Nathan Green 0:00
Hi my name is Dr. Nathan Green from the department of Statistical Science here at UCL. And this is the Sample Space podcast from the department. I'm very pleased to be joined today by Dr. Hakim Dehbi who's also at UCL in the Comprehensive Clinical Trials Unit. And we're going to be talking about a presentation that he gave to the department few months ago about controlled backfill in oncology for dose finding trials. So before we get stuck into that Hakim, can you introduce yourself a bit more, please?
Hakim Dehbi 0:49
Thanks for having me, Nathan. And thanks for inviting me to this podcast. So yeah, sure. Who am I? I'm head of statistics currently at the comprehensive clinical trials unit at UCL. I am a medical statistician by background but prior to that life, I was an engineer. I studied at UCL. My journey started at UCL at the Stats Department 10 years ago. And after that MSC I went to Imperial where I did a master's of research in cardiovascular science, and PhD in biostatistics. And then I came back to UCL and I've been at UCL since then, as medical statistician, firstly, the cancer trials unit, and now at the comprehensive clinical trials units. And my job is really is varied, I am involved in teaching a lot. At the same time, I was twice that decision. I don't do much trial anymore. In terms of analysis, I'm supervising my colleagues. And I'm working mostly on ground developments, which is working out an appropriate trial design for a research idea. And working out the sample size and developing the applications with the with the medical team. And that's about it. And a lot of admin, obviously, like all of us.
Nathan Green 2:04
Yeah, the more important you are, the more admin you have. So you're probably very important. That's right. And my objective somehow is to become less important. So I would use my admin. Yeah, yeah. Learn from me. Okay, that sounds great. So I'm going to talk about a particular project you've been doing. I mentioned, it's called controlled backfill in oncology. Some people might not know what any of those words mean. Could you give a little bit of a background for what the problem is with this whites problem that people are interested in?
Hakim Dehbi 2:36
Yeah, sure. Absolutely. Well, thanks for the question. The motivation for this paper that was published a couple of months ago now is that in those finding trials, most of the methodology was developed in the era of cytotoxic agents. So cytotoxic agents effectively are chemotherapy agents, and what they do is that they destroy the cancer cells. By doing so they create extra toxicities for the patients. But there was an easy surrogacy between toxicity and efficacy. If you observe side effects on the patients, you can assume reasonably that the drug is doing its effects, it is actually destroying the cancer cells. But nowadays, things have changed a lot in the field of oncology. In particular, we've got immunotherapy, we've got small molecules and other mechanisms that are such that the surrogacy between toxicity and efficacy is not maintained to the same extent anymore, it might well be that as smaller a lower dose level is as efficacious as a higher dose level and this has been seen in practice in some immunotherapy trials. So given this changing background in those finding trials in oncology, depending on whether we are with cytotoxic or non cytotoxic agents, we need to relax some of the working assumptions of these trials, we have to think about not only the maximum tolerated dose, but also the recommended phase two dose which takes into account efficacy, which also needs to look at the other dimensions of the problem including pharmacokinetic pharmacodynamic. And and therefore, somehow we need to collect information on the lower dose levels in a given dose finding study. The objective is not to go as fast as possible through the levels and more, as was the case with cytotoxic agents. It sounds intuitive, but what are the advantages of lower doses then, if you give a lower dose level Firstly, it's less costly. Secondly, there is probably less toxicity cytotoxic agents have side effects. Non cytotoxic agents also have side effects they may not be as immediate as cytotoxic agents. They may be delayed in the medium or in the long term but they also have side effects by giving a lower dose, you reduce the risk of side effects. Therefore, you reduce the risk of non compliance or treatment interruptions or treatment actually stopped stopping, you reduce the costs of the drug. And you might actually not use much in terms of efficacy or obtain the same efficacy as higher dose levels.
Nathan Green 5:20
Maybe I'll to give it to I don't know the word for these things never to like give it to some more patients. Maybe some patients, you could only have higher efficacy or higher toxicity doses?
Hakim Dehbi 5:32
Absolutely. Yeah. If we recommend a lower dose level, it might be tolerable for a greater proportion of patients.
Nathan Green 5:39
Okay, great. How I'm thinking about it is you're kind of you're not making the assumption that just more is better. Yes. And so you want to somehow be able to define the shape of this this dose response curve, rather than just saying it's just seemed like linear monotonically increasing thing? Up until the point? It's just too toxic?
