Adaptive DoseResponse Studies

1
Adaptive Dose-Response Studies

• Inna Perevozskaya
• Merck Co,Inc.

2
Acknowledgement PhRMA Adaptive Designs Working
Group

• Co-Chairs
• Michael Krams
• Brenda Gaydos
• Authors
• Keaven Anderson
• Suman Bhattacharya
• Alun Bedding
• Don Berry
• Frank Bretz
• Christy Chuang-Stein
• Vlad Dragalin
• Paul Gallo
• Brenda Gaydos
• Michael Krams
• Qing Liu
• Jeff Maca
• Inna Perevozskaya
• Jose Pinheiro

• Members
• Carl-Fredrik Burman
• David DeBrota
• Jonathan Denne
• Greg Enas
• Richard Entsuah
• Andy Grieve
• David Henry
• Tony Ho
• Telba Irony
• Larry Lesko
• Gary Littman
• Cyrus Mehta
• Allan Pallay
• Michael Poole
• Rick Sax
• Jerry Schindler
• Michael D Smith
• Marc Walton

3
Upcoming DIJ Publications by PhRMA working group
on adaptive designs

• P. Gallo, M. Krams Introduction
• V. Dragalin Adaptive Designs Terminology and
  Classification
• J. Quinlan, M. Krams Implementing Adaptive
  Designs Logistical and Operational
  Considerations
• P. Gallo. Confidentiality and trial integrity
  issues for adaptive designs
• B.Gaydos, M. Krams, I. Perevozskaya, F.Bretz Q.
  Liu, P. Gallo, D. Berry C. Chuang-Stein, J.
  Pinheiro, A. Bedding. Adaptive Dose Response
  Studies
• J. Maca, S. Bhattacharya, V. Dragalin, P. Gallo,
  M. Krams, Adaptive Seamless Phase II / III
  Designs Background, Operational Aspects, and
  Examples
• C. Chuang-Stein, K.Anderson, P. Gallo, S.
  Collins. Sample Size Re-estimation A Review and
  Recommendations

4
Dose-Response Paper Overview

• Motivation Challenges in evaluation of
  dose-response
• Summary of key recommendations from PhRMA
  dose-response workstream
• Overview of traditional dose-response designs
• Overview of adaptive dose-response methods in
  early exploratory studies
• Adaptive Frequentist approaches for late stage
  exploratory development
• Developing a Bayesian adaptive dose design
• Monitoring issues and processes in adaptive
  dose-response trials
• Rolling dose studies

5
1. Motivation Challenges in Evaluation of
Dose-Response

• Insufficient exploration of the dose response is
  often a key shortcoming of clinical drug
  development.
• Initial proof-of-concept (PoC) studies often rely
  on testing just one dose level (e.g. the maximum
  tolerated dose)
• Additional exploration of dose-response
  typically done later (Phase IIb trials)
• Adaptive designs offer efficient ways to learn
  about the dose response and guide decision making
  (dose selection/program termination)
• It is both feasible and advantageous to design a
  PoC study as an adaptive dose response trial.
• Continuation of a dose response trial into a
  confirmatory stage in a seamless design is a
  further opportunity to increase information on
  the right dose(s)
• Adaptive dose-response trial may offer deduction
  in the total clinical development timeline

6
1. Motivation (cont.)

• Efficient learning about the dose response
  earlier in development could reduce overall costs
  and provide better information on dose in the
  filing package
• This review primarily focuses on phase Ib and II
  study designs
• Applicable to endpoints that support filing or
  are predictive of the filing endpoint (e.g.
  biomarkers).

7
2. Key recommendations of the PhRMA Adaptive
Dose-Response workstream

• Consider adaptive dose response designs in
  exploratory development.
• Consider adaptive dose response designs to
  establish proof-of-concept
• Whenever possible use an approach that
  incorporates a model for the dose response.
• Consider seamless approaches to improve the
  efficiency of learning
• Define the dose assignment mechanism
  prospectively and fully evaluate its operational
  characteristics through simulation prior to
  initiating the study.

