Latest / Cardio Pod: Heart of the Matter / Importance of Subgroups in Heart Failure Clinical Trials - Part 2

Transcript

Importance of Subgroups in Heart Failure Clinical Trials - Part 2 · Cardio Pod: Heart of the Matter

• 0:00
  Good morning, good afternoon, good evening.
• 0:08
  I'm Chris O'Connor, the principal investigator of the Heart Filler Collaboratory, and we're
• 0:13
  here with Heart of the Matter, which is the HFC podcast.
• 0:18
  And I'm with my very distinguished friends and colleagues to talk about an exciting session
• 0:22
  that we had last week on subgroups and subpopulations, special populations, and really how do we
• 0:29
  handle these in clinical trials today?
• 0:33
  And I want to just introduce my distinguished colleagues before we jump into the heart of
• 0:39
  the matter of subgroups.
• 0:40
  Bill Abraham, who is a professor at Ohio State University, Joanne Lindenfeld, professor at
• 0:46
  Vanderbilt University, Scott Solomon, professor at Harvard Medical School, Brian Claggett,
• 0:53
  senior statistician at Harvard, and Mitch Sopka, chief of Heart Filler at the NOVA Health
• 0:59
  System, and Mona Fusat, who is a member of the Duke University faculty and also a member
• 1:06
  of the Food and Drug Administration.
• 1:09
  Welcome all to Heart of the Matter.
• 1:11
  We had this session on Friday where we talked about subgroups and special populations, and
• 1:17
  really we talked about, first of all, terminology, which is so important in clinical trial conduct.
• 1:23
  And when we got a group together and said, okay, these terms are different and they matter.
• 1:31
  And when we talk about subgroups, we're talking about characteristics in a clinical trial
• 1:36
  population that can be measured and that can be organized in a fashion for a particular
• 1:43
  analysis.
• 1:44
  And so when we talk about subgroups, we're talking about characteristics such as age,
• 1:48
  ejection fraction, maybe biomarkers, natriuretic peptides, they could be a genetic profile,
• 1:55
  they could be comorbidities such as the presence or absence of diabetes, but they are different
• 2:00
  than when we talk about special populations such as cardiac amyloidosis, hypertrophic
• 2:05
  cardiomyopathy, and genetic cardiomyopathies in which there are actually a pre-specified
• 2:12
  hypothesis on why that population of patients may respond differently or uniquely in a clinical
• 2:21
  trial.
• 2:22
  And so I think as we start this conversation, we want to make sure that we understand the
• 2:28
  difference between subgroups and special populations.
• 2:33
  And I'll turn to Scott Solomon to make some of the first comments on, you know, we've
• 2:38
  done a large number of trials, many of us in this room have been involved in many trials,
• 2:45
  and we conduct a trial, we try it, we set up a pre-specified subgroups, usually 10 or
• 2:51
  so that we're going to do an analysis on.
• 2:54
  And then we come to the end of the trial and we look at these and some are interesting
• 2:59
  and some are not, but how do we get more rigor in thinking about subgroups and how do we
• 3:08
  prioritize ones that may be of particular interest to the audience?
• 3:14
  Yeah, Chris, that's a great question.
• 3:17
  And when we tackle, I think on Friday to a large extent, when we do a clinical trial,
• 3:26
  we usually are asking one particular question.
• 3:29
  That's our primary hypothesis.
• 3:32
  We're determining whether we can reject, essentially reject the null hypothesis for one specific
• 3:38
  question overall in that study.
• 3:42
  And then we, as you said, divide our population based on a whole lot of different groups,
• 3:51
  men and women, diabetes, no diabetes, older, younger, et cetera.
• 3:57
  We may have 10 of these or 15 of these or 20 of these in a trial.
• 4:01
  And then we usually report those data in our primary manuscript.
• 4:07
  We do that primarily to look for consistency.
• 4:12
  But occasionally we see evidence of heterogeneity.
• 4:16
  Occasionally we look at these data and we say, wow, the people with blue hair behave
• 4:21
  differently than the people with green hair.
• 4:25
  And we have to ask ourselves, is that a real finding?
• 4:30
  Is it statistically robust?
• 4:33
  What does it mean?
• 4:35
  What does it mean for the therapy that we're testing?
• 4:37
  And the implications are different, of course, if the trial is very positive or maybe marginal.
• 4:47
  It may be different based on the extent of the benefit that we see in a particular subgroup.
• 4:54
  And then the real question is, when do we really believe it?
