WHITE PAPER Adaptive Clinical Trial Design James Rosenzweig and David McSorley

Clinical trials in the pharmaceutical industry are incorporating adaptive design methods with greater frequency, as the economic resources needed for clinical research have expanded. These designs leverage accumulating information during a trial in real time and have the potential to reduce the costs and streamline the time frames for clinical trials in drug development, particularly in the earlier phases during proof of concept and dose selection.

Adaptive design is usually defined as the According to the FDA guidelines, an adap- changes should be made during the course use of accumulating data obtained during tive clinical trial can involve: of the study. Adaptive designs use interim the conduct of the trial to decide how to • Analysis at decision points during the analyses of the accumulating data from modify aspects of the trial as it progresses, trial to stop or to adjust patient accrual within an ongoing study to modify various without affecting its validity and integrity aspects of the trial and then continue (Gallo et al., 2006). This form of research • Interim evaluations to determine if under the modifications. The different design involves a prospectively planned the trial should be stopped early—be- types of adaptations in study design have opportunity to modify the parameters of cause of a determination of success, often been put into categories, which the study based on analysis of interim data demonstration of futility, or finding of helps clarify the specific issues to be dealt while the study is being performed. The unacceptable harm to subjects with (Chow and Chang, 2008). However, prospective planning of key time points • Hypothesis reversal of noninferiority to categories may overlap and studies may for interim analysis, and the parameters for superiority or vice versa combine multiple strategies, as seen in the change in the study, have to be set before following: • Discontinuation of arms or doses, or the study is underway. the changing of doses while the trial is 1. Adaptive randomization Adaptive clinical trial designs were further underway These designs allow for changing the characterized by the US Food and Drug • Modification of the randomization rate randomization schedules of a study by ad- Administration (FDA) (FDA, 2010). to increase the probability that a patient justing the probability of treatment assign- The FDA divides adaptive designs into is allocated to the most appropriate arm ment based on prior assignments in order two categories—“well understood” and to either avoid an imbalance of important “less well understood.” Those considered Others have argued for a more liberal patient characteristics between treatment well understood have a record of being definition of adaptive design, which allows groups or to increase the likelihood of performed in the past, with established not only for prospective adaptations, but being assigned to a particular treatment statistical methods and familiarity with the for concurrent (ad hoc) and retrospec- group. Adaptive randomization schemes FDA from previously approved studies. tive adaptations (Chow, 2014). The use include treatment-adaptive (TA), covari- Less well understood designs fail to meet of Bayesian methodologies can enable ate-adaptive (CA), and response-adaptive these criteria. greater insight on which options for design (RA) randomization. Treatment-adaptive

The power of knowledge. The value of understanding. WHITE PAPER Adaptive Clinical Trial Design

