In Episode 36 of "In the Interim…", Dr. Scott Berry and Dr. Don Berry analyze the Phase II trial of Lecanemab (BAN2401) in Alzheimer’s disease, focusing on the application of adaptive Bayesian methods following persistent failures in Alzheimer’s drug development. The conversation covers the specific design features of five active arms, response adaptive randomization, and a longitudinal Bayesian model driving interim decisions, as well as direct operational and statistical challenges encountered during the trial. The hosts address regulatory proceedings, critique from "experts" regarding adaptive methods on noisy cognitive endpoints, and the direct alignment of the trial’s Bayesian 18-month efficacy estimates with the subsequent Phase III results and regulatory approvals. Key Highlights Alzheimer’s drug development context: Widespread Phase III failures prompted a retreat from conventional trial designs and a demand for greater rigor and adaptability. Lecanemab Phase II methodology: Five active arms, two dosing schedules, response adaptive randomization, and adaptive interim analyses at every 50 patients enabled real-time adjustment and efficient dose evaluation. Bayesian modeling and imputation: Use of a longitudinal model to address missing data, forecast 12- and 18-month outcomes, and inform both allocation and stopping criteria. Operational adaptations: The design accommodated unplanned safety restrictions, such as stratified randomization for APOE4-positive participants after ARIA signals. Expert skepticism: Addressed Paul Aisen’s concerns about adapting to noisy interim cognitive data, emphasizing safeguards against erroneous stopping or success. Regulatory outcome: The 18-month efficacy estimates from Bayesian modeling during Phase II matched Phase III findings; FDA granted accelerated approval based on amyloid reduction and later full approval after Phase III confirmation. For more, visit us at https://www.berryconsultants.com/
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Judith: Welcome to Berry's In the Interim podcast, where we explore the cutting edge of innovative clinical trial design for the pharmaceutical and medical industries, and so much more. Let's dive in. Scott Berry: Welcome everybody. Back to in the interim today we're actually gonna take you. In the interim, we're gonna take you into many interims, uh, a unique look back, uh, at a trial, a drug, a disease. And I have, uh, a wonderful guest, Don Berry is here, my father, to join
me, and we're gonna talk a little bit about the story of Band 2 4 0 1. And the, the story of a phase two trial and the before and the after. So welcome back to In the Interim. so let's go back to approximately 2013 and the, sorry, state of Alzheimer's disease drug development. Something like 14 straight failed phase three trials, uh, promise of drugs.
But all of these trials failing and take us back to that time and set the stage a little bit, Don. Don Berry: Okay. Um, it was, as you indicate, uh, a sorry time, but Alzheimer's, uh, researchers to their credit. Understood. That they were doing something wrong. They didn't know what. Um, and the Alzheimer's Association put together a, a round table, uh,
a meeting in Washington, which had the goal to study and learn from the bit of a workshop, uh, the 14. Latest phase three trials in Alzheimer's, all of which failed. And so there was a, each of the trialists were asked to present
what the trial was and to the extent that they could, why it failed. Was it the drug, was it the trial? They had invited me to give a presentation. I mean, they were aware of, uh, IFI two, which, um, which was a trial in, in cancer. But, uh, did, uh, you know, smart things along the way, and so they.
Like, uh, many, uh, uh, researchers at the time sort of salivating over doing something like ice flight tool. So they, I presented, uh, my analyses of what the problem was. Um, one of the, uh, I won't go through all 14 of course, but one I want to mention is a large. A pharma company, uh, that had a drug that they were looking at for
Alzheimer's, and they had a phase two trial, and it led to a large phase three trial that bombed completely. The person presenting was one of the instrumental people in the company who was part of the design of the phase two trial. It was part of the analysis to go forward into phase three, and he
admitted, said explicitly, you know, we didn't know that we had a good drug here. We recognized that, that, uh, we might fail. The data would not all that. Uh, uh, consuming of course, what the. The, the, the, the down part was in the cognition and, uh, function and, you
know, the drug was getting in, it was doing what it was supposed to be doing and the biomarkers that they could measure. Um, but it was not, you know, big enough so to speak, uh, to show cognition and function. And, uh, so it was an, an admission and the, and the person wa was asked, okay, so why did you do the trial? Scott Berry: trial? Don Berry: Why did you do the phase three trial if you didn't have enough evidence
that the drug was going to have an effect? And he essentially said, well, yes, we were playing the lottery. Um, that is, we're taking a chance. And, you know, the chance might be best based on a, a phase, uh, uh, three trial that Scott Berry: that. Don Berry: just randomly, you know, 5% of the time, even though you don't have anything, you have a, a positive trial.