8
Key recommendations of the PhRMA Adaptive
Dose-Response workstream (cont.)

• Stop the trial at the earliest time point when
  there is enough information to make the decision.
  
• A committee must monitor the study on an ongoing
  basis to verify that the performance of the
  design is as expected
• Engage the committee early in scenario
  simulations prior to protocol approval.
• Leverage the information from disease state and
  exposure-response models to design studies.

9
3. Traditional methods to explore dose-response

• Fixed-dose parallel-group designs
• Target
• average population response at a dose
• shape of the population dose response curve
• Downside
• potential to allocate a fair number of patients
  to several non-informative doses
• sample size considerations often limit the number
  of doses feasible to explore
• Fixed dose cross-over design,
• Forced titration design,
• Optional titration design

For both efficacy and safety
Primarily aimed at learning about individual
dose-response
10
4. Adaptive dose-response methods for early
exploratory studies

• Review traditional phase I designs with respect
  to estimation of the Maximum Tolerated Dose (MTD)
• Discuss novel adaptive design approaches aimed at
  improving the relatively poor performance of
  traditional designs
• Majority of these methods originated from
  oncology
• There is methodological overlap with other dose
  response methods (late stage)
• applicability can be generalized from MTD
  determination to learning about the dose response
  profile for any defined response (e.g.
  tolerability, safety or efficacy measure)
• More work generally needs to be done to extend
  applicability beyond the area of cancer research
  

11
Specifics of oncology Phase I trials

• Typically very small, uncontrolled sequential
  studies in patients
• Binary outcome toxicity response
• Designed to determine the maximum tolerated dose
  (MTD) of the experimental drug
• Design challenges are driven by severe side
  effects of cytotoxic drugs, limited number of
  patients available
• Certain degree of side effects is acceptable, but
  every effort should be made to minimize exposure
  to highly toxic doses
• Balance between individual and collective ethics
  maximum information from the minimal number of
  patients
• Be open-minded the designs presented here
  originated in Phase I oncology
• But can be potentially be useful for other early
  development trials (e.g. efficacy assessment
  dose-ranging, POC)

12
Statistical modeling of efficient learning about
MTD

• Schacter et al. (1997) a well-designed phase I
  study will identify a dose at which patients can
  be safely treated and one which can benefit the
  patient.
• Assumption monotone relationship between the
  dosage and response
• Two different philosophies in MTD definition
• 1.Risk of toxicity is a sample statistic,
  identified by the doses studied
• e.g., 33 design MTD is highest dose studied
  with lt 1/3 toxicities
• 2.Risk of toxicity is a parameter of a
  dose-response model
• e.g., dose associated with 30 incidence of
  toxic response
• Two different approaches in designing phase I
  clinical trials

13
Summary of available methods for phase I
clinical trials (Rosenberger and Haines, 2002)

• 1. Conventional (standard) method (Simon et al.,
  1997 Korn et al., 1994)
• 2. MTD as a quantile vs. conventional method
• a) Random walk rules (RWR)
• Durham and Flournoy (1994)
• b) Continual reassessment method (CRM)
• OQuigley, Pepe, Fisher (1990)
• c) Escalation with overdose control (EWOC)
• Babb, Rohatko, Zacks (1998)
• d) Decision-theoretic approaches
• Whitehead and Brunier (1995)
• e) Bayesian sequential optimal design
• Haines, Perevozskaya, Rosenberger (2003)

Bayesian Methods
14
Learning about MTD current and novel designs
summary

• Key feature Prior response data used for
  sequential allocation of dose/treatment to
  subsequent (group of) subjects
• Up-and-down type designs utilize only last
  response in decision rule
• Conventional 33 designs for cancer (traditional)
• Random-walk-rule designs
• Bayesian type designs all previous responses
  from the current study are utilized
• Continual Reassessment Method
• Other Bayesian approaches
• Common goal limit allocation to extreme doses of
  little interest

15
1. Current practice in MTD estimation
conventional 33 design for cancer

• Designed under philosophy that MTD is
  identifiable from the data
• Patients treated in groups of 3
• Designed to screen doses quickly no estimation
  involved
• Probability of stopping at incorrect dose level
  is higher than generally believed (Reiner,
  Paoletti, OQuigley 1999)

