Two New Developments in East

Transcription

1 Two New Developments in East 1. PREDICT: Predicting the Future Course of a Trial 2. MAMS: Multi-arm Multi-stage Trials Cyrus Mehta East Workshop JSM 2015

2 PART 1: PREDICT 1. PREDICT is primarily a tool for use during interim monitoring although it can also be used at design Gme if preliminary data are available 2. It predicts the future course of the trial by uglizing current data as the basis for simulagng future trials 3. It consists of two tools: PIPs and EnEv 1. PIPs creates an intuigve graphical display of possible future outcomes and helps DMCs to make decisions regarding early terminagon for fuglity or efficacy 2. EnEv uglizes site- level data on enrollment and pagent level data on outcomes to produce accurate forecasts of future enrollment and event rates 2

3 PIPs: Predictive Interval Plots PIPs are a series of repeated confidence intervals generated by simulagng the future course of the trial condigonal on the current data The intervals are sorted and stacked to provide an intuigve graphical display that facilitates early terminagon decisions PIPs focus on es#ma#on rather than hypothesis tes#ng, thereby introducing clinical relevance into the decision making process References: (1) Li L, Evans SR, Uno H, Wei LJ. Sta$s$cs in Biopharmaceu$cal Research, 2009, Vol 1, No 4. (2) Evans SR, Li L, Wei LJ. DIA Journal, 2007, vol 41. 3

4 EnEv: Enrollment and Event Forecasting EnEv provides a unified framework for simulagng pagent and event arrivals over Gme PaGent arrivals are modelled at the site level as independent Poisson processes Event arrivals are modelled at the pagent level under the assumpgon of exponengal survival, taking inputs from the pagent arrivals model As new data arrive, the Poisson and exponengal survival model parameters are updated by Bayesian methods References: (1) Anisimov V, Fedorov V. Sta$s$cs in Medicine, 2007, Vol 26. (2) Torgovitsky R. Cytel Internal Technical Report for Serono Reflex Study,

5 Outline of this Presentation MoGvaGon for PREDICT differengagon from other tools in East Two examples of use of PIPs Example 1: early terminagon for fuglity Example 2: early terminagon for efficacy One Example of use of EnEv Use of inigal enrollment plan at design stage Use of updated enrollment plan and pagent- level data at interim monitoring stage 5

6 1. Use of PIPs for Futility Stopping Design of a Non Small Cell Lung Cancer Clinical Trial Design Parameter Value Primary Endpoint OS α 0.05 Power 80% One Interim Analysis O Brien- Fleming spending fn Median (Control Arm) 10 months Median (Experimental Arm) 12.5 months Hazard RaGo 0.8 Sample Size 888 Events 633 Plan to enroll for 18 months and follow for 12 more months 6

7 Design Summary This design has no fu#lity boundary 7

8 Interim Analysis after 258 Events Interim analysis occurred ahead of schedule due to rapid enrollment 8

9 Should we stop for futility? 9

10 PIP at 258 events under HR=1.01 Only 1 out of 1000 simula#ons had their upper confidence bound below HR=1 10

11 PIP at 258 Events under HR=0.8 Even if the true HR=0.8, the chance of success is only 29%. Moreover only 5% of future simulagons show a point esgmate for HR that is smaller than The trial was terminated for fu#lity 11

12 Use of PIPs for Efficacy Stopping ACTG A320: AZT+Epivir vs AZT+Epivir+Crixivan Design Parameter Value Primary Endpoint Time to AIDS or Death a 0.05 Power 93% 1 interim look at 25% informagon Haybigle- Peto; p=0.001 Event Free Survival (Control ) 67% at 18 months Event Free Survival (Experimental) 75.25% at 18 months Hazard RaGo 0.71 Sample Size 1750 Events 410 Plan to enroll for 36 weeks and follow for 48 more weeks 12

13 Design Summary 13

14 Interim Analysis after 96 Events At this interim look the p- value was Thus the HP boundary was barely crossed 14

15 Kaplan-Meier Plot of Data Can the trial be terminated with these early results based on just 25% of the informa#on? 15

16 PIP with the Observed Data (HR= ) Although the observed upper bound was barely below HR=1, the future intervals are much narrower and 100% of their upper bounds are below HR=1 16

17 PIP with Design Specified HR=0.71 Only 3/1000 future simulagons had upper bounds exceeding HR=1 Also 95% of point esgmates were less than 0.76; sgll clinically meaningful 17

18 Sensitivity Analysis: PIP with HR=0.75 Even if we assume HR=0.75 (on the borderline of clinical significance), only 112/1000 future intervals fail to be below HR=1. Thus DMC recommended trial termina#on. 18