So, um, uh, he was, uh, I, I, I thought surprising, and this happens a lot, uh, that you. You don't know that you, you sort of hope for the best. So we're in World Series mode. So, uh, we have a, a, a pitcher that's a great pitcher, uh, and you're swinging for the fences and you're hoping to both hit the ball and hit it right and hit it into the fences.
So there's a lot of randomness. Uh, let's hope for the best. Um, the result of that was at, at the meeting was Veronica Vinky, who was at the time at Eisai. Um, and she asked if she could discuss with me, uh, their circumstance with, uh, the. The, uh, ban, uh, 24 0 1. Um, and uh, so I agreed and I visited them and, uh, they were enthusiastic
and that's really what it takes is enthusiastic part of the company. They were enthusiastic about doing something different, um, and um, they ended up doing something. Different. And Scott, back to you. Scott Berry: Okay. Okay, so sorry. State of drug development question of whether, whether it's the drugs, whether the trial designs, whether it's the development plan, whether it's actually patients and endpoints.
Do these patients actually have Alzheimer's disease? Are we measuring the right endpoints? Lots of blame going around. So as you mentioned, uh, ASCI and Veronica Lavinsky, Andy Satlin, Chad Swanson, SBA Dadada Jinping, uh uh, also. Uh, was involved in this Jinping Wong, so a large number of people, large effort. We built a phase two trial and the notion was this is, we
want a better phase two trial. We don't want to just jump into a phase three trial. So I'll, I'll describe the trial somewhat briefly, but the trial design had five different arms of band 2, 4 0 1, the drug. It had two different frequencies. The bimonthly frequency had three doses, and the once monthly frequency had two doses, so five active arms and one placebo arm.
In order to incorporate the five arms and not have a much larger trial response. Adaptive randomization was used in the trial, so interim analysis would take place and allocation would be focused on the arms that were doing better. The endpoints in the trial sort of plays an interesting role here. 12 months and 18 months change from baseline in a cognitive measure in this patient population called Adcoms.
Adcoms is a collection of, uh, questions from MMSE, Addis Cog and CDR sum of boxes that was thought to be, uh, highly sensitive in this part of the disease. And so this was the, the endpoint in the trial. In order to drive the adaptations, the RAR, we built a model, a longitudinal model that would look at three months, six months, nine months, 12 months, 15 and 18, and forecasting forward where the patient was to identify within patient that
forecast to identify how well the arms are working and to drive the adaptation. So a key part of this is what we'll call the longitudinal model. Within that at, uh, adaptive analysis, the trial could stop for futility. They wanted to be realistic about this, given the, the, the, sorry state of, of Alzheimer's development. They also wanted triggers for accelerating drug development. So if a particular trigger was hit, they would go quickly to phase three.
If, if it wasn't futile and it wasn't this trigger of phase three, it would go to the next interim analysis and it would get more data and more data during the course of this. So this is largely the trial design within it. Now, the sample size also was, was adaptive 300 was the minimum. And then 800 was the maximum. If we didn't know the answer yet and we kept going, we'd enroll to 800.
We would then continue interim analyses as we're getting exposure through 18 months. The trial design. Now, this is a unique trial in terms of disclosure. Uh, ACI has published. The result of every one of the interim analyses during the course of this trial. And there were 16 of them during the course of this trial. So many interim analyses during the course of this trial, and I think we'll
take you through the story of the trial, what happened in the going forward, but any other comments about the design before we sort of walk through the, the, the trial and what happened, Don? Don Berry: So the, um, the interesting thing is, is, is what happened in the trial? uh, no, I think you just get into that. Scott Berry: Yeah. Yep.