16
33 design for cancer Pros () Cons (-)

• has been around for a long time, properties
  well documented
• - estimates MTD at 20 toxicity level
• Weili He, et al., show how to estimate MTD at
  intended 30
• level
• - derived estimates conditional on doses
  yielded by design
• not well suited to yield any efficacy info (not
  suitable for estimating any rate of response
  other than 30
• yields little info above MTD

17
2. Random Walk Rules (RWR) or biased coin
designs

• Nonparametric model-based approach MTD is a
  quantile of a certain dose-response distribution
  but there is no underlying parametric family.
• Biased-coin design generalizes the up-and-down
  approach of the conventional method can target
  any response rate of interest (not only 30).
• Similarity patients are treated sequentially
  with the next higher, same, or next lower
  doses
• Difference rule for dose escalation (probability
  of next dose assignment depends on previous
  response)

18
2. Random Walk Rules (cont.)
19
Random Walk Rule Example

• An example of particular case of RWR with P1
• Developed in MRL in 1980s for efficacy
  dose-ranging studies
• called up-and-down design then
• Applied to simulated dose-ranging study in
  dental pain (full description in back-ups)
• Demonstrates 50 reduction in sample size
  without big loss in precision of estimates of
  dose-response compared to parallel group design

20
Pros () Cons (-) of Random Walk Rules

• Simple and intuitive to explain easy to
  implement
• flexible enough to target any level of response
• assigned doses cluster around quantile of
  interest ( MTD)
• - Consequently, some patients will be assigned
  above the MTD (concern for oncology only)
• minimizes observations at doses too small or
  too large, in comparison to randomized design
• - derived estimates conditional on doses yielded
  by design
• derived info useful to design definitive
  studies
• simulations indicate estimated response
  proportion at each dose is unbiased
• Have workable finite distribution theory
• Reliable MTD estimates can be obtained using
  isotonic regression
• - May not converge to MTD as fast as some
  Bayesian methods (wider spread of doses)
• But, for practical considerations (safety),
  slow dose escalation is desirable

21
3. Designs based on Bayesian methods

• Continual Reassessment Method (CRM)
• Escalation With Overdose Control (EWOC)
• Decision Theoretic Approaches
• Bayesian Optimal Sequential Design

22
Continual Reassessment Method (CRM)

• Most known Bayesian method for Phase I trials
• Underlying dose-response relationship is
  described by a 1-parameter function
• For a predefined set of doses to be studied and a
  binary response, estimates dose level (MTD) that
  yields a particular proportion (P) of responses
• CRM uses Bayes theorem with accruing data to
  update the distribution of MTD based on previous
  responses
• After each patients response, posterior
  distribution of model parameter is updated
  predicted probabilities of a toxic response at
  each dose level are updated
• The dose level for next patient is selected as
  the one with predicted probability closest to
  the target level of response
• Procedure stops after N patients enrolled
• Final estimate of MTD dose with posterior
  probability closest to P after N patients
• The method is designed to converge to MTD

23
Continual Reassessment Method (cont.)
Choose initial estimate of response
distribution choose initial dose
Update Dose Response Model estimate Prob.
(Resp.) _at_ each dose
Obtain next Patients Observation
Next Pt. Dose Dose w/ Prob. (Resp.) Closest
to Target level
Stop. EDxx Dose w/ Prob. (Resp.) Closest
to Target level
Max N Reached?
no
yes
24
CRM Design example (1)

• Post-anesthetic care patients received a single
  IV dose of 0.25, 0.50, 0.75, or 1.00 µg/kg
  nalmefene.
• Response was Reversal of Analgesia (ROA)
  increase in pain score of two or more integers
  above baseline on 0-10 NRS after nalmefene
• Patients entered sequentially, starting with the
  lowest dose
• The maximum tolerated dose dose, among the four
  studied, with a final mean posterior probability
  of ROA closest to 0.20 (i.e., a 20 chance of
  causing reversal)
• Modified continual reassessment method (iterative
  Bayesian proc) selected the dose for each
  successive pt. as that having a mean posterior
  probability of ROA closest to the preselected
  target 0.20.
• 1-parameter logistic function for probability of
  ROA used to fit the data at each stage
• Dougherty,et al. ANESTHESIOLOGY (2000)