• 4:57
  And can we believe it enough to act on it?
• 5:01
  Or do we have to go and do a whole nother trial?
• 5:06
  And so we presented data, we presented some concepts, some data from trials that we all
• 5:13
  know well on Friday that included three important heart failure trials, the Paradigm trial,
• 5:21
  the Paragon trial, both with Sucubitriol valsartan, and then the Galactic HF trial.
• 5:27
  And I guess we also presented some data from the Deliver trial as well.
• 5:33
  And I'm going to turn over in a second to Brian Claggett to describe these a little
• 5:39
  bit better because he presented these data and actually served as the primary statistician
• 5:44
  for most of these studies.
• 5:47
  But we have examples of a trial where all the subgroups look pretty much the same.
• 5:53
  And we have a trial where there are examples where the subgroups look different, where
• 6:00
  we wonder if in fact there's some group that works better than others.
• 6:06
  And these can be very different.
• 6:08
  And the interpretation can be very different.
• 6:10
  Brian?
• 6:11
  Yeah.
• 6:12
  I think our...
• 6:13
  I mean, our starting example is we showed Paragon and Paradigm.
• 6:16
  It was the same drug being investigated in both, just one in reduced EF heart failure,
• 6:22
  the other one in preserved EF.
• 6:24
  And we kind of started from the normal forest plot figures where you see all two dozen or
• 6:30
  three dozen or however many subgroups that you've got.
• 6:34
  And I guess our standard practice has been to do an interaction test for every subgroup
• 6:41
  variable and to...
• 6:43
  Maybe we get to include that on the forest plot.
• 6:45
  Maybe not, but you somehow report either there were this many number of significant interaction
• 6:52
  tests and then somebody afterwards has to decide how much they care about those p-values
• 6:58
  because invariably you'll find one or two that are a little bit below 0.05.
• 7:03
  And then it kind of becomes this clinical and statistical argument where if you found
• 7:09
  something that you didn't expect and we argue about, well, is it a false positive?
• 7:13
  Or if we didn't find something that we kind of did expect and we say, oh, well, we were
• 7:17
  underpowered.
• 7:18
  And if you did find something you did expect, then you can say, yeah, I knew that was exactly
• 7:23
  what was going to happen.
• 7:24
  But it's kind of a game that we play afterwards, but we don't really know exactly what we should
• 7:32
  be taking away from these forest plots.
• 7:35
  I think that's one issue that we need to get better about is how much we can expect to
• 7:40
  learn from these things, knowing that you are likely going to find something that looks
• 7:44
  significant and when you should react to that or not react to it.
• 7:48
  And so in paradigm, I think we broadly interpreted those results as very consistent benefit across
• 7:55
  all the subgroups.
• 7:56
  In paragon, we interpreted there as being some plausible, reasonable sense of heterogeneity,
• 8:04
  largely based on ejection fraction.
• 8:09
  And Brian, I think you nicely shared with the audience during that or our think tank
• 8:16
  the rigor you have to think about when you're looking at whether a signal is real or noise
• 8:23
  when you're looking at subgroups, which I think is just a mess.
• 8:27
  I always remember the ISIS one trial and Peter Sleitz saying that he showed that the thrombolytic
• 8:34
  therapy in an anchor aspirin, I think it was aspirin.
• 8:37
  It was a factorial design, had its highest effect in Sagittarius and its lowest effect
• 8:43
  in Aquarius because they tested for astrological signs.
• 8:47
  But they had done, like you said, hundreds of subgroups.
• 8:50
  And so what is the rigor that a clinical trial team must put around those interaction terms
• 9:01
  of those subgroups?
• 9:03
  Like you said, if you have O5, we get excited about an O5.
• 9:07
  I get excited about a 0.1 for an interaction term because it's so underpowered.
• 9:12
  But really when you've done 20 comparisons, Professor Bonferroni is going to turn over
• 9:20
  in his grave, right?
• 9:22
  Right.
• 9:23
  You're almost guaranteed to find something significant at that point, right?
• 9:26
  Yeah.
• 9:27
  No, I think usually in clinical trials, we kind of defer to type one error and say we
• 9:32
  want to go out of our way to avoid false positive findings.
• 9:36
  And so if something is underpowered, we say, well, that's too bad.
• 9:39
  You should have designed a bigger trial or design another trial.
• 9:43
  And if it's real, you'll find it again in the next study.
• 9:48
  And so I think type one error usually is our primary concern.