Yeah, let's, let's, I, I know where you're going and let's, let's not quite go there yet. Um, within it. Okay. So all of this is, has been disclosed. Um, interestingly in the course of this is that. Uh, in the trial design, I was unblinded and part of a group that was reviewing the interim analysis. Don was blinded. Don did not see the data. He was involved in the design. We were both involved in the design, but he was. Blinded to any results going on during the course of the trial.
The trial was conducted by a company called Ella that they ran the interim analysis, they, they curated the data, ran the interim analysis, and then I was involved in reviewing them with Christine Broo, who's also here at Berry Consultants. So the first interim analysis in the trial happens 196 patients enrolled. It happens in the date of this, I don't want to get this wrong. The date of this is March 2nd, 2014.
So the trial design, as Don described, happened in the time of this, and at that time, randomization is updated for each of the arms in the trial. Even at this first interim analysis, the randomization. And this is published, um, with this, and we'll, we'll add a link to this is randomization to the two. Higher doses is increased and each of the other three doses is decreased.
The placebos held constant, the trial does not hit futility, uh, it can't hit success at this time or, or flagging to ACI to go to phase three. So the trial continues on. One comment about, uh, flagging to ASCI is they were interested if there was a high posterior probability that their effect was at least a 25% slowing.
That was kind of this important trigger for them. If they were highly likely of a 25% slowing, they wanted to know that, be told that, and they would accelerate to phase three. If there was a low probability of a 25% slowing, they would stop for futility. Okay? That continues on every 50 patients. So at two 50 and 300 similar analyses go on now at, at, um, 300 is where, uh, an
an unplanned event happens in the truck. Don Berry: So can I interrupt you and, and, and say a little bit more about, uh, Scott Berry: Yep. Don Berry: the 300, um, when we talk about these things, a lot of people say, oh, well you need of course, complete data. Um, this is not complete data. This is all the patients in the trial. Nobody is at the, uh, 12 month point at this time. The first analysis. Um, but we're using the longitudinal model to predict what's gonna
happen in the, in, in the future. So, um, and the, the most interesting thing that happened at this first analysis was that the algorithm was already learning that the two lowest doses, Scott Berry: doses, Don Berry: um, Scott Berry: um, Don Berry: were not effective. And so it essentially zeroed out. It didn't, it wasn't quite zero. Uh, the first time the adaptive randomization, um, and it was built
so that they could come back if the longer, uh, term data pointed out that, you know, you're missing something and, and now it's the low doses are actually doing well that never happened. Um, the something to address Scott and something that, um. The journalists addressed and quoted people in the research community was the many interims.
I mean, it was just taken as a negative, you know, you're, you're doing all these things and you're gonna be misled. And, uh, there was lots of criticism about simply the multitude of it, of analysis. Uh, would you comment on that? Scott Berry: Yeah, and, and, and we'll come back to this, uh, in, in part of the story, um, and it's general and, and now there's this pervasive viewpoint that it's too complicated, these number of interim analysis, but
every 50 patients and analysis was done, and a key part of the design is. We're, we're only stopping at that point. If we have strong evidence, a 95% chance that we have at least a 25% slowing or less than a seven and a half percent chance that we have a. 25% slowing. So, and, but if, if we're not there, we're gonna continue on in the trial. So, and then we're gonna update the randomization as Don described in that.
So this trial and its ability to update was critical and the analysis every 50 patients enrolled was critical as we'll. Describe what happened in, uh, as the trial goes. Don Berry: So, as Scott mentioned, um, the, the, fact that these are available, the interim analysis are available, one of the places is, um, a publication that, uh, I was first author and Scott was last author in JAMA Network.
Uh, open. And you, what you should do, uh, dear viewers, is go to that, go to the supplemental material, and you can actually watch a movie. You can, uh, show, uh, a movie of what happened in all of these analyses and how did things change Naturally, there's correlation from one to the next, but you, you're looking at the tendency of, uh, of. Of where it's going, what it's doing, it's assigning to the higher doses.