25
CRM example (1) results

• including the 1st patient treated
• (MTD), i.e., estimated mean posterior
  probability closest to 0.20 target
• extrapolated

26
CRM example (1) results
Posterior ROA Probability (with 95 probability
intervals)
1.0
0.8
0.6
0.4
0.2
0.0
27
Escalation with overdose control (EWOC)

• Assumes more flexible model for the
  dose-response curve in terms of two parameters
• MTD
• probability of response at dose D1
• Similar to CRM in a way the distribution it
  updates posterior distribution of MTD based on
  this two-parameter model
• Introduces overdose control predicted
  probability of next assignment exceeding MTD is
  controlled (Bayesian feasible design)
• Assigns doses similarly to CRM, except for
  overdose control this distinction is
  particularly important in oncology
• EWOC is optimal in the class of the feasible
  designs

28
Decision-theoretic approaches

• Wide class of methods characterized by
  application of Bayesian Decision Theory to
  address various design goals
• shorter trials, reducing number of patients,
  maximizing information, reducing cost etc.
• Similar to CRM parametric model-based approach
  where the posterior distribution of model
  parameters is updated after addition of each new
  patient
• Uses gain functions depending on the desired
  goal
• Constructs a design maximizing the gain function
  using the set of action (pre-selected doses).

29
Decision-theoretic approaches (cont.)

• Whitehead and Brunier (1995) Loss function
  minimizes asymptotic variance of MTD
• Two-parameter model with for dose response with
  prior distributions on the parameters
• Posterior distribution estimates of the 2
  parameters used to derive next dose, i.e., that
  estimated to have desired response level
• Most versatile
• CRM and Bayesian D-optimal designs can be written
  as special cases
• Can be extended for simultaneous assessments of
  efficacy toxicity
• Patterson et al (1999) and (Whitehead et al
  (2001) extend this methodology in looking at
  pharmacokinetic data with two gain functions.

30
Bayesian D-Optimal Sequential design

• The methodology is similar to decision-theoretic
  approach, i.e. principally concerned with
  efficiency of estimation
• A two parameter model is used with logistic
  link function defining the dose response curve
• Based on formal theory of optimal design
  (Atkinson and Donev, 1992)
• Optimality criterion chosen to minimizes variance
  of posterior distribution of model parameters
• Similar to EWOC, a constraint is added to
  address the ethical dilemma of avoiding extremely
  toxic doses

31
Bayesian D-Optimal Sequential design (cont.)

• General methodology developed for the case when
  the dose space is unknown (continuous dose space)
• Case when doses are fixed in advance is
  particularly important in practice (discrete dose
  space)
• Sequential procedure developed consisting of
• Pilot design stage for seeding ( small group of
  subjects dose assignments based on prior
  information only)
• Subsequent assignments for each patient chosen in
  accordance with D-optimality criterion to
  maximize information from the design
• Posterior updated after each response and
  affects future dose assignments

32
Simultaneous assessment of efficacy and toxicity

• Penalized D-optimal designs
• (V. Dragalin and V. Fedorov, 2005)
• Non-Bayesian
• Accomplish learning by sequential updating of
  likelihood function afetr each patients response
• D-optimality criterion (maximizing Fishers
  information) is driving the design
• Optimization subject to constraint (reflecting
  ethical concerns, cost, sample size etc.)
• Flexibility of constraints and bivariate model
  allow to address a number of questions involving
  efficacy and safety dose-response curves
  simultaneously

33
5. Adaptive Frequentist approaches for late stage
exploratory development

• These designs more typically applicable to phase
  II studies
• Strongly control type I error rate
• 2 sources of multiplicity in adaptive
  dose-response trials
• Multiple comparisons of various doses vs. control
• Multiple interim looks at the data

34
5. Adaptive Frequentist approaches for late stage
exploratory development (cont.)