and covariate-adaptive designs aim to An example of a group sequential design and education to implement this. A total balance the treatment groups with respect that employed adaptive elements was of 1,375 patients agreed to participate in to patient characteristics by changing the the Diabetes Control and Complications the EDIC study, the open-label follow-up way the subsequent patients are assigned to Trial (DCCT) (DCCT, 1993) and its to the DCCT, which continues to this day. a treatment group, whereas response-adap- follow-up, the Epidemiology of Diabetes The EDIC study continues to monitor the tive randomization aims to increase the Interventions and Complications (EDIC) DCCT patients for level of complications probability of being assigned to the treat- study (Nathan et al., 2005). The DCCT and cardiovascular events. Although the ment group with more favorable responses. was a multicenter, randomized clinical level of glycemic control in the two groups Response-adaptive schemes can therefore trial designed to determine whether or came together not long after the random- cause imbalances in patient characteristics not very tight control of blood glucose ized phase was completed, it was found that may require subsequent adjustment for (experimental therapy at the time) could that those in the intensive arm for only 6.5 these imbalances resulting in a combined reduce or prevent the microvascular and years continued to have fewer microvascular RACA design (Ning and Huang, 2010). macrovascular complications of type 1 complications for at least 30 years of fol- Adaptive randomization schemes have the diabetes, when compared with the standard low-up, and macrovascular benefits as well. most utility in small (n < 100), early-phase insulin treatment of the day. The study had trials—where equal probability randomiza- parallel arms consisting of subjects with 3. Flexible sample size re-estimation tion may not produce the desired balance no complications and those with very early This design enables the size of the sample in patient characteristics among treatment complications. The study was initiated in in the study to be changed or re-estimated groups—as the designs quickly become im- 1982. A feasibility phase with 278 subjects based on unblinded interim effect size data practical for large or longer-duration trials. was completed in 1985, and an interim and often may be included as one of the analysis determined that the experimen- adaptations in an adaptive group sequen- 2. Adaptive group sequential design tal intervention was safe and effective in tial design. In a fixed-sample study design, Classical group sequential methods use improving glycemic control by a sufficient the sample size is determined before the repeated significance testing on accu- margin to allow for full recruitment, which study and is based on prior estimates of the mulating groups of enrolled patients to expanded to 1,441 subjects. A specific clinically meaningful effect size between the decide whether to stop or continue a trial concern had been that the intervention treatment and control groups that can be based on established stopping boundar- to intensify glycemic control would be achieved for a specified power and type I er- ies for each test that maintain the overall accompanied by an unacceptably high risk ror rate. It is not uncommon for effect sizes type I and type II error rates across all of severe hypoglycemia; this was found not to be initially specified incorrectly, resulting tests. Type I error occurs when the null to be the case. An independent data, safety, in an underpowered design, especially if the hypothesis is true and type II error occurs and quality committee (DSQ) followed the variability turns out to be larger than initially when the null hypothesis is false but is results on a regular basis while the investi- specified. As a result, it may be desirable to not rejected. The concept of adaptive gators were blinded. adjust the sample sizes based on accrued design allows for additional changes to a data while a trial is underway. However, any study protocol as it progresses as a result The full randomized controlled clinical sample size re-estimation should be planned of analysis of interim data (Bauer and trial phase was stopped prematurely after in advance and done using appropriate Köhne, 1994). These include potential a mean follow-up time of 6.5 years, when group sequential methods so as to preserve modification, deletion or addition of treat- the benefits of intensive treatment were the type I error rate. ment arms, re-estimation of the sample found to be incontrovertible by the DSQ size, change of study endpoints, changing and not likely to be reversed over time. At 4. “Drop the losers” of dose and/or duration of treatment, that point, subjects on intensive control When multiple treatment arms are used, it and modification of randomization sched- were encouraged to continue and those is often helpful to have a multistage design ules. Adaptive group sequential designs originally assigned to conventional treat- to enable the investigators to drop arms combine the concepts of both early ment were advised to switch to intensive that are shown to be inferior to others. stopping and re-engineering of the design treatment. During the closeout phase of This design is sometimes referred to as based on the observed early results. the trial, they were provided the resources selection design or “pick up the winners,”

The power of knowledge. The value of understanding. WHITE PAPER Adaptive Clinical Trial Design

as it also allows adding additional arms decreased. If one subject experiences DLT, method of Branson and Whitehead (2002) (Chow, 2014). Typically, it is the first then three subjects are added. If only one gave accurate and consistent results, with an stage of a two-stage design, in which the of the six subjects experiences DLT, the advantage to the IPE method. inferior arms are dropped according to dose can be increased, and if two or more criteria specified in the beginning of the have DLT, then the dose is decreased. 7. Adaptive hypothesis design study. The winning treatment groups go It is possible to make potential changes to on to the next stage of the study. It is also Increasingly, dose-finding studies have the hypothesis of a study based on interim possible to use different analytic approach- utilized an iterative model-fitting process, data that is collected. This can be done by es (e.g., Bayesian predictive probabilities often called the continuous reassessment applying the closure principle (Marcus et vs. frequentist hypothesis testing) for the method (CRM), to find the MTD. A al., 1976) to the hypotheses of interest and progression criteria between stages, so the number of studies have shown though testing each of them by using an appro- study needs to be designed to have suffi- simulations that CRM model-based designs priate combination test (Bretz et al., 2006). cient power for those stages using a fre- are more accurate and effective than the Some examples include changing from a quentist hypothesis testing approach (e.g., 3+3 design. They are able to determine the single hypothesis to multiple hypotheses or at the end of the trial). It is possible that MTD more quickly, and a greater percent- a composite hypothesis, switching from a dropping or adding the wrong dose groups age of the subjects treated in these studies superiority hypothesis to a noninferiority could lead to loss of valuable information are found to be at or near the MTD. These hypothesis, switching between the null and that would have been helpful at the end of newer CRMs employ dose-escalation algo- alternative hypotheses, or changing the the study. Because of this, it is important rithms that emphasize overdose control, or primary and secondary endpoints. that optimize the time to event or late-dose to use valid and well-considered decision 8. Phase 1-2 or phase 2-3 seamless rules or criteria for selecting dose groups. toxicities to refit the dose-toxicity curve after each dose level’s toxicity outcome is trial design Adaptive seamless design is used to com- 5. Adaptive dose finding observed (Garrett-Mayer, 2006). In gen- Study designs that employ adaptive finding eral, it has been shown that these designs bine the aims and objectives of what would are often used in phase 1 or 2 studies in do not pose major safety concerns. On av- normally be considered separate trials into order to determine the maximum tolerated erage about 25 to 35 subjects are required one study. Most likely, a phase 1 study dose (MTD) of the medication, which is of- to test 5 or 6 dosage levels (Iasonos and would be combined with a phase 2 study of ten used as the optimal dose for later-phase O’Quigley, 2014). the same compound. Similarly, phase 2 and clinical studies. It helps to avoid having too phase 3 studies can be combined. Typically, many subjects exposed to dose-limiting tox- 6. Adaptive treatment switching a phase 1 trial to establish the MTD of a icity (DLT), and a small number of subjects In this situation, the design of the study drug can be combined with an early phase can be used to identify the MTD. To achieve can permit the investigator to switch the 2a trial to investigate the efficacy of the this end, careful selection of the appropriate patient to an alternative treatment if there is drug at that dose. A phase 2b dose-ranging initial dose is important, as well as the dose evidence of lack of efficacy, progression of study can be combined with a confirma- range and parameters for dose escalation or disease, or safety problems with the initial tory phase 3 trial with more subjects and dose reduction. medications. This is commonly used in on- investigational sites, and perhaps different cology trials because of compassion issues endpoints. It could be set up as a two-stage In early-phase oncology studies, it is often related to the consequences of withholding study, with the interim analysis serving as a difficult to balance toxicity with clinical a possible beneficial treatment. The statisti- decision point for whether the trial should effectiveness. One dose-finding method cal analysis must also adjust for the treat- be stopped or expanded. With a seamless commonly used the 3+3 design. With ment switching. In an evaluation of nine phase 2/3 design, valuable information can this procedure, three subjects receive a methods that adjust for treatment switch- be obtained in the first stage that could help particular dose of study medication. If no ing, Fox et al. (2011) found that only the in decisions made during the conduct of the patients experience DLT, then the dose is rank-preserving structural failure time (RPS- second stage, in particular which dose(s) to increased by a predetermined amount. If FT) model of Robins and Tsiatis (1991) retain in stage 2. Because this design would two or more experience DLT, the dose is and the iterative parameter estimation (IPE) allow for use of data acquired from both

