Subscribe
Share
Share
Embed
Communicable takes on hot topics in infectious diseases and clinical microbiology. Hosted by the editors of CMI Communications, the open-access journal of ESCMID, the European Society of Clinical Microbiology & Infectious Diseases.
Communicable E57: PK/PD, part 2
[00:00:00]
Thomas: Hello and welcome back to Communicable, the podcast brought to you by CMI Communications, ESCMID's open access journal covering infectious diseases and clinical microbiology. My name is Thomas T�ngd�n. I'm an infectious disease physician and professor at Uppsala University, Sweden, and an associate editor at CMI Comms.
This is the second episode of our new series on optimal dosing and PK/PD called We Will Make You Love PK/PD. Today, we'll discuss evidence supporting dosing decisions and tools that help us move from best guess dosing to dosing supported by PK/PD data and principles.
Thomas: It's a pleasure to introduce my co-host, Rekha Pai Mangalore, who is an infectious disease doctor at Alfred Hospital, Melbourne, and editorial fellow at CMI Comms
Hi everyone, and I'm so excited to be back. this episode we will be explaining concepts that many clinicians hear about all the [00:01:00] time but may not be fully across.
Rekha: So pop PK models, PK/PD modeling and simulations, probability of target attainment, cumulative fraction of response. They can all sound so technical, so abstract, and can be quite intimidating. So today we want to break them down and make them practical
I'm very excited to introduce our first guest, Iris Minichmayr.
Thomas: an assistant professor at the Department of Clinical Pharmacology at the Medical University of Vienna, and lead of the Clinical Pharmacometrics Group She completed her pharmacy studies at the University of Vienna and pursued her PhD at the Freie Universit�t in Berlin.
After a research stay at the University of Queensland in Australia, Iris spent four years in the pharmacometrics group at Uppsala University, Sweden. Since a few years, Iris works at the intersection of clinical care and drug development.
Thomas: She's involved in the analysis and planning of clinical and translational studies using PK/PD modeling and simulations, but also in therapeutic [00:02:00] drug monitoring and antimicrobial stewardship activities. Welcome to Communicable, Iris.
Thank you, Thomas. I'm very happy to be here.
And I would like to welcome a fellow Australian, Phil Selby.
Phil is a senior lecturer in pharmacy practice at Adelaide University and a clinical pharmacist on acute leukemia and bone marrow transplant unit at the Royal Adelaide Hospital in South Australia. In twenty twenty-four, Phil completed his PhD on the prevention and management of cytomegalovirus and fungal infections in allogeneic stem cell transplant patients.
His specific interests include infections in patients with hematological Therapeutic drug monitoring, management of acute leukemias and allogeneic stem cell transplantation. Welcome, Phil
Thanks, Rekha. It's fantastic to be here.
Thomas: Now that we introduced ourselves, let's take a few minutes for an icebreaker question to get to know our guest a little bit more., if you were an antibiotic, which class would you belong to and why? What would your mechanism of action [00:03:00] be?
Would you have a large or small volume of distribution, and what's your half-life? Rekha, do you wanna start?
I have chosen doxycycline
why I like it is because it is a broad-spectrum, second-generation tetracycline antibiotic, a good volume of distribution, and long half-life, so you can do twice daily dosing or even once daily dosing,
It's now really famous because we use it for doxy PEP to prevent gonorrhea, syphilis, chlamydia, so atypical infections and, finally, even mycobacterial infections, so Mycobacterium marinum.
Rekha: So that's my absolute favorite antibiotic. Of course, you need to be mindful like any other antibiotic, sitting upright when you take a dose or, you know, don't go out and get photosensitive. But otherwise, just a nice clean little small tablet,
Thomas: I'd like to think about myself as a narrow-spectrum penicillin, like a drug that's straight to the point, very focused, low toxicity, saving most of the microbiome, which could be a good thing.
I guess a limitation might be the short half-life, but then [00:04:00] again, you can use prolonged or continuous infusion to get a higher drug exposure, a high probability of target attainment. that's gonna be my pick. H-how about you, Phil?
Yeah, so, it was a bit of a tough one coming up with what I would, be, but I went with linezolid, and that's 'cause I likes the challenge of a difficult case.
Phil: So, I work with acute leukemia bone marrow transplant patients. There's always something challenging there. it's a somewhat versatile antibiotic as well, lots of different indications. also got quite a few layers of complexity below what we thought it to initially have. So, we thought it was pretty standard dose for everyone, but it's actually a bit more complex in terms of its PK than we initially thought with, , linezolid TDM now being recommended.