Uh, and control, along with it control had the same probability as the highest, uh, uh, probability experimental, uh, uh, drug. Um, and it's, it, it's very instructive. It's, it's not, these are obviously not independent. We're taking, uh. Uh, uh, these serial looks, uh, and it's one of the few studies that, uh, all of
the interim analyses are are published in. Back to You, Scott. Scott Berry: Yep. So the first interim analysis, and, and Don described that it, it's triggered by patients enrolled and I, I'll back up. The, the, the, there's two Bayesian models that are critical to this. One is a dose response, so the 12 month change from baseline and Addis Cog is modeled in a two dimensional, normal, dynamic linear model over the five.
Arms and placebo. The other one that turns out to be even more critical is we built a model between each visit and 12 months for the correlation between them. Built a linear model between each one of these, that a patient's change from baseline at 3 6 9, wherever their last exposure was, is used to estimate what is their value going to be at 12 months.
Multiply, imputed through the model to estimate the dose response and drive conclusions. So, as you're flipping through this, as Don describes at 300 patients, it's a bit of a a fun movie to watch at 300 patients. The data's at its worst relative, the arms to placebo. The placebo's actually at that point estimated with high uncertainty to be the best arm.
And the probability that the 10 bimonthly dose is better than placebo, I'm sorry, is as a 25% improvement over placebo is 18% had that have dropped below 5%. The trial would've stopped for futility, and that's the closest it's ever got to futility. It moves on to three 50. So now by the way, we are in the, uh, September 26th, 2014.
So we've moved on about seven months. They've enrolled 250 patients, they're on their fourth interim analysis, and in enrollment is temporarily paused in the trial. There's a safety committee that's, that's looking at the data and they notice a, not a, a, a radiological finding. It's, we now understand this in the diseases aria, and it is a radiological
finding that looks like microbleeds in the brain, and specifically for the highest magnitude dose, which is 10 bimonthly. They are noticing this for patients that are a POE four positive, they are asymptomatic. So these are not symptomatic bleeds in the brain. They're asymptomatic, but they were concerned about what is this? To the extent that they said ASCI could not enroll that high dose, one of the
five arms could not enroll patients A POE four positive on that arm. So we've built this adaptive randomization. For, um, all of the different arms, one of the stratification factors is a POE four positive. At that time, the decision is made by people blinded in the trial that we're just gonna separate that randomization for a POE four positives and a OE four
negatives, and there'll be different randomization, adaptive randomizations, which will zero out the 10 bimonthly. And the trial continues Don Berry: Uh, more over the patients who had been assigned to that. Within the last period of time would be stopped. So it was, you know, it was like a, an a bomb dropping on the trial. And, uh, a I was very concerned.
They knew, of course, about this, they got the message from the regulators and they had to do it. Uh, and they talked to me about, you know, canceling the trial. It was such a huge, uh, negative Scott. Scott Berry: So the, the, the, this is one of those cases where an unplanned adaptation was able to, very quickly, by the way, Ella did a fabulous job that now it's randomizing patients with this, this, uh, separated randomization
for them and the trial continues on. At this point, by the way, at 350 patients in it wants to randomize. It is absolutely zeroed out. The three low doses, the 2.5, the five monthly, and the five bimonthly, it wants the two high doses. So a POE four negatives are getting the 10 bimonthly at a high rate. The positives are getting, uh, the only, the 10 monthly at that point,
and it is already focused at this point in the trial on the high doses. And now the estimated efficacy, the probability of at least a 25% slowing is about 55%. Again, if that's 95, they would've signaled to ACI back in 2014 to run phase three, but it did not signal. It says, you know, the data's good, but keep going. Its estimated efficacy. That point is approximately a 25% slowing.
This continues on, and interestingly, and so now at this point, only about 15 patients per dose have passed 12 months. And it's estimating that that high dose has about a 25% slowing through this longitudinal modeling. The early patients enrolled are starting to move Past 12 months, nobody. Don Berry: Data and the Scott Berry's model for missing data, Scott Berry: Yeah, so on the 10 bimonthly dose, a POE, four positives that had less than six
months exposure were removed entirely. So there were different kinds of missing data. There were patients in the trial who went away. There were patients that were forced to go off drug, and then there were patients that were enrolled so frequently that we just didn't have their 12 month data. All three of these types were being imputed based on their last cognitive status in the multiple imputation model.