• Classical group-sequential design (Jennison
  Turnbull, 2000)
• Planned SS or information may be reduced if trial
  (or arm) stopped early
• At each interim looks test statistics compared to
  pre-determined boundaries
• Multi-arm trials Stallard Todd, 2003
• Adaptive design (Jennison Turnbull, 2005 Bauer
  Brannath, 2004)
• More flexible in adaptation-gtmore suitable for
  multi-stage framework
• Allowed adaptations may include ? sample size,
  modifying patient population, adapting doses

35
5. Adaptive Frequentist approaches for late stage
exploratory development (cont.)

• Further methods
• Use standard single-stage multiplicity
  adjustment some doses may be dropped (Bretz et.
  al, 2006)
• Combining phase II/III using a surrogate (Liu
  Pledger, 2005 Todd Stallard, 2005)

36
6. Developing a Bayesian adaptive dose design

• Key feature of all Bayesian methods updating
  information as it accrues posterior updates)
• Calculating predictive probabilities of future
  results
• Assessing increment in information about
  dose-response curve depending on next dose
  assignment
• Type I error is not the focus, but can be studied
  via simulations
• Downside computational complexity

37
6. Developing a Bayesian adaptive dose design
(cont.)

• Modeling is critical
• In general, no restriction on the model other
  than it must have parameters
• Prior distribution is put on model parameters
• can be non-informative
• or incorporate objective historical information
  appropriately
• simulations used to evaluate robustness w/respect
  to choice of prior
• As the data accrues, distribution of unknown
  parameters is updated (posterior)

38
6. Developing a Bayesian adaptive dose design
(cont.)

• Bayesian approaches are standard in Phase I
  cancer trials
• The methods reviewed earlier were presented in
  somewhat restrictive context
• binary response
• strong safety concerns (upper-end dose
  restriction)
• monotonic dose-response curve
• specific parametric family for model
• More general Bayesian designs are gaining
  popularity in phase II dose-ranging studies

39
6. Developing a Bayesian adaptive dose design
(cont.)

• Example ASTIN trial (Berry et. al 2001)
• Flexible model not restricting shape of D-R
  curvenon-monotone allowed
• Two-stages dose ranging (15 doses pbo) and
  confirmatory
• Incorporated futility analyses
• Long-term endpoints were handled via longitudinal
  model predicting patients long term endpoint
  using patients intermediate endpoint
  measurements
• Key advantages of BD vs. fixed (in general)
• finds the right dose more efficiently
• More doses can be considered
• If futility analysis is used -gt may save
  resources

40
7. Monitoring issues and processes in adaptive
dose-response trials

• All adaptive trials raise issues/concerns about
  credibility of the trial conclusions
• It is beneficial to have a separate body without
  other direct trial responsibilities to review
  interim results recommend adaptations
• Other precautions need to be taken
• limiting disclosure of specific numerical
  information and/or statistical methodology
• These recommendations are especially important in
  trials with potential regulatory submission (even
  if it is not confirmatory)

41
8. Rolling Dose Studies

• Broad class of design and methods that allow
  flexible, dynamic allocation of patients to dose
  level as the trial progress
• Not a distinct set of methods
• Rely more on modeling and estimation rather than
  hypothesis testing
• Examples include Bayesian, D-optimality, and many
  more
• Comprehensive simulation project by PhRMA RDS
  working group is under way
• Developing different RDS methods
• Evaluating and comparing to traditional fixed
  dose finding approaches

42
9. Conclusions

• Recommend routine assessment of appropriateness
  of AD in CDPs
• Opportunity to efficiently gain more information
  about D-R early in the development (POC) for
  maximum benefit
• More streamlined Phase III trial plan
• Reduction in timelines cost
• More information at the time of filing
• AD are not necessarily always better than
  traditional fixed dose
• There are many choices of ADs
• Extensive planning, simulations, etc.
• Added operational and scientific complexity
  should be justified
• Planning is extremely important
• Limited examples of AD are available in the
  literature

