Alex Dmitrienko: Adaptive designs have
been used in dose-finding studies in oncology for several decades
(traditional up-and-down designs, continual reassessment designs,
etc). What are other popular approaches to designing adaptive
dose-finding studies and in which areas have they been used?
Dr. José Pinheiro and Dr. Frank Bretz: There are many alternative
approaches to designing and implementing adaptive dose-finding
studies. These designs can include Bayesian decision methods,
frequentist p-value combination methods, adaptive dose allocation
involving optimal design criteria for interim decisions, parametric
and non-parametric modeling approaches, etc. The common feature
of these adaptive approaches is that decisions on how to allocate
future patients to dose levels are based on data observed up to
the decision point(s). Those adaptive decisions may include dropping
dose levels altogether, or including new ones. Two years ago,
PhRMA formed a working group on adaptive dose-ranging studies,
which has investigated in a comparative simulation study several
types of adaptive dose-ranging designs, with a focus on Phase
II trials. Conclusions about the comparative performance of the
methods investigated and recommendations on the practical use
and potential gains associated with adaptive dose-ranging studies
have been released in the working group's
white
paper.
In principle, adaptive dose-finding designs can be applied to
any therapeutic area or indication. However, greater benefits
will be obtained in indications for which an endpoint is available
that can be reliably used for decision making (e.g., the clinical
endpoint itself, or a validated biomarker) and that can be measured
relatively quickly compared to the recruitment interval. Neurological
and respiratory indications often satisfy this criterion and indeed
there are examples of adaptive dose-finding studies in both. Oncology,
with its increasing availability of biomarkers, is another therapeutic
particularly well-suited for this type of designs.
What are the main objectives of adaptive designs in dose-finding
studies? Optimize the design to maximize the amount of information
about a single target dose (e.g., a dose that delivers a prespecified
level of efficacy such as ED90), optimize the design to maximize
the amount of information about the overall dose-response relationship
for an efficacy outcome variable or optimize the design to obtain
the best possible estimate of the therapeutic window (efficacious
doses with an acceptable safety profile)? The answer is most likely
indication-specific. For what classes of indications does each
of these objectives makes most sense?
Understanding and adequately representing the dose response profile
of a compound, with respect to both efficacy and safety, is a
fundamental objective of any clinical drug development program.
Proper understanding of this relationship is crucial for two critical
decisions required during the drug development process: (i) whether
there is an overall dose response effect (proof-of-concept), and
(ii) if so, which dose level(s) should be selected for further
development (dose-finding). Thus, maximizing the amount of information
about the target dose of interest or the entire dose response
shape is of primary interest for any dose-finding study.
Adaptive designs offer efficient ways to learn about the dose
response as information accrues, using it to guide decision making,
from dose allocation during the trial, to which dose to select
for further development, or whether to discontinue a program.
It is both feasible and advantageous to design a proof-of-concept
study as part of an adaptive dose-finding trial. The continuation
of a dose-finding trial, adaptive or not, into a confirmatory
stage, through a seamless adaptive design, is a further opportunity
to increase information on the correct dose earlier in development,
and thus reduce the total duration of the clinical development
program. Beyond these program level considerations, adaptive designs
have also to be tailored to the needs at the study level. Depending
on which goal is set as primary, whether safety issues are a serious
concern, and which target dose is of interest, different types
of adaptive designs can be applied and should be compared extensively
(via simulations) at the design stage of a clinical study.
Under what conditions (or for what types of indications) do
clinical trial researchers need to consider joint modeling of
efficacy and safety outcomes in adaptive dose-finding studies?
Can you comment on the use of clinical utility indices to combine
key efficacy and safety variables?
If a dose-finding trial is based on a drug which is considered
to be safe for the dose range under investigation, the adaptations
during trial conduct can focus primarily on efficacy considerations,
like choosing the dose level with the best benefit/risk ratio,
or using the optimal dose allocation design for estimating the
dose-response profile for efficacy. Otherwise, additional constraints
related to safety have to be built into the decision-making at
the interim looks. If safety is a particularly important concern
in the investigation of a given compound, it should accordingly
play a prominent role in any interim decision process. Joint modeling
of efficacy and safety can be a powerful tool towards incorporating
the total information available up to each interim look. It can
provide useful information about possible correlations between
efficacy and safety endpoints, helping in the characterization
of a benefit-risk profile. While the necessary methodology has
not yet been fully developed, one inherent problem lies in the
very nature of considering a bivariate (or possible multivariate)
response vector: Should a safe dose level be preferred over a
slightly less safe dose level, which in turn shows a much higher
efficacy? Clinical utility indices, which try to map efficacy
and safety endpoints for each patient into an aggregated univariate
measure, may help with decision making, but remain controversial.
They require a common agreement from the entire clinical team,
as well as some level of "validation" from other clinical experts.
Difficulty in clinical interpretation remains one of the main
barriers for the wider use of clinical utility indices.
Under what conditions do clinical trial researchers need to
explore adaptive dose-finding studies based on early markers of
efficacy (e.g., biomarkers or partial responses)?
The decisions at the interim looks using the accumulating data
in an ongoing study need to be taken quickly and reliably. In
many cases, however, the follow-up time needed to observe the
clinical endpoint of interest is large compared to the overall
recruitment interval. For example, time-to-event endpoints (e.g.,
progression-free survival) are often employed as primary clinical
endpoints in oncology trials (with a median value of several months,
or even years), although the enrollment time for the entire study
might only take one year, or even less. Performing an adaptive
design based on the complete information per patient is then not
feasible, since at the time when enrollment has finished one might
have only a few patients who have completed the trial. In such
cases, early readouts of the final clinical endpoint, or the use
of biomarkers may be the only viable solution. The underlying
assumption is, of course, that they can be used as surrogates
of the final endpoint for the purpose of deciding on the subsequent
adaptations. It is, of course, critical for the success of the
adaptive approach, that the surrogate assumption be carefully
evaluated and, whenever possible, validated.
Under what conditions do clinical trial researchers need to
consider continuous (each new patient is assigned to his/her dose)
and group-sequential (a group of patient is assigned to a dose)
allocation of patients in adaptive studies?
Irrespective of whether continuous or group-sequential allocation
is adopted, adaptive designs assume that the flow of information
to the data analysis center is sufficiently rapid (e.g., instant
data capture and fast transmittal) to allow adaptations to be
performed according to the prescribed methodology. Electronic
data capture coupled with an effective computer supported data
vetting and cleaning is almost a must. Also, in applications where
drug supply is affected by the adaptations, an interactive voice
response system (IVRS) is needed. The use of such technologies
are thus indispensable pre-requisites for any adaptive study.
The decision of whether continuous or group-sequential allocation
should be adopted has therefore to be considered in a broader
context, taking into account the larger logistical challenges
associated with a continuous allocation. There will be a trade-off
between the total number of patients, the recruitment rate, the
time to follow-up and the statistical benefit associated with
updating the overall information after each patient being observed.
Considerations of these types are specific to the particular application
and need to be evaluated carefully for each trial.