It 500 patients enrolled. So this is, um, June 19th, by the way. These interims took two or three days to run. The data was there, the, the algorithms were run that largely took a day. They were reviewed in 12 hours a day. Uh, they were signed, that they were good to go, and they were implemented immediately. The information was passed to a data safety monitoring board, but
these were done essentially fully automated with a review to make sure nothing weird was going on and. The trial continued. At this point, it's estimating about a 25% slowing, which happens to be in this really interesting part for aci. If there's high probability, it's better than 25, they wanna run phase three. If there's low probability, they would wanna stop for futility, and it's hovering around that uh, point.
Don Berry: And that's on the basis of the imputation of the missing data. Had you looked at just the raw data and ignored this, you would've had something like, I don't know, 10% or something, a lot less than 25% and the the DSMB um. they happy with the the model, or did they react to the model because everything's driven by this model. Scott Berry: Yes.
They, they, they had no concerns with the model. Um, in it, we had, we had, we actually, uh, created and we only interacted with them infrequently. The interims took place. We said, statistically they're, they're, they're good. And they signed off on it and continued. It was this one aspect of the a OE four positives, and that continued throughout the trial. This continues on at every one of these interim analyses, and so the
trial gets to its full enrollment. It actually over-enrolled 856 patients, uh, ended up over enrolling in the trial. And at this point in the trial, this, this takes place. Full enrollment takes place in September of 2016. All patients are enrolled at this point. The only thing it can stop for futility. And it never got close to futility again after that third interim, or it could signal to aci.
Now, there was the ACI built in a really cool part of this trial that this all is being built, uh, is being triggered on 12 month change, but all patients are being followed through 18 months. And the trial's gonna run and get full exposure through 18 months. And at 12 months, this was billed publicly as, uh, you know, we'll find out if this is gonna go to phase three. And at this point they have that. If there's an 80%, they lowered the threshold.
If there's an 80% probability of a 25% slowing, they wanted to go to phase three in that. And again, they could stop for futility. They found out at this point it's not 80% and the actual value, let me find out what the actual value of this was at the time, um, that this interim analysis happened. Oh, shoot. Uh, at the time this interim analysis happens when all patients
are enrolled, the probability of at least a, um, uh, 25% slowing was 70. Was 65%, so that needed to be 80 and it was 65%. So it didn't jump that 80. So ACI publicly disclosed this that we're continuing exposure in the trial. We don't know the answer yet.
And interestingly, I'll, I'll go to a snapshot of. A, an article that was written by the Alzheimer's in the Alzheimer's forum about this result. And they interviewed you, Don, and asked you about this. And they also interviewed a neurologist, a well-known Alzheimer's trialist, uh, Paul Aen. So they asked you about this result that at 12 months, we don't know the answer yet, and Donna's blinded.
And he says the ban 2 4 0 1 trial results so far to the extent that we know them, and all he knew was the trial was continuing. It wasn't 80%, and it wasn't less than seven and a half indicate the therapy is neither extremely effective nor clearly ineffective on balance and in context of the, sorry, state of drug development in Alzheimer's disease. That's good news.
Was Don's quote at this, and we are now six months from full readout of the trial. Uh, and all they know is it's not 80 and it's not less than seven and a half. Paul Azen takes an opposite view, and, and I'll read you his quote and I think it's kind of interesting. Um, so what do these results mean for the field? Not everyone is convinced that adaptive trials are the wave of the future. Paul Aisen of the University of South Carolina, uh, California in San Diego
commented that while adaptive designs have proven useful in the oncology field where endpoints are more objective. They may be problematic in ad trials, Alzheimer's disease trials. The great and now quote, the greatest concerned about a concern about Bayesian adaptive approaches in ad trials relates to our cognitive clinical outcome measures. These measures are highly variable and noisy, adapting
on interim incomplete data. Risks chasing noise and making erroneous decisions regarding efficacy or futility. Don Berry: So I'm, I'm just compelled to say I know that it has noise. I know this variability. I recognize it. I modeled it. I live with it. And I can tell and build into the algorithms the uh, and the design,
what's real and what's noise. Scott Berry: And what, what was so misunderstood by Paul is that, what did the adaptive design say? It said, we don't know the answer yet. Let's keep going. We need more data. Uh, we, we, we don't know the answer, and his criticism was, they're, they're too risky of making erroneous decision, claiming success or futility, when in fact
it knew that the variability and it knew keep going, go to 18 months, uh, in it. So think there was this sort of doubt and it was built in and part of it. Okay. Flash forward now six months. And the trial finally reads out 18 months. And the fi by the way, the, the randomization, so the adaptive randomization, what's the role of this adaptive randomization there?