The power of knowledge. The value of understanding. WHITE PAPER Adaptive Clinical Trial Design

stages, there can be some economies of it is imbedded in a trial to modify patient data is obtained outside of the trial, an scale. Costs can be saved through com- eligibility after the interim analysis, the adaptive design can utilize it in the course bining of evidence across the two stages. statistical methodology must account for of the study. It is quite common for the Sample size can be reduced in comparison how the data collected before and after standard of optimal care in treatment of a to running two separate studies. There the interim analysis will be combined and particular disease to change while a study is would be no lead time between the two analyzed at the end of the study. in progress. For ethical reasons, and because studies so that time can be saved. Instead of of the possible effect of the new treatment starting anew with institutional review board An example of this design has been the on the outcomes of the trial, it may be approval, site recruitment, and subject “Biomarker-integrated Approaches to advisable to consider altering the protocol enrollment, the process would be expanded Targeted Therapy of Lung cancer Elimi- to allow addition of the new treatment in seamlessly in the second phase of the study. nation (BATTLE) trial (Kim et al., 2011). an unbiased way to all of the groups under However, extra planning is necessary and This trial included patients with stage IV investigation. Design that makes use of the statistical methodology must account recurrent non–small-cell lung cancer. The interim data can give investigators new for potential biases and multiple looks at the primary endpoint was the 8-week disease options to modify or re-evaluate the trial data, and how to combine the data from the control rate (DCR). Four biomarker profiles while it is underway. Adaptive design can different stages to make sure that the overall were examined, and four different drug enable investigators to respond to positive validity of the study can be maintained. therapies were employed, with one therapy or negative surprises from data obtained targeting each biomarker profile. The trial while the study is in progress. It can be used 9. Biomarker-adaptive design looked at the four biomarkers and aimed to to stop the study earlier when it becomes Biomarkers may be collected in some identify their predictive roles in providing clear that there is no benefit to the treat- studies to detect normal function or assess better treatment efficacy in terms of the ment. In general there are many features of pathogenic or pharmacological processes DCR to patients in the trial based on their adaptive design that may help to decrease in response to the therapeutic agent under biomarker profiles. A Bayesian hierarchical the length of studies and shorten the time investigation. However, they should not be model was used to adaptively assign patients of development of investigational drugs. confused with primary endpoints. A bio- to one of four treatment groups using the However, it is important to recognize that marker that correlates well with a clinical patient’s biomarker profile to estimate the as part of the design, all potential adapta- endpoint can be considered a prognostic posterior probability of the DCR. The study tions that may be undertaken during a trial biomarker. These can be used to identify also had an early stopping rule for futility, should be prespecified with the objective of information about the natural course of in order to drop the potentially inferior improving the likelihood of a successful and the disease being studied, irrespective treatments from the options available for informative trial. of whether the subject is randomized to newly enrolled patients. The study overall the treatment in question. At the start of had a 46% DCR and identified a higher rate The biggest challenge with these designs is the study, prognostic biomarkers can be in sorafenib-treated patients with a specific managing the additional logistical complex- used to stratify patients by good or poor biomarker (KRAS mutations), although ities and operational details that, without prognosis or disease severity, for efficiency sorafenib also resulted in more treatment considerable preplanning and careful execu- of recruiting, or for subgroup analyses to discontinuations and dose reductions. tion, could impair the validity and integrity identify the degree of expected responsive- of a study. It is extremely important that the ness or sensitivity to the treatment being Conclusion interim analyses be done by an independent studied. They should not be used to select Overall there are many potential advantages group (e.g., data monitoring committee) in the particular treatment under study. Bio- in using adaptive designs in clinical studies. order to reduce the possibility of introduc- markers can be used to identify a particular They can help in earlier selection of the tion of bias. There may be operational bias- therapy for use during the study in affected most promising patient characteristics or es introduced when using adaptations, and patients. They are most often used in therapeutic options. If there are mistaken if care is not taken, these may change the exploratory studies to identify the appro- assumptions that have been made prior to trial into a different one that can no longer priate criteria for patients to be selected the study, adaptive design can be used to address the original questions that need to for participation in later trials. However, if correct them midstream. If relevant new be answered.

