For more infomation >> Özgür Can Çoban - Yaralandım (Official Video) - Duration: 5:09. 
For more infomation >> Utku Uzun & Ali Usta - ÇOCUK ADAM (Official Video) - Duration: 2:40. 
[MUSIC PLAYING]
Hi.
My name's Scott Berry.
I'm a biostatistician at Berry consultants.
And I'm going to walk through examples
of utility functions, modeling efficacy and safety, efficacy
and tolerability, using utility functions.
I'll show you several examples of these and good practices.
So the first example I'm going to talk about
is a trial in stroke.
The endpoint at 90 days in this example
is the modified Rankin score.
You can see, on the table on the right, the seven levels
of the modified Rankin score.
It ranges from level zero, which is no symptoms, perfect status,
neurological status, down to level six, which is dead.
And you can see that the levels vary, one, two, three, four,
five, and six.
An interesting thing about this is
the number does not classify the relative difference
in the states.
It's just a classifier.
You could label these A through G as the label for the level.
The question comes up about how do we analyze this?
What is the statistical analysis of this particular endpoint?
And there are a number of different ways to analyze this.
We had a particular trial where this was the end point.
And the comment was made that let's analyze it
where we lump in zero, one, and two as one
particular dichotomous category, and then three, four, five,
six is separate.
And so we dichotomize it.
The graph on the right shows you the relative utility
of those seven states.
It's equal to be zero, one, and two,
and equal to be three, four, five, and six.
And the comment was made is let's analyze it mathematically
this way.
We separate them mathematically, and then
when the trial is over, we'll think about clinical importance
in the trial.
This-- I didn't like this.
I felt like this was bad science.
It separates the statistical analysis from clinical benefit.
So I feel like this is a statistical mathematical test
that doesn't tie in well to the design
to the clinical question.
As clinical trials are becoming more innovative,
we can't do this that we wait till the end of the trial
to address the important issues.
Adaptive trials, response adaptive randomization,
enrichment--
I'll show you some examples of these--
you need to be able to classify that upfront and prospectively
so that the design does smart things during the course of it.
So we need to address these questions upfront.
Utility functions are very nice ways to do that.
So here is an example trial.
It's called the DAWN trial.
It was a trial in a clot clearing
device for post stroke.
Now, this is approved by the FDA within eight hours,
but this was a trial looking at subjects beyond that period.
It actually enrolled between six and 24 hours.
And clinically, it looked for a mismatch.
The primary endpoint in the trial
is 90 day modified Rankin score.
We propose the primary analysis to be a utility weighting
of those seven outcomes, in part, because
of the complexity of the design itself.
The design had enrichment steps to it,
where we might stop enrolling certain types of strokes
if the device was not beneficial for that.
This is an adaptive enrichment step.
There was also early stopping for futility
in success in the trial.
These are prospectively defined rules
that we need to have set up so we can't do this aspect of, oh,
let's do a mathematical test and think hard about it afterwards.
We need to incorporate that prospectively upfront.
So in this trial, we created a utility function.
So what is a utility function?
A utility function is a mapping from a set of parameters.
So in this example, it could be the probability
of these seven outcomes maps it to the real line.
What is nice about this is it maps
what could be multiple end points into a single measure.
This utility measure on the real line
doesn't have units it's not in terms
of dollars or anything else.
It's a relative value of doses or arms
that we can directly compare on that measure of utility.
So let's look at that example of the DAWN trial.
On the graph, on your right side over here,
you'll see the dichotomous utility
is the red, where we value zero, one, and two the same.
In this one, the blue dots represent the utility score
that was used of each particular outcome, the utility
weighted MRS, and you can read about it in the citation here.
You can see the value of these within the zero, one, and two
have monotonically decreasing states,
but a two has different utility than a one.
It's a worse clinical state, and one is worse than zero.
And the height of these is relatively important.
You can see the drops from zero, one, and two
are relatively smaller than the drop from three to four
or four to five, which clinically
is a more important drop.
This is the utility that was used
in the DAWN trial for analysis.
The DAWN trial ended its stop for success early in the trial.
It never enriched.
So it kept enrolling the large population.
You can see here this figure and table
from the publication of the paper shows you the data.
It shows the relative number of zeros
through sixes on the active device arm and the control arm.
And then the bottom table shows the primary analysis
of the utility weighted MRS. The posterior
probability of superiority of the active arm, which
was greater than 0.99, and the relative difference
on the utility scale.
So you can see what a primary analysis of a utility
looked like.
It was really important in this trial
to incorporate the relative value of these six states
so that we incorporated the efficacy as well as safety.
The potential that we could create more negative outcomes
is directly captured by the end point in the trial.
Now, utility functions.