• 9:53
  And so one of the proposals, knowing that there's always this multiple testing issue
• 9:58
  when you go down your list of subgroups is there is this global interaction test that
• 10:04
  has been used a couple of times, in particular in Paragon.
• 10:09
  And we said, well, instead of going through and testing every single subgroup, why don't
• 10:13
  we do a single overall test for heterogeneity, for effect modification?
• 10:21
  And if you can pass that test at the 0.05 level, then you're at least allowed to continue
• 10:26
  on talking about people who respond better or worse than other people.
• 10:30
  But if you can't pass that first test, then we should just stop and not torture the data
• 10:35
  any further.
• 10:36
  So I think that was maybe the first thing that we can think about implementing and using
• 10:42
  as a way to gauge how much we should trust any of these findings.
• 10:47
  What about the fact that many of these subgroups that we look at are correlated with other
• 10:53
  subgroups so that when we look at them individually, we're only seeing one dimension.
• 10:59
  But often when we look at them together, we'll see that there may be, for example, more women
• 11:08
  in the high ejection fraction category or vice versa that might be actually what's driving
• 11:13
  the subgroup.
• 11:15
  And your approach that we took in Paragon, for example, where you used a multivariable
• 11:21
  model to include all the interaction terms and the covariates does account for that,
• 11:29
  correct?
• 11:30
  Right.
• 11:31
  I think there's two possible scenarios, you can imagine.
• 11:34
  I know one scenario that comes up, we didn't talk about this on Friday, but occasionally
• 11:40
  you'll see regional interaction, right?
• 11:43
  Most studies have four or five geographic regions.
• 11:45
  It does appear to be something different by region.
• 11:48
  The explanation is always, well, it's not region.
• 11:51
  It's that the BMI was different or this region enrolled younger patients.
• 11:56
  It's always that that's a confounder or masking something else.
• 12:00
  And so if you do have all the variables in the same model, in theory, that regional effect
• 12:06
  would go away and it would be more present in whatever variable is actually responsible
• 12:11
  for the difference.
• 12:13
  In Paragon, we kind of had the opposite example where we had two potential effect modifiers.
• 12:19
  We had sex and we had EF.
• 12:22
  And so conceivably it could have been that, you know, knowing that these two things are
• 12:25
  correlated with one another, maybe they were both appearing marginally significant, but
• 12:30
  the signal was actually being driven by one more than the other.
• 12:34
  It turned out that it was actually both were independently more significant that the women
• 12:41
  in the trial in Paragon generally had higher EFs, the men had lower EFs, and that when
• 12:48
  you put both in the model, you could see that it was women were responding better despite
• 12:52
  their higher EFs.
• 12:54
  Patients with lower EFs were responding better despite being a more male-dominated cohort.
• 13:00
  And so it was really by putting both in the same model that you had more confidence in
• 13:05
  both of those findings.
• 13:07
  The other thing we learned from that exercise was that when we categorize some of these
• 13:17
  things that are essentially continuous measures, we lose power.
• 13:22
  And when you look at them continuously, like ejection fraction, you end up gaining power.
• 13:30
  And so one of the tricks would be to look at the continuous variables, things like ejection
• 13:37
  fraction, GFR, age, as continuous variables rather than as categorical arbitrarily cut
• 13:46
  variables.
• 13:47
  I like that.
• 13:48
  I mean, I think we're getting to a place where we can standardize the approach, which is,
• 13:55
  I think, what your team has done, but it's not done uniformly.
• 13:59
  Let's pretend that we are an advisory panel with Dr. Abraham and Dr. Lindenfeld, and the
• 14:07
  regulators are Dr. Fusat and Dr. Sopka, and Dr. Claggett and Dr. Solomon are the investigators.
• 14:15
  They came to you with Paragon, their heart filler preserved EF trial, finished with an
• 14:20
  overall P value on the primary endpoint of.06.
• 14:24
  Brian did this fancy-.058.
• 14:26
  .058, please.
• 14:27
  .058.
• 14:28
  .058.
• 14:29
  Don't make a trauma any worse than it was.
• 14:33
  .058.
• 14:34
  And Scott has a t-shirt in that room in his office somewhere that has that.
• 14:41
  And Brian did this fancy global test on the subgroups, and it was highly statistically
• 14:48
  significant, the global, so he went down and they found that sex and EF were, the interaction
• 14:55
  terms were highly important and that there was more effect in greater signal on the lower
• 15:01
  end of preserved EF and women.
