Bayesian statistical models could help address recruitment challenges, but experts agree that sponsors must first understand – and be prepared – for the additional work required to implement them effectively.
Bayesian models use prior data to update the probability of a hypothesis as new evidence emerges. This distinguishes them from standard frequentist models commonly used in clinical trials, which typically evaluate each study independently without formally incorporating prior information.
Go deeper with GlobalData
Discover B2B Marketing That Performs
Combine business intelligence and editorial excellence to reach engaged professionals across 36 leading media platforms.
The US Food and Drug Administration (FDA) has issued draft guidance on the use of Bayesian models to clarify best practices for incorporating them into drug development programmes.
These models are not completely new to the FDA, with the agency releasing guidance for medical devices in 2010, and some mention of the method in rare disease guidance. This draft guidance is, however, the first time it has been specified by the agency for use in all drugs and biologics.
Drug development lends itself to Bayesian statistical models, says Andrew Garrett, EVP Global Scientific Operations at ICON, due to its linear structure, from preclinical trials, through Phases I to III. This means that evidence accumulates over time that can be used to support Bayesian models.
Other regulators are also moving to incorporate Bayesian approaches. On 30 January, the European Medicines Agency (EMA) also released guidance titled “Use of Bayesian methods in clinical development,” while regulators in China and Japan are already permitting their use.
US Tariffs are shifting - will you react or anticipate?
Don’t let policy changes catch you off guard. Stay proactive with real-time data and expert analysis.
By GlobalData
The FDA’s recent move, according to Dennis Akkaya, chief commercial officer of myTomorrows, a global patient recruitment company, signals the agency is becoming more flexible in its approach to drug development, mimicking regulatory agencies elsewhere.
“The FDA is signalling that it wants to be, on certain occasions, more flexible and accelerate drug development. It is kind of mimicking some other more flexible regulatory strategies that we’ve seen other nations do, for example, in Japan and China, where real-world data is used to a much greater extent,” Akkaya adds.
Borrowed data a major benefit
The guidance will allow sponsors to use “borrowed data”, for example, incorporating data from Phase II trials in Phase III studies, or historical data, an area that should be of particular interest to sponsors, says Tim Clark, Bayesian statistical expert and VP Clinical Science at ICON.
“Where Bayesian comes into its own is where you can use historical information, what we call an informative prior. When you use an informative prior, and you combine that with your data, it can result in sample size efficiencies and greater precision in the treatment difference you are trying to estimate,” Clark explains.
This is particularly beneficial for rare disease trials where recruitment is a persistent challenge. By enabling more effective use of prior safety and efficacy data, as well as historical control data, sponsors can design studies that are more likely to meet recruitment targets while maintaining robust evidence standards.
While it is particularly beneficial in rare diseases, Clark believes that this guidance suggests the FDA is hoping that sponsors will utilise these methods in more prevalent diseases.
It also shows that the FDA is continuing to be more accepting of the use of real-world data (RWD) in studies, says Akkaya – something that has been growing considerably in recent years with the use of alternative trial designs such as single arm studies instead of the gold-standard randomised controlled trials (RCTs).
“There is a shift towards relying more on RWD to find a solution within rare diseases, looking at what we can learn from what sits outside of a trial. This guidance, on top of others, like the expanded access guidance, RWD guidance, guidance around inclusion and exclusion criteria, the plausible mechanism pathway, all of this is signalling that we can learn from additional data sources,” says Akkaya.
Borrowed data also has benefits when sponsors are looking to broaden patient populations, for example, when seeking to gain approval of a drug approved in an adult population into a paediatric population.
“If a sponsor knows how the drug works and that it is safe in the adult population, then they want to extend treatment to the paediatric population,” says Garrett.
“The question is, do you have to do the same sort of studies in a paediatric population? It might be quite hard to find enough children to take part, particularly given caregivers’ considerations, and there is a desire to limit the number of children put into clinical trials. Bayesian methods allow sponsors to augment the prospectively collected paediatric data with the current information on adults, and this can result in the need for fewer paediatric patients.”
The approach can also be used when companies seek to launch in different countries, allowing data integration from previous studies and its use alongside a smaller population from a particular region.
Timeline increases
Using these kinds of methods does have its drawbacks, however, as they increase trial preparation time due to the additional work required to collect the data and ensure it is of the required standard to meet the agency’s requirements, says Garrett.
“There’s a lot of preparatory work that needs to be undertaken when you use Bayesian methodologies. You have to pre-specify your methodology and justify your prior beliefs, which often means running simulations,” Garrett explains.
It doesn’t have to be used on every endpoint, however, Garrett added, noting that it could just be utilised in just the primary endpoint if a sponsor wanted to. This would reduce the time slightly, as these considerations and preparations would only be required for one endpoint instead of the whole trial design.
“If you then want to use Bayesian methods for all endpoints, including the safety endpoints and for every analysis, it becomes a lot of work,” Garrett adds.
Clark agrees that it would increase the preparation work for a study but believes that the benefits outweigh the additional work required and may shorten the study overall. “Companies must be aware that there is more work involved in doing a Bayesian study compared to a frequentist study. Therefore, they must be prepared for that, but when you incorporate external information into the analysis, it can result in sample size reductions – and with a smaller sample size come shorter timelines,” Clark says.
As well as this, Akkaya believes it will allow sponsors to investigate more ultra-rare disorders, where they may have previously not been able to overcome recruitment challenges.
“It might be easier to pursue smaller, more complex kinds of patients that you’re looking for in a trial. That is something that has been a big challenge, but this guidance might make it easier to do this because you’re allowed to borrow information from patients that might not initially meet the criteria to get into your trial,” Akkaya says.
Draft guidance remains in its infancy
While this represents a positive signal for the industry overall, significant hurdles remain. Sponsors must develop a clear understanding of how these methods work, what types of data regulators will accept, and which study designs are best suited to their application. As companies begin to adopt them, implementation approaches are likely to vary, Clark believes.
“We’re at the beginning of the journey with Bayesian statistics in pharmaceuticals now that the authorities have released guidance,” Clark says.
“This shows a greater interest in the agency part for sponsors to use these methods. I am sure that new ways in which they can be effectively used, resulting in less of an upfront burden, will be developed, as well as new metrics for deciding whether certain models are appropriate or not. As a result, the way we look at Bayesian analysis now may be quite different in five years’ time.”