And it's probably underused a bit due to some irrational fears regarding its toxicity as well. So I guess mechanism-wise, it does inhibit protein synthesis at the 50S ribosome, so it's pretty keen to get in early and to try and prevent problems before they occur. It's also got quite a large volume of distribution, so it gets to where the infection is, it gets to the crux of the problem, and it's got a moderately long [00:05:00] half-life, so it stays around long enough to get the job done, but not too long so that it makes things awkward,
Thomas: That's a very good choice. We're all very different. Love it. Yes. And we're curious to know about your choice, Iris.
So as a pharmacometrician, I work with code, so I'd probably also be a protein synthesis inhibitor because it stops the translation of bad data into bad results somehow. So maybe a modern next generation oxazolidinone.
Iris: So if this were a job interview, I'd probably say I am a broad spectrum reserve antibiotic, system-oriented, dependable under pressure, highly active against complex problems, hopefully clinically valuable. I have a large volume of distribution because I can easily move into diverse, difficult environments, and I don't really like being trapped in just one compartment.
my half-life probably also maybe 15 hours. Convenient to [00:06:00] handle but still adjustable. Although I would like to stress I don't require close monitoring. Would you like to hear about toxicity as well?
Thomas: I don't think you're a toxic Iris .
Iris: Yeah, I would say I'm mostly well-tolerated, although overexposure may occasionally cause confusion, and I can sometimes cause fatigue in modelers because my PK is entirely nonlinear
Thomas: That's pretty advanced. Excellent answer.
Rekha: So now that we know you guys a little bit more, I think we should get into this episode. I would like to start with you, Iris, given that you just gave a beautiful pharmacometrician answer to antibiotic choice, I'd like to ask you, what kind of quantitative methods, are in PK/PD? What, does one mean when we talk about quantitative methods?
Iris: So quantitative PK/PD methods quantify the relationship between drug concentrations in the body and drug effects over time. So we have, pharmacokinetics models, PK models on the one [00:07:00] hand that quantify how the body affects a drug. So a drug after administration of a dose. So we measure concentrations over time, for example, in plasma, and then we quantify the processes of drug absorption, distribution, metabolism, and excretion in the body using PK parameters.
for example, clearance or volume of distribution. And these parameters can then be used to support dosing decisions. PD models or pharmacodynamic models quantify how the drug affects the body. So they characterize the relationship between drug exposure and effect. For example, based on measurements of biomarkers or based on a clinical response.
So PK/PD models characterize how exposure drives efficacy and toxicity. For example, when it comes to antibiotics, they would describe the relationship between antibiotic concentrations and bacterial killing, or also the [00:08:00] relationship between concentrations and adverse events.
Thomas: We talk about pharmacometrics a lot. So can you explain that in more simple words?
Iris: Pharmacometrics is basically an umbrella term for quantitative methods, including modeling and simulation. So we can use these, pharmacometric methods to describe, to quantify, and to predict these PK/PD processes or also other physiological or pathophysiological processes and the interactions with a drug.
And there's also other models apart from PK/PD models, for example, those predicting disease progression
when we talk about quantitative methods and what are the main tools that we should think of or picture here?
There's a wide variety of different quantitative tools. So to give an overview maybe let's focus on pharmacokinetics.
Iris: So the analysis of concentration time data. The simplest case is the analysis of a single concentration time [00:09:00] profile from one patient. So you have a peak after the end of infusion and then an exponential decline afterwards. If many concentration measurements are available over time, we do not necessarily need a mathematical model, but we can perform a so-called non-compartmental analysis.
with this method, we can directly derive pharmacokinetic measures from the observed concentration time profile. For example, we look at the elimination phase, and we can derive an elimination rate constant from the terminal slope of the profile. We can also determine the area under the concentration time curve, and based on that, we can calculate additional parameters such as clearance or volume of distribution or terminal elimination half-life.
So this method is very simple. It's relatively fast to perform, but it also has limitations because it requires rich sampling, so several concentrations [00:10:00] over time, and it also assumes linear pharmacokinetics. or it assumes linear pharmacokinetics which might not be adequate for drugs with complex pharmacokinetics.
Linear pharmacokinetics means that the PK parameters are constant over time. So for example, clearance is constant over time, and that means that an increase in the dose is directly proportional to an increase in ex-exposure. So if I double the dose, I would have double the exposure. ,
Rekha: So Iris, can you tell us a b- little bit about compartmental analysis?
Yeah. So, we talked about the analysis of one concentration time profile. So we could also use compartmental analysis for this. and in contrast to non-compartmental analysis, these compartmental methods describe the concentration time profile using mathematical model.
