What does Bayesian software do for vancomycin kinetics?

[Monsterdaddy]

Hi Folks, anyone here use a PK software program with Bayesian "features?" I am trying to get a handle on what Bayesian estimation actually does for vancomycin dosing. Depending on the PK software, it should be able to estimate the steady state AUC, peak and trough levels by using a population Vd estimate based on the patient's characteristics from a single sample measurement (curve fitting by interpolation).

I can do the same thing with a simple Excel spreadsheet and the Solver Add-In. The exception is the population Vd estimate. But I am questioning why I even need a Bayesian estimate because my Excel calculates patient specific PK parameters. I would think patient specific is better since it reflects a patient who may already be sick and thus is a departure from population measurements. (And heck, even I can make an informed guess that the initial Vd is probably between 0.6 and 0.8 95% of the time.)

So I am hoping for more insight from people who are familiar with Bayesian estimates as to the true benefits of this. Thanks so much!

---

Members don't see this ad.

 

[lord999]

/s Surely you understood that Melanie Jordan is a complete idiot who can't do basic addition, and if they gave her a PhD in pharmacokinetics, that should give you an idea of how useful the subject is in practice. /s

Short Summary:
Your instincts are right. Don't be too cute, take the trough and peaks and then get a corrected crude estimate. Vancomycin is not that NTI in most cases to matter. Also, whoever did the population PK didn't have their thinking cap on when it came to application.


Long answer:
Honestly though, there are extremely few edge cases (burns come to mind) that you'd do the mathematics. Bayesian is a guess based on some prior, and you don't get enough information to update the prior that you wouldn't get out of a nonogram anyway. Leave Bayesian to population PK rather than patient level.

To a upper secondary student, I'd explain Bayesian like this.

1. Guess a whole number between 1 and 100. Each number has a 1/100 probability of being selected. You get to know the answer as right, low, or high.

This is the following strategy I chose:
Guess 1: 50 - High
Guess 2: 25 - Low
Guess 3: 36 - Low
Guess 4: 40 - High
Guess 5: 38 - High
Guess 6: 37 - Correct

6 guesses.

2. Ok, how many guesses did it take you to get the right number? How did you shift your strategy to account for the right number?

When you choose a guess, that's a Bayesian prior. As you gain more information, that Bayesian prior becomes better calibrated, but you have to get information about that guess on how close you thought you were to try the next time. That "closeness" is called the posterior. You try to take strategies to minimize the posterior so that your prior guess gets closer and closer until it basically sticks at a number.

There are several different ways (algorithms) to get here. The strategy of halving is actually named as the Newton-Raphson. There's actually a more efficient guess method called Gibbs sampling that does this when there are multiple sets of numbers to guess (lottery style).

By the way, the philosophy underlying both the standard frequentist method and this Bayesian guess and correct method is something harder called likelihood. The basic way to explain likelihood is the following, if owlegrad, giga, BMB, and others took guesses at the number and followed their own strategies, the number of guesses for each person over the entire subforum would have specifically a normal distribution, and where the mean and median are, that happens to be the likely number of guesses for any 100 number guess (it's actually between 6 and 7 guesses if done right), and that likely number in theory is what underpins likelihood and its robust work as maximum likelihood estimation.

Why I bring this up is that the cost of collecting that much information is not patient friendly or useful, you are better off doing the norm and taking the trough and peaks and dealing with it analytically rather than using models. Again, there sometimes is better information like burns where you have to do it the hard way, but you know when those cases occur.

Bayesian estimates are only marginally useful for establishing better baselines in complex cases where there are major interfering factors that are unknown unknowns (in math terms, variability is extremely heterogenous, in clinical terms, this patient is a trainwreck and you don't know the extent of the perfusion and distribution problems).

But then again, I'm liable to get some ID pharmacist on my case for this, but even in the journals, patient application of population PK has always been very tenuous at best besides basic therapeutic/toxic floors and ceilings.

