SLIDE 1 Operating Characteristics For Clinical Trial Design
Kenneth R. Hess, Ph.D. Department of Biostatistics
1
2019 Statistical Practice in Cancer Conference
SLIDE 2 Outline
Hess: 5 min: Introduction and overview 20 min: Dose finding phase I OCs and BOIN design Lee 5 min: Shiny Applications 20 min: Toxicity and efficacy monitoring for single arm phase II studies Both: 10 min: Q and A
2
SLIDE 3 Trial Design Operating Characteristics
OCs elucidate how a design performs under various scenarios For simple randomized trials, OCs = power curves For single arm studies, appropriate OCs depend on design Fundamental part of trial design (calibration) Should be planned and executed carefully Metrics should be chosen appropriately Generally based on computer simulations Important to “stress test” designs OCs should be included in protocol
3
SLIDE 4 Phase I Dose-finding in Oncology
Goals: Assess DLTs and Estimate MTD/RP2D Assume toxicity and efficacy increase with dose Start with lower dose and escalate sequentially Typically use cohorts of 3 patients Dose finding based on observed DLTs
4
SLIDE 5 3+3 Design (1950s)
Classical design, still used in > 90% of phase I studies Simple and transparent Treat patients in groups of 3 Escalate if 0 DLTs in 3 pts Expand with 3 more patients if 1 DLT in 3 pts Stop if > 1 pts with DLT in 3 or 6 pts Total N determined by # of pts with DLTs
5
SLIDE 6 Start at minimum dose level Treat 3 patients at this dose level # with DLT Treat 3 patients at this dose level # with DLT MTD exceeded MTD exceeded Go to next higher dose level Maximum dose level? MTD not found >1 of 3 1 of 3 >1 of 6 1 of 6 0 of 3 No Yes
3+3 Design
SLIDE 7
Newer Design Options
mCRM – modified continual reassessment method
– Fits probability model for DLT-dose curve (O’Quigley, 1990)
mTPI – modified toxicity probability interval
– Fits probability model to maximize chance of selecting dose with DLT rate in interval around target (Ji, 2010)
BOIN – similar to mTPI but better (Liu, 2015) mTPI2 – revision to mTPI to fix flaw (Guo, 2017) Keyboard – same as mTPI2 mCRM = model-based; others = model-assisted
SLIDE 8 Obstacles to Use of Newer Designs
Simplicity and transparency of 3+3 design Difficulty explaining model-based designs Difficulty implementing model-based designs Difficulty accepting simulation results as proof
8
SLIDE 9
Computer Simulation of Trials
DLT = binary event; binomial distribution Specify scenarios of true DLT rates E.g., dose level 1: 10%, 2: 25%, 3: 45% Generate random data for sequences of 3 pts Follow algorithm until MTD is found Repeat 10,000 times Generate metrics
SLIDE 10
6 Simulated Trials for DLT rate = 10%, 25%, 45%
Exp # Results MTD # Pts
1
0/3, 1/3+1/3 Level 1 9
2
0/3, 1/3+0/3, 2/3 Level 2 12
3
0/3, 3/3 Level 1 6
4
2/3 Exceeded 3
5
1/3+0/3, 0/3, 2/3 Level 2 12
6
0/3, 0/3, 1/3 + 2/3 Level 2 12
SLIDE 11
Results from 1,000 3+3 Simulated Trials
MTD below dose level 1: 9% MTD not reached: 0.1% Dose Level 1 2 3 4 5 DLT Probability 10% 20% 30% 40% 55% Selection Probability 27% 35% 20% 7% 1% Mean # Pts 4.4 4.6 3.3 1.6 0.4
SLIDE 12 Phase I Study OCs
Use variety of scenarios with MTD at different dose levels Include scenarios where MTD between dose levels Include scenarios where MTD well below first dose level Include scenarios where MTD well beyond last dose level When possible also include randomly generated scenarios
12
SLIDE 13 Model-assisted Designs
Easier to understand that model-based designs More transparent that model-based designs Easier to generate trial designs in practice Easier to implement when running trials Easy to use, online Shiny applications available
13
SLIDE 14
Bayesian Optimal Interval (BOIN) Design
New model-assisted design Described in Yuan Y, et. al. 2016 CCR 22:4291 Any DLT rate target can be specified Maximum N is specified Easy to implement Table to guide dose-escalation
SLIDE 15
BOIN creates three distinct probability regions and looks to see into which region the observed data fall.
λ𝑓 λd
SLIDE 16 BOIN Design Parameters
Target DLT rate, φ λ𝑓 and λ𝑒 are predetermined boundaries based on φ Maximum number of patients overall Maximum number of patients at any dose level Cohort size Dose elimination threshold; default = 0.95 Ref: S Liu, Y Yuan JRSSC 2015 64:507
16
SLIDE 17 Start at the lowest dose Treat a patient or a cohort of patients Reach the maximum sample size? Compute DLT rate at current dose Retain the current dose Stop the trial and Select the MTD Escalate the dose De-escalate dose Yes No ≥ λ𝑒 ≤ λ𝑓
Within (λ𝑓, λ𝑒)
BOIN Design
SLIDE 18 BOIN Dose Escalation Rules
Number of Patients Number of DLTs 3 6 9 12 15 18 21 24 27 30 E E E E E E E E E E 1 S E E E E E E E E E 2 D S E E E E E E E E 3 DU D S S E E E E E E 4 DU D S S E E E E E 5 DU DU D S S E E E E 6 DU DU D D S S S E E 7 DU DU D D S S S E 8 DU DU DU D D S S S 9 DU DU DU DU D D S S 10 DU DU DU DU D D S 11 DU DU DU DU DU D D 12 DU DU DU DU DU DU D 13 DU DU DU DU DU D 14 DU DU DU DU DU DU 15 DU DU DU DU DU DU E: Escalate to the next higher dose; S: Stay at the same dose; D: De-escalate to the previous lower dose; DU: De-escalate to the previous lower dose and the current dose will never be used again in the trial
BOIN Table of Decision Rules for Changing Dose Levels Based on Observed Toxicity Data
SLIDE 19 Design Performance Comparison
Metric 3+3 BOIN mCRM
33% 49% 49% Pts treated at MTD (%) 26% 31% 34% % selected dose <16% DLT 40% 25% 20% % selected dose >33% DLT 8% 12% 15% Risk of treating < 6 at MTD 43% 28% 28% From: H Zhou et al, CCR 2018 (target DLT rate = 25%)
SLIDE 20 Conclusions
OCs are essential in phase I study design 3+3 has obvious statistical short-comings Need to communicate these problems better Need to explain simulations better BOIN design: better than 3+3; easier than CRM Free software for BOIN – includes protocol template
20
SLIDE 21 Thank you for your attention!
21