Hakim Dehbi 6:01
That's correct. Yeah. So what you've described is indeed the previous paradigm, but it is changing, right? Okay. So how new is this? This concept of backfill patients has appeared in the literature and in the field of applied clinical trials over the last three to five years, I would say, it was not done in a formal way. We just noticed in protocols that backfill patients were mentioned. And it wasn't clear how many backfill patients would be used. In addition to the dose finding patients, it wasn't clear on which those level they would be allocated to. And it wasn't clear what was the objective of backfill patients. Intuitively, we all understood the reasons for backfilling. But there was no statistical framework around it. And that is the reason why we thought we could actually write a paper to provide this statistical structure and to also take into account the patient's perspective, if you are a patient if you've got a certain type of cancer, and if you join a clinical trial, you want to be treated ethically, you don't want to be treated on those level one. If at that point in time in the study, it has already been shown that those level one is not efficacious or not as efficacious as other those levels, it's great to identify tipping point, if you want where the dose efficacy curve become becomes flat. However, if it becomes flat, if there is a tipping point, but and obviously for this to be to be to be achieved, you need to collect information both on the left and on the right of that tipping point, but not to the expense of the patient's right. Therefore, we need to somehow derive a statistical structure that would allow us to identify the point in time where this is not a reasonable idea anymore.
Nathan Green 7:53
Okay, that's really good. I suppose I've been guilty of just seeing patients and patient data as just numbers in a, you know, in a calculation, but actually to consider their experience as a patient and what's best for them individually. So you touched on it. So what is your this statistical approach that you've come up with in order to do this?
Hakim Dehbi 8:14
So, during the course of the study, we monitor various parameters in addition to toxicity. Toxicity is often in those assessed by the presence or the occurrence of those limiting toxicities. So these are side effects of the drugs that are considered too pronounced for further increase in the dose level. So we've got these DLTs and we you know, monitor their occurrence in the various patients at the various dose levels, but in addition to that, we collect information on the various other markers, including pharmacokinetic, pharmacodynamic, and potentially response rates. So in oncology trials, they are various parameters, including what we call objective response. And this is made of complete response, which is proper disappearance of the cancer lesions or partial response. And both of them together can be seen as a binary endpoint. Yes, no, at the level of the patient has the drugs worked effectively in simple terms, and and we monitor that during the course of the study. And then we use a statistical technique we use Changepoint models to to assess the presence or absence of a plateau at the end of the study. However, during the course of the study, we use Bayesian statistics, and the simple beta binomial model for statisticians in the in the audience of the podcast, to work out the probability that the efficacy at those level one is less than the efficacy or the other those levels combined. Once we've got a posterior probability that the efficacy is reduced compared to the other dose levels, and we set a threshold that let's say 80%, then we discard those level one from the set of potential doses for backfilling. And this is an iterative process. Once level one is discarded, the first available those level for backfilling becomes those level two in the study. And we repeat this during the course of the study. And at the very end of the study, we use the change point model. To identify the presence or absence of a plateau, we fit two models and with model selection with it one model with a change point a model without a change points. And if there is a change point, if model fit is better with a change point, then without a change point, then we conclude that there is a plateau potentially, instead of recommending the maximum tolerated dose, which is a dose associated with a certain percentage of side effects, we would recommend a dose level that is lower the dose level at the start of the plateau effectively efficacy wise.
Nathan Green 10:48
So what does the change point mean, the change point models, I've seen a lot of these paths on changepoint models where we have maybe a latent event where the rate has changed at some point where you don't directly observe that rate, you have observed the the number of events. And so the change point would be the point that the model with the best fitting model would say that the frequency of events has increased, you don't mean that by a change point?