43
References

• Dragalin V. Adaptive designs terminology and
  classification. Drug Inf J. 2006 (submitted)
• Rosenberger WF, Haines LM. Competing designs for
  phase I clinical trials a review. Stat Med.
  2002212757-2770
• Durham SD, Flournoy N. Random walks for quantile
  estimation. In Gupta SS, Berger JO, ed.
  Statistical Decision Theory and Related Topics.
  New York Springer1994467476.
• OQuigley J, Pepe M, Fisher L. Continual
  reassessment method a practical design for phase
  1 clinical trials in cancer. Biometrics
  19904633 48
• Dougherty TB, Porche VH, Thall PF. Maximum
  tolerated dose of Nalmefene in patients receiving
  epidural fentanyl and dilute bupivacaine for
  postoperative analgesia. Anesthesiology
  200092(4)1010-1016.
• Babb J, Rogatko A, Zacks S. Cancer phase I
  clinical trials efficient dose escalation with
  overdose control. Stat Med. 1998171103-1120.
• Whitehead J, Brunier H. Bayesian decision
  procedures for dose determining experiments. Stat
  Med. 199514885-893.
• Patterson S, Jones B. Bioequivalence and
  Statistics in Clinical Pharmacology London
  Chapman Hall2005.
• Whitehead J, Zhou Y, Stevens J, Blakey G. An
  evaluation of a Bayesian method of dose
  escalation based on bivariate binary responses. J
  Biopharm Stat. 200414(4)969-983.

44
References (cont.)

• Haines LM, Perevozskaya I, Rosenberger WF.
  Bayesian optimal designs for phase I clinical
  trials. Biometrics 200359561-600.
• Dragalin V, Fedorov V. Adaptive designs for
  dose-finding based on efficacy-toxicity response.
  Journal of Statistical Planning and Inference
  20051361800-1823.
• Jennison C, Turnbull BW. Group Sequential Methods
  with Applications to Clinical Trials. London
  Chapman and Hall2000.
• Stallard N, Todd S. Sequential designs for phase
  III clinical trials incorporating treatment
  selection. Stat Med. 200322689-703.
• Jennison C, Turnbull BW. Meta-analyses and
  adaptive group sequential designs in the clinical
  development process. J Biopharm Stat.
  200515537-558.
• Bauer P, Brannath W. The advantages and
  disadvantages of adaptive designs for clinical
  trials. Drug Discovery Today 20049(8)351-357.
• Bretz F, Schmidli H, König F, Racine A, Maurer W.
  Confirmatory seamless phase II/III clinical
  trials with hypothesis selection at interim
  General concepts. Biom J. 2006 (in press).
• Liu Q, Pledger WG. Phase 2 and 3 combination
  designs to accelerate drug development. J Am Stat
  Assoc. 2005100493-502.
• Todd S, Stallard N. A new clinical trial design
  combining phases II and III sequential designs
  with treatment selection and a change of
  endpoint. Drug Inf J. 200539109-118.
• Berry DA, Müller P, Grieve AP, Smith M, Parke T,
  Blazek R, Mitchard N, Krams M. Adaptive Bayesian
  Designs for Dose-Ranging Drug Trials. In Gatsonis
  C, Carlin B, Carriquiry A ed. Case Studies in
  Bayesian Statistics V 99-181. New York
  Springer-Verlag2001.

45
Questions?
46
Backups
47
Back-up set 1 UpDown DesignDefinition

• Yields distribution of doses clustered around
  dose with 50 responders (ED50)
• 1st subject receives dose chosen based on prior
  information
• Subsequent subjects receive next lower dose if
  previous subject responded, next higher dose if
  no response
• Data Summaries
• proportion of responders at each dose
• continuous data via summary statistics by dose
• Inference based on conditional distribution of
  response given the doses yielded by the dosing
  scheme
• 5 MRL examples from 1980s

48
Up Down DesignSimulated from Past Trial Results

• Single-dose dental pain study (total 399
  patients)
• 51 placebo patients
• 75 Dose 1 patients
• 76 Dose 2 patients
• 74 Dose 3 patients
• 76 Dose 4 patients
• 47 ibuprofen patients
• Primary endpoint is Total Pain Relief (AUC)
  during 0-8 hours post dose (TOPAR8)
• UpDown design in sequential groups of 12
  patients sampled from study results sorted by AN
  within treatment.