 

[Monsterdaddy]

That's what I'm thinking. In fact, I would argue Bayesian is even more useless in complex cases because they are the statistical outliers.

I got to thinking about this whole Bayesian estimate thing because it's applied after you obtain one level from a patient. If you have that level and you know the doses and their times of administration, you can calculate PK that is patient specific. But folks want to use that level and apply a population Bayesian estimate -- why??? Do you not trust the results? Because you will ignore perfectly good patient specific results and substitute a Bayesian estimate.

And in the case of vancomycin AUC kinetics, the Bayesian estimate is just a guess for Vd. Well, I can guess Vd too! And here's the biggest kicker. In most cases, if you calculate AUC by picking a range of Vd (say 0.5-0.9) all the AUC's you calculate are statistically the same (a lucky quirk but the results are valid -- true Vd is someone inside the range!)

Anyways, I'd love to hear more insight on Bayesian. Maybe I am missing some benefit.

 

[owlegrad]

So I am not familiar with Bayesian but I will say it drives me nuts when students argue with me about how other preceptors have taught them how to dose Vancomycin. I literally could not care less what someone else taught you, I am telling you to only use population data for the initial dosing and once you have a trough use the patient specific data to guide the dosing after that. I am shocked how many students persist in using the average Vd rather than the one actually demonstrated, in real life, by the patient. Do people have no critical thinking skills at all?

And then to defend it as “that’s what my last preceptor told me to do!”. Even if I believed that (I don’t) why would that make it acceptable for you to do it wrong?

 

[Monsterdaddy]

LOL owlegrad that's a good analogy! BTW, how are you calculating Vd at your facility? I would guess two sample measurements for actual half-life and then using the steady state formula to back into Vd?

 

[owlegrad]

LOL owlegrad that's a good analogy! BTW, how are you calculating Vd at your facility? I would guess two sample measurements for actual half-life and then using the steady state formula to back into Vd?

Hmmm. I guess I have to blow up all my street crew and admit I use Global RPh. They have a calculator for initial dosing and a separate calculator for dosing based on levels.

I would love to go back to basics and relearn the pharmacokinetics basics but in truth I haven’t needed them yet. Global RPh has served me very well so far.

 

[lord999]

So I am not familiar with Bayesian but I will say it drives me nuts when students argue with me about how other preceptors have taught them how to dose Vancomycin. I literally could not care less what someone else taught you, I am telling you to only use population data for the initial dosing and once you have a trough use the patient specific data to guide the dosing after that. I am shocked how many students persist in using the average Vd rather than the one actually demonstrated, in real life, by the patient. Do people have no critical thinking skills at all?

And then to defend it as “that’s what my last preceptor told me to do!”. Even if I believed that (I don’t) why would that make it acceptable for you to do it wrong?

/s No, because critical thinking is not necessary or desirable in your average pharmacist. Makes them easier to manage. /s

More seriously, its fairly horrifying to watch pharmacists present "research" and " statistics " as it usually demonstrates their lack of literacy and numeracy. Bayesian is a statistic honed on a belief. If you are presented with better information in the posterior sense (and actual patient numbers probably being the absolute best), the hell you tell me you don't make use of it...sigh, idiots.

It is so great that the systems and current pharmacopeia are so wide in their safety indexes that most would have to commit major and multiple serial blunders to harm someone. Doesn't give me much faith in most pharmacists actually helping anyone, but hey, we're in the George Jetson future.

Please use the nonograms and automatic calculators. There's no general reason to risk doing this manually. That said, it's pathetic that we're at the level of cargo cult with respect to any chemical or pharmaceutical knowledge now.

And to Monster, I had the same PK professor as you did. I've got plenty of stories about how seriously she screwed up exams and basic math. Last I heard, she was put in charge of Pharmacal Science, which is probably the worst appointment made in my career.

 

[Monsterdaddy]

ROFL, that's nothing to be ashamed of. Are you using this one for dosing by levels?