That's an example where we have these seven
levels of a single modified Rankin score.
Other common uses of utility functions
are to incorporate efficacy and safety or efficacy
and tolerability directly in a single measure.
And again, as we do innovative designs
and we're doing adaptive designs,
we can incorporate all of these into a single endpoint
so the adaptive design does smart adaptations
in the course of the trial.
We're used to doing things like an ED90, which
is, in a way, a poor man's utility function
to say, oh, let's get 90% of the efficacy
and not push it too high and target that,
because we don't know about tolerability and safety.
This is a way of directly capturing
safety and tolerability in your end
point of interest, the utility.
Let's look at an example of this.
So this is an example efficacy curve over a range of doses.
You can see these 10 doses across there,
and the black line on this curve represents the probability
of a positive efficacy outcome.
Let's use an example of a treatment for insomnia,
that this represents the proportion
of patients that have a good night's
sleep after taking the drug.
That's efficacy of the drug.
You can see the yellow line here represents the ED90,
the effective dose 90.
Whether that's the right dose depends entirely
on the other endpoints, the tolerability.
So let's look at potential tolerability curves.
So I've added here three different colored curves
that represent the probability of a tolerability issue that
arises.
In this example, it's easy to think about
that you're overly fatigued the next morning after taking
that particular treatment.
There is a blue curve here, and a green curve here,
and a red curve here representing
the three different levels that could
be for the different doses.
You would love it to be the green curve, because that has
a less chance of tolerability.
In this example, in the green curve,
you would push the dose higher, because you
can get more efficacy with less relative tolerability, where
the light blue curve, you would want
to push the dose down farther.
You still get efficacy, but you get less tolerability.
The ED90 is the same dose regardless.
So we could attach utility functions
to the relative weight of efficacy and tolerability
in this example.
Let's look at several examples of this.
It will use a simple utility function, and as you can see,
where the utility function is, the differential
between the placebo and each dose on efficacy,
subtracting off the added probability
of a tolerability event.
So we're really looking at the difference
between the black curve and each of the tolerability curves
with an equal weighting of efficacy and tolerability.
You could imagine a coefficient of tolerability making
them more important or less important,
or more complicated functions, but this
is a simple function, really, of looking
at the space between the two curves is the relative value
or utility of a dose.
So with that utility function, the graph
on the right side of the slide here
shows the utility of each of the doses using the utility
function we describe in equal weighting of efficacy
and tolerability.
On the bottom of the graph, on the x-axis,
you can see the maximum utility dose
accounting for both of these endpoints,
and in the case of the blue to the green,
you can see, the higher the dose,
the more utility when tolerability is less.
And these are all different than the ED90.
So this is a really nice tool for phase two trials,
for adaptive trials to incorporate efficacy and safety
or tolerability into a single endpoint.
I'll describe an example of a phase two trial that
ran using this technology.
The drug is Trulicity.
Here is a reference you can look at
to look at the design of the paper.
It was a phase two, three seamless trial
that had seven doses in phase two
with response adaptive randomization.
During the course of the trial, every two weeks,
the randomization probabilities for the seven doses
were updated to favor doses that were performing better.
The way we judged better was based
on four different endpoints all combined
together into a single utility function that
measured the therapeutic benefit to a patient
of each particular dose.
An algorithm selected the doses to seamlessly move
to phase three.
It also spanned other phase three trials.
Those doses selected by that utility function
are now the marketed doses of Trulicity.
The example in that trial, the utility functions
are shown on the right of the slide.
You can see how we combine together
the utility of an efficacy endpoint change from baseline
in HBA1C relative to an active comparator.
We combine heart rate, and blood pressure, and weight loss all
into a single utility.
The way we did that in this trial
was by multiplying each of the four components
you see on the right of your slide
into a single utility function.
And you can see, for example, if heart rate or blood pressure
had too high of a difference from placebo,
the relative value of the dose would be near zero,
and we would go back to doses that
were smaller that had high efficacy,
but had a better profile for safety or for weight loss.
This was used in an adaptive trial
in that particular example, and you
can see the power of this within an algorithm.
So to sum up what I've talked about,
utility functions can be incredibly powerful tools
for capturing multiple endpoints into a single function.
Very powerful as we're looking at innovative designs,
adaptations, algorithm based decisions in this.
Now of course, the difficulty in these trials
is creation of the utility function,
eliciting that utility function for your particular scenario.
This can be challenging and to represent
whose utility, are we looking at patients,
are we looking at financial, are we
looking at relative value of doses?
But this is real--
these are really the right scientific questions
we should be addressing to drive the design
into the right space.
Thank you.
[MUSIC PLAYING]
Đăng ký:
Đăng Nhận xét (Atom)
Không có nhận xét nào:
Đăng nhận xét