• 15:06
  And they got a label enhancement based on that.06.
• 15:12
  Now Joanne and Bill, how do you feel about that?
• 15:17
  They had a negative trial.
• 15:21
  Brian comes in with all this fancy hoopla, statistics and global chunk tests and this
• 15:26
  way and that way, and before we know it, there's a label enhancement.
• 15:32
  Yeah.
• 15:33
  Do you think they got it right?
• 15:37
  After looking at it, I think they got it right, but I think there's always a little bit of
• 15:41
  question when you have an initially negative trial.
• 15:45
  But I think other data sort of supports it, and as we look for other supports and other
• 15:49
  studies, I think we see more support of that.
• 15:52
  So I think as we see other things, like for instance with the mineralic corticoid receptor
• 15:58
  antagonists and other things that sort of look similarly, then we start to believe it even
• 16:03
  more.
• 16:04
  Yeah.
• 16:05
  I think for me, Chris, the challenge is in the biological plausibility.
• 16:10
  And we see some agents that are effective in the HEP-PEP population across the full
• 16:15
  spectrum of HEP-PEP ejection fractions like STLT2 inhibitors.
• 16:20
  And then we see others such as an ARNI or perhaps a mineralic corticoid receptor antagonist
• 16:27
  that seems to demonstrate diminishing returns with the increasing ejection fraction.
• 16:33
  And I think without a confirmatory study, we've really got to put this together in some
• 16:39
  biologically plausible story.
• 16:41
  That may exist.
• 16:42
  Some have advanced it, but that is one of the potential, I think for me, prerequisites
• 16:48
  of interpreting subgroup analysis is to understand whether or not this is founded in some biological
• 16:56
  reality.
• 16:57
  Well, you know, Chris, I would make an argument that we shouldn't probably use such a binary
• 17:05
  term like positive or negative study.
• 17:09
  The difference between a p-value of 0.06 and 0.04 is negligible.
• 17:18
  We're really talking about a few events on either side.
• 17:21
  So to argue that if we had had seven, it would have been seven more events going the other
• 17:28
  way in this trial with thousands of events, we would have had a quote unquote positive
• 17:34
  trial, everything would have been different.
• 17:37
  That doesn't make a whole lot of scientific sense.
• 17:40
  But I think what is clear is that this wasn't an incredibly robust treatment effect in the
• 17:51
  Paragon trial.
• 17:53
  And the reason we think that was the case is because it worked better in some people
• 17:57
  than others.
• 17:58
  I think the plausibility thing was so important there because it wasn't just about the statistics.
• 18:04
  It was about the fact that we had just done another 8,500 patient trial called Paradigm
• 18:10
  in Heart Cellular with reduced ejection fraction that was the adjacent population.
• 18:15
  So everything kind of kept going as you went up in ejection fraction.
• 18:20
  We had seen the exact same thing with mineralocorticoid receptor antagonists in TopCAD and Candesartan
• 18:27
  in the CHARM trial.
• 18:28
  So neurohormonal agents that petered out as ejection fraction went up.
• 18:34
  And so I think if it were just the statistics, I would agree 100% that we really have to
• 18:41
  be very circumspect about that.
• 18:45
  When we saw that result, we got very nervous about the trial that was ongoing at the time
• 18:49
  we were doing called Deliver with the SGLT2 inhibitor Dapagliflozin.
• 18:54
  And so what we ended up doing was changing the primary endpoint to split alpha between
• 19:02
  the patients in the overall group and the patients in the subgroup with an ejection
• 19:06
  fraction under 60%.
• 19:09
  We were able to divide the alpha between those two groups.
• 19:12
  And then we thought we were so clever having done this, and it turned out not to matter
• 19:18
  at all.
• 19:20
  The benefit did go across the spectrum.
• 19:23
  And Bill, I don't know why that is the case with SGLT2 inhibitors, but they're not neurohormonal
• 19:29
  modulators.
• 19:30
  They're certainly different types of drugs.
• 19:33
  And I believe behave differently from the vasodilator neurohormonal type drugs that
• 19:40
  we've been using for so many years.
• 19:43
  But that is one way you could prospectively do that.
• 19:47
  You can choose a subgroup and then apply some alpha.
• 19:52
  And Brian also presented some suggestions for how you might do that, either by being
• 19:58
  really smart and knowing which subgroup you want to choose, or by doing it more agnostically
• 20:04
  and ascribing some alpha to some heterogeneity and then figuring out after the fact where
• 20:10
  that heterogeneity actually exists.