Iris: and these models represent the body in a really highly simplified way as a system of [00:11:00] connected buckets or compartments that are kinetically homogeneous, and the drug can distribute between these buckets or between these compartments. And maybe it's important to say that, these compartments usually don't correspond to specific anatomical structures or organs, but they are rather functional spaces within the body that behave similarly from a kinetic perspective.
And that's different from so-called PBPK models, , physiologically based pharmacokinetic models where compartments are designed to represent organs or tissues. so as central compartment of a model typically comprises blood and also highly perfused tissues and so-called deep compartments or peripheral compartments, they represent less well perfused, spaces, such as adipose tissue or bone, for example
Thomas: Yeah. as a clinician [00:12:00] who's not making these models on an everyday basis, it's like you said, it's, sometimes hard to describe what you mean by compartment, and we have these equations of drugs or even bacteria moving from different stages or between different compartments.
I don't know if you can answer this, but, how much would you say in, in modeling is related directly to physiology or organs? Like, you discussed this a little bit. are they different disciplines do you take that into account, or is it purely a mathematical model of what's happening with the drug,
Iris: So the extent of physiology in the model depends on the type of models and also on the data we have at hand. So Many models are empirical, so they're completely dependent on the input data. And, if we have what we call mechanistic data for example, information about a certain transporter or an elimination process, for example, the renal elimination but maybe also elimination via [00:13:00] an enzyme, then we can incorporate these mechanistic insights into the model and we would then call the model semi-mechanistic .
Thomas: Yeah, that's a new word. Semi-mechanistic. So that's, when you add, things you actually know is happening, to the model. It's not only purely a mathematical equation.
Iris: it also depends on the types of measurements we have. For example, we can measure concentrations not only in plasma, but also in tissue. For example, in adipose tissue or in more lean tissue, muscle tissue. And if we have these concentrations, we could include these data in dedicated compartments.
For example, a compartment for fat tissue and a compartment for muscle tissue, and then we can more physiologically describe the distribution into these spaces.
Rekha: So you're kind of describing the journey of the drug through these tissues using mathematical concepts. Is that sort of the takeaway for a medical person? Very [00:14:00] simplistic.
Iris: Yes. The journey and especially how long the drug stays in these compartments and how fast it is eliminated from the body.
That depends on the drug itself, but also on the characteristics of the patient
so we discussed, now how to describe the pharmacokinetics in an individual. you said before, to make a really good model, you'd need rich sampling, you need repeated samples. but we often talk about population pop PK models. how do you develop them and how are they useful?
So let's picture a scenario. So we administer the same dose of a drug to many patients, and then it's, really highly unlikely that we observe the same concentration time profile. But instead, we will see an entire band of diverse profiles. So some patients will exhibit high concentrations, or they might be at risk of toxicity, and other patients might have low concentrations and maybe be at risk of therapy [00:15:00] failure.
And there's different methods how to analyze these diverse profiles. one very simple possibility is to simply pool all concentration measurements together and pretend that they stem from one individual. This is called naive pooling, and this method ignores the variability between the patients, but it may still be useful when you have, for example, only one measurement per patient A second option to do population PK analysis is to analyze each individual of the population separately.
Iris: So we determine PK parameters for every patient one by one. And afterwards in the second step, we summarize these individual parameters across the population. So we calculate means or variants or other descriptive statistics, and that's why the method is called two-stage. So first, one-by-one analysis, and then the statistical summary.
This approach [00:16:00] is fairly simple, but it also has limitations. like non-compartmental analysis, it requires rich, and ideally balanced sampling for each individual, which is often difficult to achieve in challenging populations, for example, in neonates.
And the third approach and the main focus of today, I think, is, population pharmacokinetic modeling. And these models, they allow us to analyze all concentration time data of all individuals simultaneously. We can describe the central tendency of the data, so what we call the typical concentration time profile or the typical patient.
So if we picture again the patient population and the band of concentration time profiles, the profile of the typical patient would essentially be the line that is running through the middle of this band of concentration time profile So when you look at the PopPK [00:17:00] publication, it's the pharmacokinetic parameters of this typical profile, so of this typical patient that we estimate and that are usually reported.
So these parameters are referred to as the typical values of parameters, typical clearance, typical volume of distribution. And maybe
Iris: To put it in an easier way, these parameters, they determine the shape of the concentration time profile. So in other words, they determine how quickly concentrations rise and how quickly they decline over time And we can not only describe this central tendency in the data, the typical patient, but we can at the same time characterize the individual profiles of each patient, and we can quantify the variability between patients.
