Rishi Bansal (Ontario & Hertford 2023) is a DPhil candidate at the Nuffield Department of Population Health with a focus on clinical trial methodology.
Tell us about your research.
I'm doing my DPhil in Population Health, it's my third year now, and my thesis is about exploring and describing the methods and performance criteria required for Bayesian methods to be applied to large-scale randomized clinical trials. Essentially, we're looking at how these methods are applied to drug and device trials to find out how well those methods perform in the real world. We’re trying to offer guidance to the medical community — how do we make sure that these trials are robust and produce good results that we can trust, and translate into actual clinical practice.
So you’re helping people who are designing clinical trials to design them better?
Yes, that's really what my research is — how do we design these trials to give them the best chance of being successful? To me, success doesn't necessarily mean that the trial is positive; rather, how do we give it the best chance of providing the medical community with a really definitive answer one way or another? That's what we really want to get at through doing trials — can we get a definitive answer that tells us whether to use a particular intervention or not?
What's difficult about setting up a clinical trial at the moment?
Problem number 1 is that they're just so expensive. Trials are really expensive to run globally, and the price tag of a clinical trial to test a new drug has only increased in the past 10-20 years. Since everything that you're doing costs so much, you become really cognizant of all the choices you're making, such as how many patients you enrol, or how many countries you include.
On top of that there are so many unknown quantities that go into the scientific and statistical elements — we don't actually know whether the drug works or not before we start the large Phase III trial that's intended to study its efficacy. We may have some idea, based on previous trials that have been done, or prior knowledge about the drug or similar drugs, but we don't actually know until we've done this large-scale trial.
That, in itself, is a controversial statement these days — there's lots of debate about how we know that something works, and whether we need trials to show that or not. Obviously, I'm biased given that it’s my area of research! But there are so many questions about the statistical quantities, the scientific quantities, the expense, and the logistics can make it really challenging. So this all comes together in a uniquely complicated mess, but that's the fun that we get to sort out.
So is it just about making trials cheaper, or more than that?
I think it's more than just making them cheaper. There are lots of things that can be done to try and reduce the expense — in particular, Bayesian methods have been put forward as one way to make trials cheaper, because you can potentially reduce the size of your trials through all sorts of statistical methods, and you can stop your trials a little bit earlier than you might be able to otherwise.
My work looks into disentangling what's happening when we make those decisions — are we actually benefitting from having fewer patients in this trial, or making it shorter than we would normally plan? In what scenarios is that completely reasonable and the trial still gives us a really strong answer to our research question, and in what scenarios is it inappropriate to stop the trial early or enrol fewer patients, even if that means it's going to be cheaper and easier to run? Might the entire medical community benefit from having a much larger, much longer running trial that's going to definitively answer our question, as opposed to one that's going to give us a more uncertain answer but might be a little bit easier to do?
Can you explain Bayesian methods to a non-scientific audience?
It's actually a really intuitive thing, because it's basically how we think about the world, or think about anything really. Bayes' theorem, which was written by Reverend Thomas Bayes, says that you have a prior probability, or some prior information about an idea, a question, or topic; you have some data that you're going to observe from an experiment, or some new information that's coming in, and this data is going to somehow modify your prior information, or belief, into what's then called a posterior belief or a posterior distribution or probability.
So it's really just as simple as that: we know things about the world just from our experiences living in it or, in a medical setting, because of other drugs, other trials, other information that we might have before. Now we have a particular scientific question at hand and we're going to run an experiment to collect new data. We're then going to take that in and see how that new data modifies what we knew about our hypothesis beforehand. Does it agree with what we knew before? Is it completely discordant from what we knew before? We can then finally put these together, in this case mathematically, to produce a posterior conditional probability that represents what we now know, after having completed our experiment.
Is applying that to clinical trials a new thing?
Bayesian methods applied to trials have been around for 50+ years. There are some fantastically written papers from the '60s, '70s, '80s, where you have these statisticians debating each other in the medical journals about whether to do things using the Bayesian method, or in the classical frequentist tradition.
This debate has been going on for a very long time,but it's really only come to the fore in the past 15 years or so, largely because a lot of the machinery behind being able to run Bayesian models is simulation, and that requires a huge amount of computing power. Prior to 2005-10, the computing power that was available to the average person running these trials was just not enough to be able to support that level of simulation. And the computer packages that were necessary to run these things were not written until the late '90s and early 2000s. And then in 2003, there was the first ever FDA approval of a drug based on a Bayesian analysis. That’s really when the medical community realized that this is something that is here, that's acceptable to regulators, that's scientifically sound, and the number of trials that have been using Bayesian methods since then have really skyrocketed. Especially during the pandemic, there were a lot trials powered by Bayesian stats, and those were all published in all the top journals. So a lot of people became more familiar with the concept and since then, it's really taken off.
But in my eyes, this is a field that's still getting its legs. There's so much to explore, there's so much that we don't know. There's so much that we haven't decided as a community on the best way to do this. The FDA only released regulations and draft guidance on applying Bayesian stats to clinical trials two months ago. This type of guidance can sometimes take a decade to actually reach its final form, so it's a really exciting field to work ins because it's exploding in popularity and relevance. That also provides so much opportunity to look into the unexplored areas and try and leave a little bit of a mark, and guide the field in a way that we think is best.
Has COVID changed the nature of clinical trials?
COVID was a really good example of what to do, what not to do. There were lots of trials that were done at the time that were really small, that did not have very many patients, that didn't run for very long, and produced results that weren't particularly helpful to anybody, because the precision and the certainty around any answer was just not nearly enough to provide what a clinician working in the hospital, who has patients who are dying in front of them, would need to know.
On the flip side, there were a few trials, including the RECOVERY trial which was run by my department here at Oxford, which were the complete opposite of that. They took a very different philosophy — they wanted to test all kinds of drugs that people were interested in, or hypothesised to work for COVID. But they would only do the trial if they could enrol thousands and thousands of patients, and wait until they were certain that they had an answer to whether or not a drug worked, before releasing those results to the public. Therefore, a lot of what we found out about how to treat COVID and what drugs do and don’t work came from RECOVERY — because they took that philosophical approach of saying that we're not going to stop running this trial until we have an answer that we feel that clinicians and governments around the world will be confident in.
So what next, after you finish the DPhil?
I'm a medic by training. I finished med school before I came here to do my PhD, and so I'll be going back to do my residency training in internal medicine for the next 5-10 years to get my clinical training to go along with the fantastic research training that I've got now. And, hopefully, I will have a career as a clinical scientist to look forward to, where you get to, in a very privileged manner, split your time between seeing patients and working in the hospital, and having dedicated time to work on research.