NON-STEADY STATE KINETICS -DOSING BY LEVELS

I verified that is 100% accurate. It only limitation is requiring 3 measurements around a single dose (or just two after a new start). I really don't know why people don't use that more often, you basically get accurate and patient specific data as early as Day 1. In fact, with this calculator you could skip initial dosing by ordering just a mg/kg loading dose then use the calculator for the maintenance dosing based on your results.

In my calculator I took it a step further so that you could have multiple doses between the first measurement (or new start) and then have the last two measurements.

 

[Monsterdaddy]

@lord999, I was trying to IM you a huge LOL on our kinetics professor but couldn't find that feature on this forum. It is still sooooo true. I had to take her final early and afterwards I told her she made it way too tricky but she blew me off. Well the average raw score of the class wound up being ~65! She had to give a ~20-25 point curve to stop from failing half the class so my raw A- became an A+++ haha.

But your point about close minded pharmacists is so unfortunately true.... I don't even want to think about how many times I argued with a pharmacist that changing a dose from 1 gm Q8 to 1.5 gm Q12 is the exact same daily dose -- they think with a lower trough it will head off AKI.....

 

[lane one]

ROFL, that's nothing to be ashamed of. Are you using this one for dosing by levels?

NON-STEADY STATE KINETICS -DOSING BY LEVELS

I verified that is 100% accurate. It only limitation is requiring 3 measurements around a single dose (or just two after a new start). I really don't know why people don't use that more often, you basically get accurate and patient specific data as early as Day 1. In fact, with this calculator you could skip initial dosing by ordering just a mg/kg loading dose then use the calculator for the maintenance dosing based on your results.

In my calculator I took it a step further so that you could have multiple doses between the first measurement (or new start) and then have the last two measurements.

It isn't utilized because 3 measurements is not realistic for the average "empiric coverage" patient. We have enough lab draw refusals as it is.

 

[WVUPharm2007]

I haven't done it for 10 years, but I recall thinking that adjusting vanc doses was one of the most overrated things in difficulty in all of pharmacy while I was a hospital staffer. People made it out to be fricking rocket science.

 

[PharmDBro2017]

I haven't done it for 10 years, but I recall thinking that adjusting vanc doses was one of the most overrated things in difficulty in all of pharmacy while I was a hospital staffer. People made it out to be fricking rocket science.

It's still just as simple now, as it was then.

Globalrph + a smidge of critical thinking ftw

 

[PolarPop]

The other problem with this push to AUC/MIC based dosing is that the vancomycin MIC part of the equation that you get from the lab is notoriously inaccurate. You could have the same isolate in three different hospitals with three different automated susceptibility machines and get three different results.

 

[Monsterdaddy]

The other problem with this push to AUC/MIC based dosing is that the vancomycin MIC part of the equation that you get from the lab is notoriously inaccurate. You could have the same isolate in three different hospitals with three different automated susceptibility machines and get three different results.

This is true which is why the new ASHP guidelines suggest to assume MIC=1 and just target 400-600 AUC.

See: https://www.ashp.org/Pharmacy-Practice/Policy-Positions-and-Guidelines/Draft-Guidance-Documents

I want to give some Commentary on the draft which I think overpromotes Bayesian software but I wanted to get a better handle on it, hence this thread. Still hoping to get some comments from people who have used Bayesian software.

 

[Monsterdaddy]

BTW, don't use this GlobalRPh calculator:

PHARMACOKINETIC (DOSING BY LEVELS FULL VERSION)

This calculator will not be accurate because it uses steady state formulas and patients may never be in steady state due the accumulation process. I've tested it and although it calculates actual elimination half-life (almost) correctly, and also the peak and trough for that dose alone, the Vd will not reflect steady state which is significant enough to throw the actual steady state peak/trough and AUC.

 

[Monsterdaddy]

It isn't utilized because 3 measurements is not realistic for the average "empiric coverage" patient. We have enough lab draw refusals as it is.