• 20:12
  Yeah, look, I think in the case of paradigm paragon, the totality of data, including a
• 20:19
  P value of 0.058, really favored approval.
• 20:24
  And I would have voted for approval on that basis without all of the subgroup analyses
• 20:28
  anyway.
• 20:30
  I think we have to look at the big picture in these clinical trials as we think about
• 20:35
  their outcome, particularly in HEPPF where there is a large unmet need and few proven
• 20:42
  therapies that improve outcomes.
• 20:46
  And Mitch and Mona, the regulatory viewpoint, did Scott get that correct?
• 20:52
  I'm going to defer to Mitch on this.
• 20:59
  I'm going to put on a hat that is mostly regulatory and say that in this case, in the case of
• 21:07
  succubitol valcartan, there were a number of things going for it that were not just
• 21:12
  the subgroup.
• 21:13
  As was mentioned, it was an adjacent population, although there was a five percentage point
• 21:18
  ejection fraction range that was left out, but there was a closely adjacent population.
• 21:25
  But I think also very supportive were the re-adjudication analyses, the primary adjudication
• 21:31
  by the trial site PIs was supportive and then the blinded re-adjudication was very supportive.
• 21:37
  And I actually think that that was probably more of what drove the ship than the fact
• 21:43
  that it was a subgroup or anything else.
• 21:47
  I think that it was viewed overall as a positive trial once those re-adjudication procedures
• 21:54
  were put in place.
• 21:56
  I do think that all the other supportive data was very helpful that you saw a consistent
• 22:01
  trend and all these issues, but I think truly it was that the trial was probably considered
• 22:08
  to be positive overall actually, much more in line with what Bill mentioned, I think
• 22:13
  then that there was some cohesive way to view this subgroup analysis as opposed to all others
• 22:21
  where we would typically want another trial to prove that it was truly effective in the
• 22:26
  subgroup.
• 22:27
  Well, and look, the FDA didn't restrict the expanded label for the higher EF group, right?
• 22:34
  And I think that speaks volumes to really what influenced FDA's decision.
• 22:39
  And that's in contrast to, for example, and we also talked about this during the workshop,
• 22:46
  the Maid It CRT trial, which looked at the heterogeneity in response introduced by cure
• 22:54
  restoration and QRS morphology, and both supported from the data analysis and also from the standpoint
• 23:04
  of biological plausibility, that is, how does one postulate CRT or resynchronization therapy
• 23:11
  improving patients with, for example, a pure right bundle branch block?
• 23:16
  Mechanistically, it's a bit of a stretch to think about how that might work.
• 23:21
  And so the data and the biological plausibility fell together, and in fact, the label acknowledged
• 23:28
  that limitation.
• 23:30
  But let's dive into that one.
• 23:32
  Thanks for bringing that up, Bill, because that was the other really fascinating trial
• 23:36
  that we discussed.
• 23:37
  And Joanne, you were part of that initiative.
• 23:41
  Maid It CRT, give us a little bit of an overview.
• 23:44
  You studied over a thousand patients.
• 23:47
  You had a very positive effect in class 2, 3 heart failure on the endpoint of heart failure
• 23:53
  events plus mortality, all driven by heart failure events.
• 23:58
  And then the company and the leadership said there's a subgroup where the signal is even
• 24:04
  greater, even though it was like 0.006, I think, in a hazard ratio, it was pretty strong
• 24:11
  on the primary.
• 24:12
  And we want to even narrow it further to the QRS greater than 150, which was about 40%
• 24:20
  of the population.
• 24:21
  And that's where it landed.
• 24:23
  Joanne Yeah.
• 24:24
  I think in the presentation that wasn't initially the company prior to the presentation wanted
• 24:28
  to present those data and narrow the indication.
• 24:30
  And the FDA and their wisdom and their frequently right said, no, we don't want you to narrow
• 24:34
  it.
• 24:35
  You can present it all, but we don't want you to narrow it.
• 24:38
  So I think that was how it was presented.
• 24:40
  And I think as we discussed, there's still some questions about that.
• 24:44
  But I think overall other data has suggested, in addition to, as Bill said, the biologic
• 24:48
  plausibility of left bundle, clearly that is the group that benefits the most.
• 24:52
  But there are probably other benefits, but the group that benefits the most.
• 24:57
  Bill But you know, interestingly, that was not even
• 24:59
  a pre-specified subgroup for Maitrecy.
• 25:01
  Joanne Yeah, it wasn't.
