-story), or, or, …

» I am not aware of this issue being discussed anywhere in the scientific literature, and the bioanalysts I have asked about it also don't have any information.

Well this thread started with the issue of manual integration. In my experience analysts are so scared about findings in an inspection, that they (erroneously) believe manual integration is forbidden data manipulation. Remember my first example (same study, but low QC sample). Here it is not a question of “perpendicular, tangential, exponential skimmed” integration but simply wrong integration. At T/R=1 of course it would mean out, but for QCs it may lead to rejection - or false acceptance – of a batch. IMHO, bioanalysts should concentrate more on the basics of data evaluation in chromatography rather than blindly accept the results of a silicon-brain. The separation in my example was OK (Rs 2.5) and the tailing pretty good for a chiral method. There is an abundance of literature on the EMG-method. Only analysts can force manufacturers of CDSs to implement better algorithms.

» From your example, I would think that above all that the problem with TM is likely to be bigger if we use it in studies to characterise for example the half-life of a new chemical entity or similar.

Yes, if you consider individual samples. But the issue of batch acceptance/rejection is applicable to all studies.

» Finally, I solemnly promise I will not make a lengthy post ever again until next time.

Same with me.

Cheers,
Helmut Schütz
[image]

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes
keshav khude
Junior

2010-08-13 09:11
(edited by Ohlbe on 2010-08-13 09:51)

@ sagark
Posting: # 5777
Views: 6,026
 

 Manual integration

Dear Sagar,

Manual integration is required only particular chromatogram/individual sample id during the batch integration, if you are maintaining proper documentation for initial integration and reintegration with proper reason and same to reflected in respective SOP, so it is accepted by any auditor.

Thanks,
Keshav D.Khude


Edit: Full quote removed. Please delete anything from the text of the original poster which is not necessary in understanding your answer; see also this post! [Ohlbe]
Back to the forum Activity
 Thread view
Bioequivalence and Bioavailability Forum |  Admin contact
18,416 posts in 3,912 threads, 1,175 registered users;
online 18 (1 registered, 17 guests [including 8 identified bots]).

The greatest shortcoming of the human race is our inability
to understand the exponential function.    Albert Bartlett

The BIOEQUIVALENCE / BIOAVAILABILITY FORUM is hosted by
BEBAC Ing. Helmut Schütz
HTML5 RSS Feed