That is a true issue. However the solution is so simple it is often overlooked, for example see this again:

NON-STEADY STATE KINETICS -DOSING BY LEVELS

People don't realize the most amazing advantage of this calculator -- random levels can be used. Nurses and techs can screw up administration and draw times because the calculator won't care.

It's absolutely true patients hate excess needle sticks. To minimize this, we can simply order random vancomycin levels with other blood draws. For example, most hospitals have scheduled morning labs -- add a random vancomycin level to that and we have already reduced the incremental draws by 1. And throughout the day there may be other labs orders to piggyback on (i.e. a potassium level) and it can ordered after the fact since most labs retain blood samples for several days, thus we have eliminated the second incremental draw.

("Random" level is such a misnomer, if you have the draw time it's no longer random we can calculate the time relationship to the dose -- just like a trough.)

So for new starts if we order one random level with AM labs and just one true incremental needle stick we basically have the three levels required to determine patient specific PK parameters on Day 2 (the first measurement in the calculator isn't needed since the level will be zero). Beats Bayesian any day 🙂

 

[lane one]

That is a true issue. However the solution is so simple it is often overlooked, for example see this again:

NON-STEADY STATE KINETICS -DOSING BY LEVELS

People don't realize the most amazing advantage of this calculator -- random levels can be used. Nurses and techs can screw up administration and draw times because the calculator won't care.

It's absolutely true patients hate excess needle sticks. To minimize this, we can simply order random vancomycin levels with other blood draws. For example, most hospitals have scheduled morning labs -- add a random vancomycin level to that and we have already reduced the incremental draws by 1. And throughout the day there may be other labs orders to piggyback on (i.e. a potassium level) and it can ordered after the fact since most labs retain blood samples for several days, thus we have eliminated the second incremental draw.

("Random" level is such a misnomer, if you have the draw time it's no longer random we can calculate the time relationship to the dose -- just like a trough.)

So for new starts if we order one random level with AM labs and just one true incremental needle stick we basically have the three levels required to determine patient specific PK parameters on Day 2 (the first measurement in the calculator isn't needed since the level will be zero). Beats Bayesian any day 🙂

You stated in your previous post that you could achieve accurate patient specific kinetics as early as day 1. That is not possible from utilizing "AM labs". I get your point about piggybacking labs onto other blood draws, but if its going to take 48-72hrs to get enough data that way, the traditional trough collection at steady state looks the same to me.

 

[Monsterdaddy]

@lane one, you are correct on AM labs since it is Day 2 but I was referring to obtaining Day 1 PK from ordering a single random level after the first dose. For example, PT is administered a ~25 mg/kg loading dose at 9 AM and you order a random level anytime at least 1 hour after the infusion is completed, i.e. 3 PM.

A single sample on Day 1 has proved very helpful in very elderly patients with poor CrCl but needs to be therapeutic ASAP. The protocol in most hospitals tend to recommend very low doses @q24 for these patients (because Cockcroft-Gault tends to underestimate actual clearance in this population) so redosing in under 12 hours sure beats waiting 48+ hours for a trough before 3rd dose. And two samples on Day 1 would be even better if we're lucky.

But the big picture is that I think pharmacists need a paradigm shift in our thinking of maintenance dosing. Instead of thinking the regimen ordered is good for days while we watch SCr for changes, ordering random levels with AM labs daily will spot an accumulation trends early and thus redosing will avoid AKI or subtherapeutic levels. I am certain levels are far more predictive of clearance changes than SCr.

I've been playing around with a Bayesian software program and I am very disturbed by the results. They are not just guessing at Vd, they are guessing at half-life, even when I input two sample measurements the program will override the actual half-life and substitute it's own guess. I am contacting the vendor to question these results.

If I had to guess, it seems the Bayesian estimate is trying to forecast the PK of the patient several days out -- this is where I got to thinking about pharmacists wanting regimens that last for days. Of course, the flaw is that current dosing and evaluation will be wrong and lead to unnecessary higher dose adjustments.