And importantly, one of the core tasks of population PK/PD modeling is to answer the question: Why are these [00:18:00] profiles different? Why do these patients differ so much regarding their drug exposure? So that might be because of different renal function or because of a different body composition, patients with obesity, for example. And with these population PK models, we can now conduct what we call a covariate analysis and find such influential covariates. And these can be important for dosing decisions
Thanks, Iris.
Thomas: it's a big chunk of information.
that I think is useful, also for people who work mainly with patients in clinical practice. So I'm gonna turn to Phil now, to talk exactly about that. If you recommend a dose or dosing regimen, what would this recommendation be based on? Can you use these population PK models to make an actual dose recommendation,
Phil: Yes. So I think dosing recommendation really comes from linking three core things, particularly when we're looking at modeling. So the first thing is probably the [00:19:00] population model, the second is patient's individual characteristics, and the third is the exposure target that we're really aiming for.
So the model really describes, as Iris said, how drug concentrations behave on average, and then how they vary with those individual patient factors. And it's what we call those covariates like weight or renal function, probably the two most common ones that we see. So for an individual patient, we're trying to use those characteristics that they have to adjust away from the average and therefore predict what the likely pharmacokinetics And then if we look at the target, you know, are we aiming for an area under the curve, a trough concentration, or a pharmacokinetic/pharmacodynamic index, like something such as with penicillins, with time above MIC. And that target can really vary depending on the clinical context and the balance between efficacy and toxicity as well.
So ideally, these targets, we inform them by exposure response and exposure [00:20:00] toxicity relationships. So we're not just predicting concentrations, we're actually also predicting important outcomes for our patients. I just think a good example of how those targets might change that I come across regularly is vancomycin.
you have quite a number of circumstances where you might use vancomycin, and most people are targeting an, area under the curve now with this. And you might have a patient who is floridly septic, is MRSA colonized, likely to have MRSA bacteremia, and you're probably gonna want to err more on the side of targeting, an exposure some sort of guarantee that you're gonna get a good amount of efficacy in treating that infection because it's, a life-threatening infection.
And you might accept a little bit more toxicity with this. Whereas if you have a patient who maybe, you know, has a coagulase-negative staph infection, they've had the source removed, say it was a line And they're just really completing the tail end of a course and are otherwise stable, you might be happier aiming for a lower exposure there where you really want to avoid causing any toxicity and noting that you probably don't need as high a target to ensure that efficacy [00:21:00] because, the cost of getting that wrong is, probably far less, in practice, essentially, we're really combining the model, the patient's data, and then the clinical target, and then we're selecting the regimen most likely to receive the desired exposure based on all those things.
And I often think of it as a bit of a diagram where all these things kind of feed into that dose. So, the model's one thing, the patient characteristics are one thing. all those kind of feed into that recommendation that you come up with.
Thomas: Mm-hmm. Yeah. so maybe a question for both of you.
in terms of the population PK models themselves, like how accurately can they predict, the drug exposure in an individual patient? What are the limitations? are there any patient groups that we often mistreat because we apply a population PK model that's not appropriate for them?
Yeah. Well, I think that there's, again, some limitations we have to be aware of with PK models, and I think these come down to the data that's used to build the model. so what is that data? What, population is that from? is quite important. The assumptions within the model, and then how well it translates to our real-world patients [00:22:00] overall.
So, the underlying data we often have can be quite, limited, and that can be both in size and diversity. often we might develop models from a very specific patient population, that, you might not be able to extrapolate to other patient populations, across settings.
and, , example of this is You wouldn't use a model that you built in a critically ill patient population with, a non-critically ill patient population so much, because we know that often the pharmacokinetics in those two populations can be quite different.
I guess the other thing that we need to think about is the sampling strategy that the model was based on. So models built on very sparse or trough-only data,, that's gonna struggle to characterize really accurately that full pharmacokinetic profile. So, if we're trying to estimate something like an AUC, that can sometimes be problematic if we've got trough-only data that the model's built on.
whereas, you know, our richer prospective data generally give you, you more reliable models to, that we can base those recommendations on. all models also do rely on those structural and statistical assumptions about things like the compartmental [00:23:00] behavior and, these are somewhat simplifications of reality.
Phil: They're useful, but they're not obviously perfect representations of our biology. even a well-developed model, we might not be able to generalize that to every single patient. and we often see patients who sit outside populations, used to build the model, and often, as has been said, that might be a patient with a very high body mass index.
It might be someone on dialysis. , One I commonly see , is patients with very low muscle mass, where creatinine-based estimates become quite misleading and inaccurate. And so if you've got a model that's using creatinine-based estimates to give you a dose, that, can often fall down there as well.