19 3. Enrollment and Event Forecasting Design Parameters for the OncoX Clinical Trial Design Parameter Value Primary Endpoint OS α (one sided) Power 90% One Interim Analysis γ(- 5) efficacy and fuglity boundaries Median (Control Arm) 5 months Median (Experimental Arm) 7 months Hazard RaGo 0.71 Sample Size 460 Events 374 Plan to enroll for 24 months and follow for 6 more months 19

20 Design Details This design will enroll 460 pa#ents and wait for the arrival of 374 events 20

21 The e-chart (ignores site level data) Assumes constant enrollment for 24 months and 6 more months for follow- up 21

22 Initial Enrollment Plan country, sites, earliest start, latest end, pa#ent/site/month, cap 22

23 Enter the Enrollment Plan in East 23

24 Enrollment and Event Output Predic#ng: enrollment dura#on (17 to 19.5 mths); study dura#on (26 to 30 mths) 24

25 Data at the Interim Analysis (187 events) InterimSubjectData.cydx Hazard Ra#o Es#mate Enter into Interim Monitoring Worksheet 25

26 Updated Enrollment Plan InterimEnrollmentPlan.cydx 26

27 Conditional Simulation Entries 27

28 Enrollment and Event Output At month 21, future simulated trials place study duragon between 33 and 36 months 28

29 Concluding Remarks First agempt to use the actual pagent- level data at interim monitoring stage for forecasgng the future PIPs future chance of a successful outcome EnEv future enrollment and event forecasts Future DirecGons: More general models for pagent arrival, and survival IntegraGon with AdapGve SSR in East DistribuGon of p- values and other stagsgcs in PIPs More addigons based on user inputs 29

30 PART 2: Multi-arm Multi-stage Design GeneralizaGon of group sequengal design to more than two arms An alternagve to the combinagon funcgon approach of Posch et. al. (SiM, 2005) Current implementagon: trial stops if any arm crosses efficacy boundary trial stops if all arms cross fuglity boundary boundary is not re- computed if an arm is dropped Under development: adapgve SSR and adapgve drop the loser

31 Mathematical Framework K- look GSD Only 1 comparison, made K Gmes K- look MAMS: K K M comparisons, made K Gmes GeneralizaGon of Dunneg s test i 1 i 1 [ ] P ( I W < e and W e ) = α 0 j j i i i= 1 j= 1 [ ] P ( I max{w...w } < e and max{w...w } e ) = α 0 j1 jd j i1 id i i= 1 j= 1

32 Example: Chronic Obstructive Pulmonary Disease Once daily bronchodilators for COPD (Am. J. Respiratory & CriGcal Care, vol 182, 2010) Compare three doses (150 mg, 300 mg, 500 mg) of Indacaterol to Placebo Endpoint: Week 12 change from baseline in 24 hour trough FEV1 Expect differences from placebo of between 0.14 and 0.18 liters with standard deviagon σ=0.5 Design for 90% power at one- sided α=0.025

33 Two-arm, Three-look GSD Requires 165 pagents/arm for δ=0.18, σ=0.5 Expected sample size under H1 is 132/arm

34 Four-arm, Three-look MAMS Requires 130 pagents/arm for δ=0.18, σ=0.5 Expected sample size under H1 is 105/arm

35 Compare the 2 arm and 4-arm boundaries 2- arm boundaries on Z- scale 4- arm boundaries on Z- scale

36 Higher hurdle with 4-arm Trial There are more chances for a false claim with 3 comparisons instead of 1 Therefore the 4- arm trial has to have stricter boundaries

37 Boundary and Sample Size Comparison The boundaries for the 4- arm trial are higher than for the 2 arm trial This compensates for the greater probability of boundary crossing under H0 But the sample size/arm is lower for 4- arm trial. Why? What would happen if the value of δ was not the same for each treatment

38 4-arm design with different δ values Same boundaries, but requires commitment of 169/arm Expected sample size under H1 is 136/arm In this configuragon, 4- arm design requires more pagents/arm than 2- arm design The higher efficacy boundary hurdle is not offset by extra opportuniges to cross the efficacy boundary because only one dose has a strong effect

39 Simulate under equal δ values Not surprisingly, the marginal probabiliges are similar for each treatment

40 More output for equal δ case 44% chance of gesng at least two doses into the label

41 Simulate under unequal δ values 65% chance that dose 1 will be shown to be efficacious

42 More Output for Unequal δ case 40% chance of gesng at least 2 doses into label

43 Future Development of MAMS AdapGve MAMS Sample size re- esgmagon Alter the number of future stages Alter the spending funcgon Parameter EsGmaGon P- values Point esgmates Confidence Intervals 43