Hakim Dehbi 11:16
You need to deal with the plateau. Right? That's yeah, absolutely needs and not in this context. In this context, the change points model would be a logistic regression model that includes the logistic transform up to the change points, and then a straight line thereafter. And imagine that you plot on the y axis the probability of response and on the y axis on the x axis, you've got those levels, you would have a logistic transform on this on this on this graph up to the change point, and then it would be flat. Okay, I see. So it's kind of to do the gradient racism thresholds, as you know, sufficiently? That's right, horizontal. Okay. Yeah, that's really interesting. Regardless of the medical fields, you may have a proportion of survivors who, for whom the drug has worked. And you would notice then that the survival curve is is going to plateau, those who've actually survived for that long are the real survivors. And there is no reason for the survival curve to continue up to zero. And it has been shown it has been shown with CAR-T cell in oncology, for example, and they were reporting the news recently, that some patients have been declared cured. Thanks to CAR-T cells and CAR-T cells have been a big a big revolution in oncology trials for specific types of cancer and UCL was was involved very much in the development of CAR-T cells. So this is where actually, you know, other challenges. We face all the challenges in in the analysis of clinical trials, where the models for survival analysis do not apply anymore, right. And the proportional hazard assumption may not be respected in such cases. This is not the point of this current paper. But it is fascinating that actually, these are the developments in in medicine that drive the changes in our discipline in statistics, or medical statistics in this case. Yeah, well, I mean, and that's a good thing. Right? So I have the impact from the the research, like the real life impact impact on modern medicine.
Nathan Green 13:22
In the interest of time, I've got a couple more questions. First of all, I'm interested in how you would allocate individuals to different doses in the trial, is it do you just randomly distribute patients?
Hakim Dehbi 13:36
It's an excellent question. No, we do not randomly allocate patient to those levels. There is a principle of precaution in those final trials, they are often first in human trials. So we have to go slowly and progressively throughout the dose level starting at the first dose level, and we should avoid those skipping. So there are various algorithms that have been or model based designs that have been developed over the years to allocate the patience to the dose levels in a progressive manner. We start at those level one and we observe what happens in terms of toxicity. If the toxicity profile is acceptable, we would proceed to level two, we would then proceed to level three, level four etc. Once we've explored the range of doses that are acceptable toxicity wise, then we try to identify the optimum level according to a certain criteria which might be those level with a certain percentage of associate with a certain percentage of toxicities. This is for what we call the dose finding patients, but we also have the backfill patients and these backfill patients are not necessarily located where we think the MTD is amongst the dose range, we would allocate them to lower dose levels. And these patients would at least what this is what we suggest we would randomise them to the lower dose levels as long as this is feasible ethically, as long as we do not have enough information to say that actually these those levels are not as efficacious as the dose levels currently given to the dose finding patients. So we've got these two groups of people, we've got the dose finding patients on the one hand, and we've got the backfill patients. On the other hand, the backfill patients would be allocated to lower dose level than where the escalation experiment is at. But with this extra ethical constraint, if you want make sense.
Nathan Green 15:30
This is my last question. And it's an obvious one. So what next for this piece of work? Where does this methodology go? Are there still some outstanding problems which you can apply this or maybe some variations of this type of modelling for different contexts?
Hakim Dehbi 15:51
Thanks, Nathan for the question actually, backfill belongs under the umbrella of amplification, we are effectively amplifying the information that is collected from a regular dose finding study at levels that are different from the maximum tolerated dose or the recommended phase to those there are other ways of amplifying those expansion cohorts is another example. One can also use pairs of those levels in the escalation phase to spread the experimentation. In terms of amplification, there is still a lot of work to do to characterise it, and to derive the most appropriate ways of amplifying. Because in early phase dose finding trials, the patient's perspective, the ethics are particularly important. We are trying to identify the recommended phase to those for the science, it's super important to obtain as much information as we can on the various dose levels, we cannot randomise. But we need somehow to derive during the course of the study various metrics to guide us, and this is what we are currently working on. I have the chance to work with John Doe quickly, which is the main author of the continual reassessment method, which is an approach that has revolutionised the way those findings. Trials are performed. It's much more appropriate, much better from a statistical perspective than the three plus three for example. And we are currently working on this with another colleague in New York, Alexa Sonos, so this is what we are busy with. Okay, because I'm still planning to do. Yes, absolutely. And I do that in my spare time, in addition to all your work, and admin.
Nathan Green 17:47
Great. I think we're gonna leave it at that. So just to say thank you very much for taking the time to talk to us. It's been very interesting. And let's go and enjoy the nice spring weather.
Unknown Speaker 17:59
UCL minds brings together the knowledge, insights and ideas of our community through a wide range of events and activities that are open to everyone.