Bioequivalence and Bioavailability Forum

Hi ElMaestro,

❝ ❝   — Our brain is an excellent pattern rec­og­​ni­tion machine.

❝ ... and this widely recognised fact also explains why regulators over the years have been so happy to see manual integration in chromatography.

… though they realized that sometimes it still is necessary (ICH M10 Section 3.3.6). Yep, has to be documented.

There is an invited comment about the comparison of r² and visual inspection (reference #4 of the post above). Since it is behind a paywall, important parts:

The implementation of the algorithm in WinNonlin (which is essentially unchanged today) is an aid to provide an initial estimate for the user who can then manually adjust the terminal slope calculation for each individual as appropriate. It was realized at the time of the original development that the adjusted r² approach might not always select the same time points that a user would select by visual inspection, which is precisely why an interactive graphical tool was included with WinNonlin. In this sense, the adjusted r² algorithm was simply an aid to increase efficiency of the analysis of pharmacokinetic data. Over almost 30 years, the adjusted r² algorithm as implemented has been unchanged, and we are not aware of any alternative algorithms that would significantly improve the accuracy or speed of the estimation method. However, we agree with the author that for some profiles the adjusted r² algorithm appears to have some bias, and we would welcome suggestions on either how to identify profiles the adjusted r² algorithm may not perform well on, improve the adjusted r² algorithm to decrease the bias, or propose an alternative algorithm that is more accurate and equally as efficient.
[…]
An important operational point to note is that the time required for analysis is an important consideration in use of any algorithm. The r² and adjusted r² algorithms take a fraction of a second to implement for a single pharmacokinetic profile. The time required to implement the visual inspection algorithm proposed by the author is not stated; however, one can assume that it might be approximately 60 seconds for a single pharmacokinetic profile. This suggests that the visual inspection algorithm is at least 60-fold less efficient than either the r² or adjusted r² algorithms. Even if one accepts the author’s assertion that the visual inspection algorithm has less bias across all pharmacokinetic data analysts, the enormous loss of efficiency for a small (<10%) improvement in bias would argue that the visual inspection algorithm is not an improvement over either the r² or adjusted r² algorithms.

(my emphases).

IMHO, this line of argument is convoluted. r² was developed as a tool to get initial estimates. Since it is known to be imperfect, the user can (and should according to WinNonlin’s User’s Guide) make adjustments. So far, so good. This is exactly what I do for ages as well.
But then the authors speak of ’efficiency’ (speed?). Oh dear, 60 seconds per profile? 🧠💤
Further, with my 30+ years of experience with the software I can say that the algo works remarkably well for IR and DR formulations. Sometimes I have to exclude an increasing last concentration. Remember that at the LLOQ A&P can be 20%. Therefore, an apparent increase might be pure chance. However, not excluding it, would bias λ_z with all of its ugly consequence (larger extrapolated AUC).
For formulations with flat profiles (CR, patches, etc) the algo fails quite often. The authors prefer not to talk about that.

---

• Weiner D, Teuscher NS. Comments on D.A. Noe’s papers on noncompartmental pharmacokinetic analysis: Performance characteristics of the adjusted r² algorithm for determining the start of the terminal disposition phase and comparison with a simple r² algorithm and a visual inspection method https://onlinelibrary.wiley.com/doi/10.1002/pst.1979 and Criteria for reporting noncompartmental estimates of half-life and area under the curve extrapolated to infinity https://onlinelibrary.wiley.com/doi/10.1002/pst.1978. Pharm Stat. 2020; 19: 113–4. doi:10.1002/pst.1999.

Dear ElMaestro,

[OffTopic]

❝ ❝   — Our brain is an excellent pattern rec­og­​ni­tion machine.

❝

❝ ... and this widely recognised fact also explains why regulators over the years have been so happy to see manual integration in chromatography.

Well, you have to acknowledge that some labs gave them excellent reasons for that reaction... Where inspectors' brains recognised patterns of biased integrations to validate failed runs. As usual, some people f*cked up, and the rest of the profession is now paying the price for it.

I have on my computer the slides of an old presentation from Sciex in 2003, comparing manual integrations and the 3 integration algorithms in Analyst 1.4 (Analyst Classic, MQII, IQA) using artificial (synthetic) chromatograms, which concluded that:
- Human integration performs better than any of the machine algorithms
- All algorithms provide consistency, human eye is not as consistent.

It also showed that Analyst Classic performed the worst on most tested aspects (except for the integration of split peaks). Interesting, considering the very, very large number of labs using it.

[/OffTopic]