So it's not overall the models are wrong, it's just that they are approximations and that we need to apply them with our clinical judgment that we're using.
Iris: I completely agree with you. it's very important that the scenario that we would like to simulate is really supported by the model. So a model that was developed for patients with normal renal function, for example, it might not be appropriate for patients with [00:24:00] severe renal impairment.
Or maybe to give another example, if you have a drug with nonlinear pharmacokinetics and you would like to simulate, steady state concentration time profiles and make a dosing recommendation, and if the model was developed only based on measurements on the first day, for example, the nonlinearity might not have been captured, so you would probably make severe mistakes, basing your dosing recommendations on this model
Rekha: I think it's important for us all to consider which model we are using and why, and understand the concepts and assumptions that go into that model. but let's follow on to some more concepts now.
So Iris, we hear a lot about simulations and PK/PD simulations, I've heard and read about them being used in clinical breakpoint setting and dosing. Can you just explain to us in simple terms what is it that we are simulating and, what are the parameters that are used, in these models when we are simulating?
Yes. so I think [00:25:00] simulations are the part of modeling where the real fun parts are. So you have developed a population PK model, and now you can address what we say are what if questions. Yeah, so you can investigate scenarios that have not been covered by the raw data or by the study. For example, if the population model was developed for twice-daily dosing, you could simulate what happens if I give a loading dose or just a once daily dose or what if, I administer continuous infusion, for example.
We can simulate for single specific patients. So for example, for what we said before is the typical individual. for a patient with a different clearance or a slower elimination or a faster elimination. So what we do is we fix the parameters, clearance and volume of distribution, for example, to the values of the final model that we develop.
Iris: And then we [00:26:00] decide on a dosing regimen we're interested in, maybe a sampling strategy, depending on the software. And then we can simulate, the concentration time profiles based on this scenario. We can also simulate with variability, so we can consider that patients are different and that we call stochastic simulations or Monte Carlo simulation. Stochastic simulations are widely used in the field of infectious diseases, for example, for probability of target attainment analysis or also for the determination of MIC breakpoints
Thomas: Thank you, Iris. so this is something we use a lot more and more, I would say, and it's, determining the clinical breakpoints, as you say. It indirectly determines the dose regimens we use for patients in order to be able to use the clinical breakpoints.
but when you make the simulations, you express the outputs in different ways. Probability of target attainment is, one way. Can you explain a little bit more how you generate these outcomes and what they actually [00:27:00] mean?
Iris: . So probability of target attainment analysis are used to reflect the proportion of a simulated patient population that achieve a specific PK/PD target, as Phil said before. So how does it work? We use the model. We simulate virtual patients we are interested in, for example, 1,000 patients.
And for each patient, we check did this patient achieve the target at a certain MIC value. So for example, if 900 out of 1,000 patients achieve a target, for example, at an MIC of two, then the probability of target attainment would be 90% for that MIC value. So we calculate the PTA for one MIC at a time.
And you might have seen such figures. So P- PTA plots, you see the probability of target attainment on the Y-axis, and you see different MIC [00:28:00] values on the X-axis. And you see that the higher the MIC is, so the less susceptible the pathogen is, the lower the PTA is.
and how high probability of target attainment do you think we should aim for?
Yeah, is a matter of debate, but, in the community, 90% PTA are often considered acceptable. Some also go for 95% or even 99%, but I would say the most common threshold is 90%
Okay, thank you.
Thomas: so that means you sort of lock the PK profile, you lock the target like 50% time above MIC or something, and then you, end up, judging when or you can use that dosing regimen you're simulating for a specific pathogen.
Iris: There is a second concept related to PTA, so that's CFR. So it's the cumulative fraction of response. this CFR does not answer the question, do we [00:29:00] reach the target at one MIC value? But it answers the question whether a dosing regimen achieves a target in the overall bacterial population.
so, CFR considers distributions of MICs. you have MIC values that are more common, and you have MIC values that are less common for a pathogen. And the CFR considers the probability of target attainment across this MIC distribution so it's basically a target attainment analysis, not for a single MIC value, but across an entire pathogen population
Thomas: so in other words, as Iris just described, we can, use PTA and CFR, but none of them actually prove that there is a clinical benefit of one dosing regimen or the other. it's more about, the likelihood of reaching a PK/PD target.
In your opinion, where do these PK/PD methods and outputs have their biggest real-world impact right now? how do we use them? [00:30:00] Do you think they're used in the right way?
it's a point because often, , we, aim for these particular probability of target attainment, but we really would like these to be linked to patient outcomes and, often that's not always the case that we have really strong evidence in that setting.