49
Simulated UpDown Designfrom completed Dental
Pain Study

• Sequential groups of 12 patients (3 placebo, 6
  test drug, 3 ibuprofen)
• First group receives Dose 2
• Subsequent group receives next higher dose if
  previous group is non-response, next lower dose
  if response
• Response (both conditions satisfied)
• Mean test drug mean placebo 15 units TOPAR8
• Mean test drug mean ibuprofen gt 0
• Algorithm continues until all ibuprofen data
  exhausted
• originally planned precision for ibuprofen vs
  placebo
• (16 groups 191 total patients)

50
Dental Pain Study Complete Results
51
Simulated UpDown Resultsfrom Dental Pain Study
data (1st 8 Groups in sequence)
52
Simulated UpDown Resultsfrom Dental Pain Study
data (last 8 Groups in sequence)
53
Dental Pain Randomized Design vs UpDown Design
Results
54
Dental Pain Randomized Design vs UpDown Design
Results
Complete Trial N399, UpDown Design N191
55
Conclusions from Simulated UpDown Design in
Dental Pain

• UpDown design is viable for dose-ranging in
  Dental Pain
• Yields similar dose-response information as
  parallel group design
• Can use substantially fewer patients than
  parallel group design
• Logistics of implementation more complicated than
  usual parallel group design
• Can be accomplished in single center or small
  number of centers

56
Back-up slide set 2 D-optimal design implemented
in a user-friendly software iDose (Interactive
Doser)

• Web-based application is available to any
  workstation equipped with a web browser
  (Rosenberger et al., Drug Information Journal,
  2004)
• Nothing to install/maintain on the client side
• Integration with other software for patient
  information is easy
• Service-oriented architecture of web-based
  application
• Addition of high-value services is easy to
  deploy, update , and maintain
• Service can be offered by external providers
• Clinician access control must be implemented
• Low security requirements no actual patient
  information transferred

57
iDose Software (cont.)

• iDose supports long transactions
• Dose and toxicity of each patent reported over
  time
• Server keeps track of clients state while waiting
  for patients response
• Statistical part is implemented in Mathworks
  Matlab product
• Matlab Server product allows Matlab to run on a
  server as an external process accessed through
  Common Gateway Interface
• Intermediate stages are preserved as a file
• Clinicians use their keyword to retrieve the
  state where they left
• Any existing access control systems may be
  layered for additional security
• All parameters entered are checked for validity
• Dynamic, context-sensitive help provided for each
  parameter entered

58
Simulated Bayesian D-optimal design for ED50
(iDose website)

• Osteoarthritis efficacy good/excellent assumed
  underlying distribution
• Dose 15 30 60 90 120 180 240
• G/E 30 40 55 65 75 75 75
• Prior estimates ED25 between 15 and 30 mg
• ED50 between 30
  and 60 mg
• 6 patients in Stage 1 for seeding purposes
• Optimal Design 3 pts at 15 mg, 2 at 60 mg, 1 at
  90 mg
• 24 subsequent patients entered sequentially at
  doses yielding minimum variance of ED50 estimate
• Responses / non-response assigned to approximate
  targeted G/E distribution above

59
Simulated Bayesian Optimal Design for ED50 -
Results
60
Simulated Bayesian Optimal Design for ED50
Summary

• Osteoarthritis efficacy good/excellent assumed
  underlying distribution
• Dose 15 30 60 90
  120 180 240
• assumed G/E 30 40 55 65 75 75
  75
• Responses 4 - 2 1
  8 - -
• pts. 13 0 4 1
  12 0 0
• observed 31 - 50 100 67
  - -
• Bayesian estimated ED50 48.7mg using only 30
  patients!!!
• However, little info about other doses due to
  nature of D-optimal design for ED50

61
Graphic Summary of Results from iDose software