• 25:02
  Absolutely.
• 25:03
  Bill That was a post-hoc finding.
• 25:05
  Bill Oh, Maitrecy, I didn't realize that, that's amazing.
• 25:10
  And it didn't come out in their primary presentation and paper, it came out subsequently.
• 25:14
  Joanne Well, the panel asked, they didn't present
• 25:16
  the data even to the FDA, the panel asked that to present those data.
• 25:22
  But the p-value for that subgroup was very small.
• 25:27
  Bill Right.
• 25:28
  Now, Dr. Fusai, you asked the think tank panel to vote on that, whether they would approve
• 25:35
  CRT for the overall population, yes, no, or would they approve it for just the subgroup
• 25:42
  QRS greater than 150?
• 25:45
  And do you remember what that vote count was?
• 25:48
  Joanne I think the majority went with the whole population.
• 25:52
  Bill They did, they did.
• 25:54
  And I think Brian raised his hand for the overall population.
• 25:58
  I'm not sure about that.
• 26:00
  You're a subgroup guy, Brian?
• 26:04
  You like subgroups?
• 26:05
  Joanne I do, yeah.
• 26:06
  Bill Why did you vote just for subgroup?
• 26:10
  Joanne No, I mean, I think if heterogeneity is there,
• 26:14
  we should care about it regardless of whether the trial was overall positive or overall
• 26:18
  marginal or overall neutral.
• 26:20
  And I think, I don't know, I feel like part of the problem that we've gotten into is that
• 26:24
  there are some scenarios where we don't talk about subgroups or we don't talk about heterogeneity
• 26:29
  at all.
• 26:30
  And that means that we only talk about it when there's a borderline trial and that gives
• 26:35
  the whole bad reputation to subgroups as a whole because it feels like it's something
• 26:39
  that we're trying to claw back a non-significant or a neutral study.
• 26:46
  And I think we should always care about heterogeneity.
• 26:49
  Bill And Chris, are we not trying to do a better
• 26:54
  job with personalized medicine and cardiology?
• 26:58
  We don't really.
• 27:00
  Most of our trials, most of our therapies are based on fairly broad populations, but
• 27:06
  I think we really do have to start thinking about which patients are going to gain the
• 27:11
  most benefit from the drugs that we use.
• 27:14
  We could probably go back and look at this with lots of the therapies that we've used
• 27:18
  for the last 20, 25 years and there are probably people who are going to benefit a lot more
• 27:24
  than others and maybe be harmed less than others because these aren't always benign
• 27:30
  therapies.
• 27:31
  Chris Yeah, I think I'm sorry, Chris.
• 27:33
  Bill Oh, go ahead, Bill.
• 27:35
  Yeah.
• 27:36
  Yeah, I was just going to pose the question, you know, when are confirmatory trials really
• 27:43
  appropriate here?
• 27:45
  We may be limiting effective therapies for patients who may benefit and in other cases,
• 27:53
  you know, we might declare a trial negative when there is a large positive subgroup that
• 27:59
  could potentially benefit in the future.
• 28:01
  So you know, whether the overall trial is positive or negative, you know, what do we
• 28:07
  do with these subgroups and do we draw firm conclusions or do we do more trials?
• 28:13
  And I understand that there may be a limited number of trials that we can do, but you know,
• 28:18
  I guess I would ask for both the clinical and the statistical view of this, is the subgroup
• 28:22
  data ever strong enough to call a definitive?
• 28:26
  Well, that's, I think you've really capsulized the question for the next podcast because
• 28:35
  we're not going to answer today for sure, but I really want to take a moment to thank
• 28:40
  our speakers and our discussion today.
• 28:43
  It was terrific.
• 28:45
  And if we take one point away from this, I think is could we standardize the approach
• 28:51
  to subgroup analysis?
• 28:52
  And I think that's what Brian really pushed forth in our session on Friday.
• 28:57
  And I'd like to see us advance that as a collaborative initiative.
• 29:02
  But more to come on subgroups.
• 29:04
  Dr. Carson, we couldn't get you on.
• 29:07
  You're driving, I know, but we're going to bring you back to talk about the V-Hef to
• 29:12
  A-Hef story, which is a 30-minute story in and of itself.
• 29:17
  And we want to make sure that we actually get to see you and not Sammy driving your
• 29:21
  Tesla.
• 29:22
  So thank you all for being here this afternoon and have a great day.
• 29:26
  Thank you.
• 29:27
  See you.