The power of knowledge. The value of understanding. References Bauer P, Köhne K. Evaluation of exper- Fox R, Billingham L, Abrams K. Eval- Nathan DM, Cleary PA, Backlund JY, Ge- iments with adaptive interim analyses. uation of methods to adjust for treat- nuth SM, Lachin JM, Orchard TJ, et al. In- Biometrics. 1994 Dec;50:1029-41. ment switching in clinical trials. Trials. tensive diabetes treatment and cardiovascu- 2011;12(Suppl 1):A139. lar disease in patients with type 1 diabetes. Branson M, Whitehead J. Estimating a N Engl J Med. 2005 Dec;353(25):2643-53. treatment effect in survival studies in Gallo P, Chuang-Stein C, Dragalin V, et al. which patients switch treatment. Stat Med. Adaptive design in clinical drug develop- Ning J, Huang X. Response-adaptive ran- 2002 Sep;21:2449-63. ment—an executive summary of the PhR- domization for clinical trials with adjust- MA Working Group (with discussions). J. ment for covariate imbalance. Stat Med. Bretz F, Schmidli H, König F, Racine A, Biopharm Stat, 2006; 16:275-83. 2010 Jul;29:1761-8. Maurer W. Confirmatory seamless phase II/III clinical trials with hypotheses selec- Garrett-Mayer E. The continual reassess- Robins JM, Tsiatis AA. Correcting for tion at interim: general concepts. Biometri- ment method for dose-finding studies: a non-compliance in randomized trials using cal Journal. 2006 Aug;48(4):623-34. tutorial. Clin Trials. 2006; 3:57-71. rank preserving structural failure time models. Communication in Statistics -The- Chow SC, Chang M. Adaptive design Iasonos A, O’Quigley J. Adaptive ory and Methods. 1991;20(8):2609-31. methods in clinical trials – a review. Or- dose-finding studies: a review of mod- phanet J Rare Dis. 2008 May;3:11. el-guided phase I clinical trials. J Clin US Food and Drug Administration. Guid- Oncol. 2014 Aug;32:2505-11. ance for industry: adaptive design clinical tri- Chow SC. Adaptive clinical trial design. als for drugs and biologics—draft guidance. Annu Rev Med. 2014;65:405-15. Kim ES, Herbst RS, Wistuba II, Lee JJ, February 2010. Available at: http://www. Blumenschein GR Jr, Tsao A, et al. The Diabetes Control and Complications Trial fda.gov/downloads/Drugs/.../Guidances/ BATTLE trial: personalizing therapy ucm201790.pdf. Accessed May 3, 2016. Research Group. The effect of intensive for lung cancer. Cancer Discov. 2011 treatment of diabetes on the development Jun;1(1)44-53. and progression of long-term complications in insulin-dependent diabetes mellitus. N Marcus R, Peritz E, Gabriel KR. On closed Engl J Med. 1993 Sep;329(14):977-86. testing procedures with special reference to ordered analysis of variance. Biometrika. 1976;63:655-60.

The power of knowledge. The value of understanding.