Out of the 856 patients enrolled in the trial, 238 go to placebo. The lowest dose gets 52 patients. The next gets 48. The other five milligram gets 89. Those are the three doses that very quickly are moved away from the two high doses, get 2 46 and 152, and the only reason it was 152 is because it
couldn't enroll a POE four positive. But at that time, out of the, uh, out of the band 2 4 0 1 doses, 587 get put on band 2 4 0 1. Two thirds of them, 67%, 398 are on the two high doses. So the response adaptive randomization, when Don goes back to the Alzheimer's and says, you need better dose finding trials with better exposure.
To understand efficacy, you need to explore dose it. It did that and through response. Adaptive randomization very quickly focuses on the high dose. Don Berry: And so you get greater precision associated with the high doses, exactly. The doses you're thinking about moving forward. So it's, it's a huge benefit. But I wanna point out that this is an attraction for patients. They can be told in the design that if there is a dose that is doing well,
we're gonna keep looking at that. And so you are, uh, if that has happened in the trial, uh, you're more likely to get a dose that's doing well. And not only in this trial, but in other trials that we use adaptive randomization. We find out that that's exactly true, that the, the end result is that the patients in the trial, uh, you know, this has been something that has been vetted around for, uh, decades, uh, in jama, et cetera.
Is are, are patients in clinical trials treated better than patients? Not in clinical trials, and some say yes, some say no, and there's data on both sides here. It's clearly the case. That if you're in an adaptively randomized trial and there is something that's positive, you're more likely to get it. So it should be an attraction to not only physicians, um, uh, but also their patients.
Scott Berry: So the trial reads out at 18 months. And the two high doses that I described with the larger sample size, the 18 month, uh, estimate for the 10 monthly. So this is the, the high dose that's given monthly, and then there's the high dose that's given bimonthly. The 10 monthly is estimated to be an 18% slowing. And its probability of being better than placebos, 93%, but
it isn't estimated to be 25. The high dose given twice monthly is estimated to be a 27% slowing. And again, think about where this is relative to their targeted 25 and a 76% probability that it's at least a 25% slowing is the estimate of that dose. So Asci at this point does a couple things. Uh, it starts a phase three trial and it selects the 10 bimonthly.
At this point, aria is much better understood, and there's actually some sense that this is a level of efficacy of the clearing of amyloid. So in the background, the other thing is the mechanism of this treatment is the removal of the pathological amyloid. It and they're doing pet scanning and they it, the high dose is absolutely removing amyloid within this and this ar Aria's thought perhaps to be in effect of this
in some sense, almost that it's showing benefit by this demonstration of Aria. Uh, it's still a concern. And, and of course, uh, uh, if symptomatic bleeding or a higher level of bleeding would of course be a bad thing within this. So they run this phase three trial, but they also submit the data to the agency, the phase two data to the agency, um, uh, and present to them. So they plan a trial of the 10, bimonthly, enrolling both a POE
four positives and negatives. And, uh, 1,795 patient trial is enrolled. The, um, the meanwhile, the FDA reviews the data, they give accelerated approval to Ban 2 4 0 1, which has now changed its name to lecan. They grant accelerated approval based on the reduction in amyloid, and there's
a stunningly cool data that if you look at the reduction in amyloid for the lower doses relative to the higher doses, and you look at the change in the cognitive values for those doses, it's incredibly predictive. And I think the fact that the multiple doses in the effect on amyloid re relative to the effect on the clinical outcome is very highly predicted by the effect on it.
