FDA is encouraging medical device firms to consider using Bayesian estimation in clinical trials to streamline regulatory timeframes.

The statistical method has the potential to shorten study duration, cut costs by reducing the number of patients, and enhance product development, CDRH Biostatistics Division Director Gregory Campbell, PhD, explained at an FDA/Johns Hopkins University-sponsored symposium on Bayesian analysis May 21.

For several years, FDA has alerted individual sponsors to the Bayesian tool during clinical trial discussions; however, the agency is refocusing on the method as it executes a "critical path" initiative to curb regulatory burden (¹ Also see "McClellan’s FDA Legacy? Decisive Role In Product Development Priorities" - Medtech Insight, 22 March, 2004., p. 8).

Bayesian analysis "might be least burdensome," Campbell remarked, "but it is really up to the companies to trade what might be a lower clinical burden for what might be a higher statistical or computational one."

FDA now has four Bayesian statisticians on staff and hopes to recruit one or two more by year-end, Campbell said. He projected that device regulatory submissions that include Bayesian modeling will increase from the current level of approximately 5%.

From 1999-2003, CDRH reviewed at least 14 original PMAs and PMA supplements in which Bayesian estimation was the primary analysis. They included applications for stents, spinal cage systems and a heart valve, the agency notes (see chart: " ² Examples Of Device Trials Using Bayesian Methods For FDA Approval ").

"We've also had many IDEs approved which have their main proposed analysis as a Bayesian one," Campbell said. "And there are lots of things in process."

FDA has been investigating premarket uses of Bayesian estimation since 1996, when Campbell - along with former CDRH Director Bruce Burlington and former Office of Surveillance & Biometrics Director Larry Kessler - first proposed new statistical tools to ease regulatory pathways.

CDRH's Office of Device Evaluation, division directors and scientific reviewers lent their support. Subsequently, FDA hired several Bayesian consultants in 1997 and hosted a related conference in 1998, co-sponsored by AdvaMed (³ Also see "Bayesian Statistics Approach Offers Benefits To Sponsors, FDA Staffers Say" - Medtech Insight, 16 November, 1998., p. 32).

The device center learned that Bayesian methods can achieve the same regulatory decision sooner, or with less data, than current clinical study formats. The tool gauges the probability of a predicted event occurring in future trials, based on the number of times the event has not occurred in previous studies.

By using quantitative historical data from the same device, or from similar devices, Bayesian estimation establishes a weighted, predictive model of a device trial's outcome. Historical data are referenced as "priors." The prediction hinges on a built-in confidence interval based on the strength of the priors used.

Medtronic Sofamor Danek's study for InFuse bone graft/ LT-Cage , for example, used "Bayesian methods...to combine results from patients seen at 12 and 24 months with results from patients seen only at 12 months to increase the precision for comparisons for success rates at 24 months," VP-Clinical Affairs Bailey Lipscomb, PhD, recalled.

"The basic reasoning for doing this is that 12-month responses should carry some information about responses at 24 months, provided that there is a correlation between the responses at the two time points," the statistician explained. The trial sought to prove non-inferiority using an approved device as a control.

"What you're really doing is changing what constitutes success in a clinical study," Campbell asserted. "The reliance now is in terms of the posterior distribution. So careful thought needs to go into choosing exactly what is the decision rule in terms of posterior probability, what is the probability of superiority, and what is the credible interval's lower bound."

"Devices often have a lot of prior information," he observed. "If you contrast the situation versus pharmaceutical products, what we have is a mechanism of action, which we usually understand, which is physical, rather than pharmacokinetic, chemical or pharmacodynamic, and it's usually local in its action and not systemic."

"If you have a new molecular entity that is discovered, it doesn't change over time. In contrast, if you have a new device that has been invented, the inventor, the engineer, the clinician is tinkering with that device, trying to improve it on a monthly basis," which results in a series of prior data, Campbell noted.

Recognizing this factor, Medtronic was one of the early converts to Bayesian estimation, after U.S. Surgical (now Tyco) and Sulzer Spine-Tech (now Zimmer) gained approval for spinal fusion cages in 1996 - three years ahead Sofamor Danek.

Searching for ways to recover lost ground, Medtronic attempted to reduce the standard clinical follow-up time for spine fusion cages from two years to one year. "We went back to FDA a year later with these ideas of doing a Bayesian type of analysis and [Campbell] and his people greeted us with open arms," Lipscomb said.

"We knocked several months off our ultimate approval [in 1999 for InterFix ] by using the Bayesian predictive methods; therefore, I became a real believer in Bayesian...a sort of proselyte, hallelujah!" Lipscomb proclaimed. "We've converted studies...to Bayesian, [and] in the new studies we're submitting, we're 100% on the bandwagon for Bayesian."

While touting benefits of Bayesian estimation, speakers at the Bethesda, Md. symposium cautioned that the tool is not a panacea. "Success stories are important, but don't approach the Bayesian inference as if this is the thing that is just going to take care of all of statistical inference and all problems," Constantine Gastonis, PhD, Brown University, warned.

Critical to the Bayesian approach is obtaining solid priors, speakers agreed. "If you only have a single prior study, the Bayesian model does not work very well," Campbell said.

"Without good prior information it is not clear that it is worth the risk for many companies, [which] are betting usually millions about whether...the Bayesian approach [and] the model turns out to be predictive."

Device areas targeted for increased Bayesian activity include drug-eluting stents, which face a growing volume of clinical data collected on each particular stent and mounting pressure to come to market quickly.

Johnson & Johnson/Cordis already has built a hierarchical Bayesian model using historical control data for a Cypher sub-study of vessels less than 3 mm, according to FDA.

"Our effort is not just restricted to Bayesian statistics," Campbell offered. "We're interested in pioneering the notion of using new methodology for diagnostic products and thinking about causal inference and how you could do that in a regulatory setting."

Bayesian estimation is hardly a new methodology. The tool derives from a probability theory developed by English theologian and mathematician Thomas Bayes (1702-1761).

The basic formula was found in his notes posthumously and published as an "Essay Towards Solving a Problem in the Doctrine of Chances" in the Philosophical Transactions of the Royal Society of London in 1764.

The method was then adopted by Pierre-Simon Laplace (1749-1827) and applied to astrophysical mechanics and medical statistics. The tool was not widely accepted, however, until the mid-20th century.

"What happened in Bayesian statistics from 1970 to 1990 is no less than a sea change...because of computing hardware and software," Campbell explained. "Statisticians in 1970 were theoreticians who were imagining a conjugate of priors.... You can now do the calculations on any prior."

CDRH Promotes Bayesian Tool To Speed Approvals; Statistical Rigor Required

More from Archive

More from Medtech Insight