, I think that both applications in terms of clinical breakpoint setting or dose optimization in special patient populations are important.
Phil: I think that they operate at slightly different levels, though. So in terms of PTA and CFR, we know they've had a major impact in this clinical breakpoint setting, and as Iris said, they'll be really providing that quantitative link between dosing, drug exposure, and then the susceptibility of the microbe that we're interested in.
So, they help define whether an organism should be considered susceptible at standard dosing. I.e., can we actually achieve with our drug the exposure that we need to kill the bug without, you know, is it actually possible to do that in the patient? Can we do it without causing, toxicity that's unacceptable?
I'm a bit biased [00:31:00] because I'm a pharmacist who deals with special populations, so I think in terms of real-world day-to-day clinical impact, that the biggest value right now is in dose optimization, and particularly in special populations and, patients like those who are on continuous renal replacement therapy or undergoing ECMO.
we know in these situations that pharmacokinetics can be highly unpredictable and somewhat variable. So I guess in those settings, our PTA-based approaches let us evaluate whether standard dosing is likely to be adequate, and if not, it allows us to then explore some alternative regimens that might allow us to, obtain our targets.
So I guess in a way, like the breakpoint setting is kind of the foundation as it tells us what exposures we should be aiming for, while the dose optimization is the application, and so it's where we use those PK/PD principles to actually improve dosing in our individual patients or our high-risk groups where we potentially have altered, physiology, altered pharmacokinetics.
I guess if we take those two examples I sort of mentioned before with chronic renal replacement therapy and, ECMO. So [00:32:00] in continuous renal replacement therapy, clearance can be highly variable, and so often our standard dosing might not reliably achieve our PK/PD targets.
Phil: And so, , if we actually run a PTA analysis on people on CRRT, it allows us to test different regimens like extended infusions, and we can select one that approves target attainment in that, patient population. Similarly, in ECMO, we know that sequestration of drugs and altered volume of distribution can affect exposure, and so these approaches help us understand whether the standard dosing is still appropriate, whether it there's not a big effect of ECMO on these things, or whether we actually need to look at changing our dosing approach.
Thanks, Will. in order to do all this, you need, of course, a population PK model then derived from the same type of patients. Sometimes you want to go even further , and take drug concentration measurement and apply model-informed precision dosing, which will be the topic of next episode.
Thomas: is that something you, you would normally use? Or are you happy with the population PK data?
Phil: Where that's available, [00:33:00] absolutely, and I think that's probably where we're moving towards in a lot of these areas. Often the average model that we come up with, while it might give us a decent estimate of that initial dose, there's often gonna need to be some tweaking.
And I think that's where our getting drug concentrations from that particular patient, combining them with the model to give us our model-informed precision dose often potentially has a real benefit and enables us to, probably optimize the dose a bit more effectively
Rekha: And we need really effective models for that though, for many different antimicrobials
Phil: Absolutely. You need the base model to be good, don't you? And you need it to be applicable to that patient population. And, and that's often the hard thing because, , you have these patient populations that behave differently, but it's very hard to recruit those to studies and do studies on those groups because your numbers are often not high.
So that does become practically sometimes quite difficult. Yeah.
Thomas: I personally feel some skepticism sometimes from clinicians who don't really understand or trust the PK/PD data and really want hard clinical evidence from [00:34:00] randomized trials ideally. Do you think we will get this?
Do you think, could be used more in large trials
Phil: i'd like to see it used more. I think pharmacokinetic modeling is becoming quite important in drug development, . we know that. and we are seeing some, of these things come through with drugs.
I know, one drug, was it ceftaroline, or one of them had actually a dose recommendation for CRRT, um, come through, which, you know, is something that we, haven't seen in the past. So I think that we're seeing a bit more of it, but we've still got a fair way to go, and it is tricky.
Phil: it's probably not, the most profit, developing kind of approach for companies developing drugs, producing something that needs a lot of therapeutic drug monitoring and things as well. So I, think that, it's very variable around the world where things are happening, and I know in Europe in particular, we, see a lot of centers doing this very well, whereas, in other places probably we don't see it done as well, as we would like.
but it is something particularly I think that, we'd [00:35:00] like to see develop, and, and I hope it will over the next few years.
Rekha: Yeah. But you know, the most common antimicrobials we use are still, those ones that were developed years and years ago. So we rely on post-marketing, our own-- developing our models in real world settings now.