Almost demonstrating, and I know we shouldn't say the word, a surrogacy of this, I think enabled them to trigger accelerated. Now the phase three trial reads out 1,795 patients reads out, and the press release, uh, for this comes out in 2022 and it demonstrates highly statistically significant benefit on CDR sum of boxes for that dose at 18 months.
And the estimated slowing of the rate of decline is 27%. It was nailed. Uh, normally we, you know, to variability we wouldn't have expected exactly, but it was, it was exactly the estimate from the phase two trial in that Bayesian modeling that's doing the longitudinal modeling, that's doing the dose response, estimated it to be a 27% slowing at 18 months and it hit it right on the head.
In that which ended up to be highly statistically significant, and FDA gave, uh, kinumab full approval for the treatment of Alzheimer's disease. Don Berry: I go back to that meeting we had with the 14 failures in phase two. Uh, it, it's hard not to wonder, uh, whether if they had done something in those 14 trials. Like this, you know, saw something that was positive, uh, and increase
the sample size accordingly. If we wouldn't have better treatments today for Alzheimer's disease. Scott Berry: Yeah, it's, it's hard to go back to this trial, but I, I describe a couple things happened that I think was powerful. If you go back to the algorithms, and you can go back to this and, and look at the posterior estimate of them when there are 500 patients enrolled in this trial, um, and this happens, 500 patients enrolled.
This happens in 2014. The algorithm is estimating this 25% reduction, and it holds pretty constant all the way through the end that it's giving this forecast. Now, uh, I, I think. A side did a great thing by getting this exposure and this decision, but it, it pretty much forecasted at that point. But I also alluded to back at 300 patients when 300 were enrolled, this analysis took place for futility.
The data were not promising. Had they run a 300 patient trial, a small phase two, they might not have ever continued development of this drug. Okay, so the alternative designs that could have been run may have led to this never being run in phase three. It's hard to know, and it's hard to know the decisions that would've been made, uh, in this scenario. Don Berry: And also had they used something other than.
A reasonable and somewhat sophisticated missing data imputation process. Instead of that 27%, uh, that, uh, uh, they observed in the phase two study, uh, and they looked at just completers, it would've been 12%. Part, and this is part possible to explain on the basis of the, the fact that the missing data were precisely, if you look at what the missing data were,
uh, for patients who had been assigned to the high dose, that distribution was very different from the missing data distribution for the, uh, uh, other doses and for example, for control. Um, and so the model recognized that, um, that it was, uh, that whether the data is missing, of course, the, the model didn't know about this, uh, regulatory, uh, folks
that, uh, mandated, uh, dropping that, uh, dose from the future part of the trial. Um, it only knew what the data were and was basing it on what the longitudinal modeling was. So it was, there, there were a number of, uh, it's sort of a perfect storm, or not a storm at all, but perfect weather, uh, that, uh, everything was, uh, moving in the right direction and, um, it gave rise to a positive conclusion.
What's the latest? I mean, the umab has been fully approved. Uh, now another, uh, phase, uh, phase three trial has been, uh, observed with, uh, similar mechanism of action. And can you say something about the current status, Scott? Scott Berry: Yeah, so Umab from Eli Lilly has been approved full approval, uh, and it demonstrates slightly better, uh, slowing of cognition, I think at, at 18 months.
But now I think there are two treatments that are disease modifying that have been approved. And interestingly now where Alzheimer's was a bit of a a, a dead end. For pharmaceutical development. There's a huge investment now in part that, that clearly that that, that affecting amyloid has an effect and it has an effect on slowing disease. Now, the, going into tau going into better effects of this there, there's much
more in the, in the way of development. I, I, I wanna point out something interesting. Don Berry: We always wanted to change the world. Scott Berry: the, yeah. Don Berry: I. Scott Berry: Yeah. And so the interesting thing is that trial design was run meanwhile, A, there were more than ACI that were interested in Don's messaging of better phase two trials and Abbott. Was one of them. And Rob Lens was the clinical, um, in, in charge of neurology at Abbott at the time.
And they ran two trials, one of them a BT 1 26, which is published. The results of that stopped for futility. And I think largely that the conclusion was that the drug didn't work. So a very similar design run in that scenario with a different drug sort of was pretty strong that the drug doesn't work. And I, I think that's the right conclusion. So the design doesn't make the drug. But hopefully the design gets the right answer for the drug, and I think it's
two examples of putting in a different profile to a very similar design. Gave very different answers at the end of the day. Don Berry: There was an interesting discussion with Rob and the, uh, clinical trial committee at, uh, uh, at at Abbott. They asked him at, when, at the beginning. So this is a complicated trial. I mean, it had many things, many uh, aspects, very, uh, much in common with, uh.
Uh, the 24 0 1 trial. Um, how much time did it take to do this? And Rob said, you're right. It took a while. It took us like six months working with the Berry's to get what we wanted. Um, but on the other hand, the trial ended two years sooner than it would have otherwise because we were doing the right thing in investigating. And, and, uh, understanding that we really did hit futility with that drug.
So it's good in both directions. Scott Berry: Yeah. Yeah. I, I guess a last thing it would be good to come back to is you made this comment, so in the trial there were 17 analyses of this data. The data's accruing, it changed the randomization, but this notion that. Uh, what's beautiful in the Bayesian perspective and is that the analysis at the end of the trial is unaffected by the fact that all these analyses take place
within it, and it makes so much sense. And, and there's this view that somehow these 17 analyses Infect the data that it no longer tells us with the same level of strength and in in viewing this and seeing the results go through and look at this and just. We should be doing this in most of our phase II trials. It's good for patients, it's good for drug development. It is not a bad thing that these analyses take place when
a snapshot of the data are done. It doesn't change the data, it doesn't change the scientific conclusion, and I think this is a beautiful example of exactly that. Don Berry: And in a way it's, uh, it's ai. You know, everybody's talking about ai, but this is a prospectively designed trial. As an algorithms, it's run it, the algorithm is like an automaton. Um, and so we can evaluate as we did, uh, the type one error associated with it,
the power of, uh, the study, um, and the, uh, again to sell at Tom Park, um, working on the logistics associated with it. Uh, you know, we, uh, uh, hearkening back Tom, first work, Tom and I were first work together in a Pfizer trial where we had daily stroke trial, where we had daily updates and the algorithm would re randomize, uh, would reassess the
randomization probabilities every day. Uh, and this kind of thing. There is a logistical aspect of this. Um, but it's possible to do, and it's possible to have something that's rigorous. And, uh, and the FDA has issued guidance to, um, uh, uh, address the question of the, uh, complex, the adaptive complex adaptive designs. Uh, so the FDA is wholly involved here.
Not only the. Neurological part of the cancer part, but other parts as well. I. Scott Berry: Yep. All right. So, uh, again, uh, jama. Network open, uh, Barry at all. Uh, Lecan map for patients with early Alzheimer's disease. Bayesian analysis of a phase two B dose finding randomized clinical trial has a really unique look at, uh, uh, a trial that, uh, with a very
cool story and a very cool impact. And so here on, in the interim. We've taken you through 17 interims and a look back at a, at a pretty cool trial. So Don, thanks for joining me. In the interim, I know you could not join me in the interim in that trial. Uh, but thanks for joining me here Till next time.
In this episode of "In the Interim…", Dr. Kert Viele, Director of Research at Berry Consultants, takes up a requested topic: published results showing that Bayesian borrowing of historical control data cannot increase...
In this mailbag episode of "In the Interim…", Dr. Scott Berry answers two listener questions and weighs in on the pace of innovation in clinical trial science. Asked what infrastructure a learning healthcare system...
In this episode of "In the Interim…", Dr. Scott Berry and Dr. Kert Viele, Director of Research at Berry Consultants, review the European Medicines Agency's draft reflection paper on methodological aspects in...
In this episode of "In the Interim…", Scott Berry is joined by Jenny Devenport and Kaspar Rufibach of Merck Group, co-authors of the EFSPI Statistical Methodology Leaders paper "Statistical Methodology Groups in the...
In this episode of "In the Interim…", Dr. Scott Berry shares the framework he presented at a CSRC regulatory science hackathon on digital twins and synthetic controls in cardiovascular clinical trials. He separates...