Thomas mentioned randomized control trials. I think POPPK models can also be used to plan better randomized control trials. For example, we can do PDA analysis and other simulations to identify the most promising scenarios, the most promising, dosing regimens, and we can investigate many, many scenarios, far more than we could ever do clinical trials.
Iris: So it would help us to select which scenario I would ultimately investigate in a randomized control trial. And I could also, of course, use the POPPK model to optimize the number of patients or the sampling strategy in that trial once the regimen has been determined.
Thomas: Thanks, Iris. that's a very good point. You can use PK/PD methods and tools [00:36:00] to plan ahead, before you do your clinical trial, and make sure you have it optimally designed. let's talk a little bit more about the limitations of simulations and modeling.
so we said before we were making a number of assumptions, who will affect the output of the simulation. What do you think are the biggest and most critical uncertainties in the models we're using today? in other words, what could be the most common reasons model-based recommendations can mislead us and don't really make sense in a clinical setting?
one big assumption, of course, is that the patient population for which we would like to apply the model or for which we would like to answer a question, that this population behaves like the population for which the model was built So I think we discussed before, it's really important that the patient characteristics are approximately the same and that, the, timeframe for which the original model was built matches your simulated scenarios, And it's [00:37:00] always important that the model is fit for purpose. So if we would like to draw conclusions from the model, it's really important that the model is good enough to support these conclusions. So we have to check that the model had been qualified for the, intended task.
Iris: So for example, if you want to simulate a population with diverse patients or with variability, it's important that the predictive performance of the model had been checked before. So, there are different model evaluation criteria, for example, you can look for in the paper. and for example, if you want to use the model for different patient population, then what we call external evaluation is important so that the scientists had proven that the model also works for patient data outside, the population that was used for model development.
There's quite a few assumptions that these models do. So they simplify, of course, [00:38:00] biology. they assume certain distributions, another strong assumption we often make is that we model drug concentrations, right? We, don't model dynamic disease physiology very often. So maybe the exposure alone does not predict the outcome. Maybe there's other factors and there is some developments that try to include, for example, immune system factors, , or, some parts of that to take that into account.
And as Phil said before, assumptions about the targets are really important, and they can entirely change a dosing recommendation. So if you use the wrong target, even if you have the right model, then you might also make a mistake with your recommendations. And of course, if we use PTA analysis and these PK/PD targets, they contain MIC values and as you all know, these MIC values, often they are unknown, and if they are known, they have some uncertainty, [00:39:00] of course.
We are not exactly sure whether it's really an MIC of 2. So there is quite a few assumptions we have to make
Phil: Yeah. So I think when you're talking from a clinician to clinician perspective, a lot of it, comes down to how transparent and relatable the assumptions from the model are. so a lot of the uncertainties like the patient population, the models built in, the sampling strategy, or how well covariates like renal function were characterized often, that's not always highly visible when you initially read the paper.
It's often hidden behind a lot of math sometimes and things for clinicians. and I guess if a clinician can't see how a recommendation was generated, it can be very hard for them to then trust it, and especially when it conflicts with maybe their clinical experience of what they've seen happen. And I think that's one of the, biggest barriers to trust in, some of these simulations generated from models.
Clinicians are very used to anchoring their decisions in things they can directly observe, like a drug level or a patient response or clinical response to a drug. So, , model-based [00:40:00] recommendations can sometimes feel a bit sort of imaginative, a bit step removed from those sorts of things, which can make it, sometimes hard for clinicians to then link, th-those recommendations from a model simulation to the clinical situation.
improving trust is really about making the models more transparent, showing how the recommendation was derived, and ideally linking it back to the patient-specific data so that it feels less like a theoretical prediction and more of an informed clinical tool that the clinician's using.
Yeah, that's interesting. when it gets too complex and there's too much math there, it really becomes difficult for someone who is so used to, a pattern recognition or knows their dosing and that dosing has worked in the past to then change their dosing to something that some software or a model spits out that they haven't looked at the, , workings of.
Rekha: I totally get that. how do we make this more understandable to clinicians? for example, if I was on a round and I had to have, looked at a sort of a model and how would I then make it more understandable to me? What are the key points that I would need to look at to say, "Okay, I understand this [00:41:00] model.
Yes, it might apply to my patient. I might give it a go because I really want to use the optimized dosing here."
Phil: so the first thing is you wanna really look at what is the population that that model was developed in, all right? And does it link with your population of the patient that you're looking at using the dosing regimen in. I think if the study population is obviously very different from your patient, and I used previously the example of a non-critically ill patient versus ICU, or normal renal function versus dialysis, then that immediately affects how applicable that particular dosing regimen might be to your patient.
second is the data and the model quality. So, you know, how was the model built and validated? was it based on rich sampling or sparse data? And was there an appropriate external validation as Iris has talked about? And I guess that gives you a sense of, , how robust those predictions might be, how much they can be relied upon.
third is performance in the clinical sense. So the other thing you wanna look for is, how well does it actually predict the concentrations or target attainment. And so we have a bunch of tools used in modeling, [00:42:00] things like visual predictive checks, that, help us answer whether the model is reliable enough to support dosing decisions and, is it gonna adequately predict what we want it to predict.
So I don't think clinicians need to understand every mathematical detail, but they do need to understand whether the model applies to their patient, whether it's been properly validated, and whether it actually performs well enough to be used in clinical practice.
Yeah, I cannot agree more. it's about the quality of the data.
Iris: Is our work data collected in clinical routine? How sure am I that, the data were adequately documented, things like that. Second, quality of the model, as Phil pointed out. There is a lot of models published in PubMed, but some of them might have flaws, or might have not been developed, according to the accepted standards, for example.
the third challenge is probably even if the model is of good quality, how to correctly interpret the model.
Sometimes [00:43:00] physicians would like to compare maybe a patient or population of their own study with a population from literature, and it's very important to understand what the reported parameters mean in such population PK models. So for example, in your own study, you found a clearance of 10, or for example, your patient has a clearance of 10 liters per hour.
And in the publication you see, okay, it's the same population but, this typical clearance is maybe 15 or maybe even higher. So you would assume the elimination is different. But it's very important to have a close look for which patient this value is true, for which patient is this reported clearance 15.
It might be that the reported values are based on a And if you take this into account, you will see that these two clearance values may actually be comparable if you [00:44:00] base the values on the same renal function.
So that sounds complicated, I know, but we try to explain these concepts, in an easier way. And we did a session at, ESCMID Global last year to provide what we call a toolkit for non-experts and where we really explain how to read a pop PK paper, how to interpret and how to compare and discuss these results c- correctly.
So if you're interested, I would encourage you to take a look at this session in the ESCMID media library.
Rekha: Great. Thanks. , Phil, do you have anything else to add to Iris' comments?
Phil: I think we often see, as Iris said, there's a lot of models being produced, and you find a lot in the literature.
I think where we're probably lacking a little bit is actually moving those models towards being really clinically applicable. how we do that is tricky, but, you know, we need to create tools that are easy to use for clinicians, that have clinician buy-in, to be able to use these models and apply them to the clinical context.
And I think that's really important. Sometimes we just stop at [00:45:00] building the model and I think we need to, go a few steps further than that, down the track.
Thomas: I think that's a very good point and currently working with the ESCMID guidelines for antibiotic dose adjustments in patients with renal impairment or hemodialysis, and I, totally understand what you mean.
There's a lot of PK studies out there, a lot of good papers with models and simulations, but exactly how to translate that to something that clinicians can use on an everyday basis is, very hard. It's gonna be expert opinion largely, based on the available data we have. , But I think that's a very good point.
All right. thanks Phil and Iris are there any other last comments from either of you that are important for us to understand?
Phil: the one thing I'd, say is just that, clinical context is just so important here. These are very helpful tools, but they always need to be used in a clinical context, and I think that that's really vital.
And, often our experience clinically is, very important still in forming how we treat patients. and just because we have these, fancy tools, we shouldn't do away with that.
Rekha: Yeah. They wouldn't replace what we do. They would [00:46:00] just be a really great adjunct to our clinical decision-making, and that's really what we are looking forward to, right?
Using a really good clinical decision-making tool to improve how we manage our patients.
Thank you. So with that, it's time to wrap up for today. thank you to our guests, Iris Minichmayr from the Medical University of Vienna, and Phil Selby at Adelaide University in South Australia.
Thomas: Thanks also to my co-host, Rekha, who will return later in this series. I hope you enjoyed this second episode of our PKPD series. In the next episode, we'll move even closer to the patients and discuss existing software for individualized model-informed precision dosing and practical challenges in implementing these tools.
Thank you for listening to Communicable, the CMI Comms Podcast. This episode was hosted by Thomas T�ngd�n, editor at CMI Comms, and Rekha Pai Mangalore, editorial fellow at CMI Comms, ESCMID's open access journal. It was edited by Dr. Katie Hostettler-Oi. Theme music was composed and conducted by Joseph [00:47:00] McDade.
Any literature we have discussed today can be found in the show notes. You can subscribe to Communicable wherever you get your podcasts.
thank you for listening and helping CMI Comms and ESCMID move the conversation in ID, and clinical microbiology, further along.