1. Phase I dose escalation studies in Oncology: a call for on-study safety and flexibility
Bill Mietlowski, Biometrics and Data Management, Novartis Oncology
KOL Adaptive Design seminar
July 8, 2011
KOL adaptive design presentation July 8, 2011

2. Outline of Presentation 2
Challenges of Phase I setting in Oncology
Design requirements
Proposed designs: algorithmic (e.g. 3+3) and continual reassessment method (CRM) vs. design requirements
Novartis Oncology standard: Bayesian logistic regression with escalation for overdose control to determine potentially unsafe doses
Protocols and dose escalation teleconferences to choose among the potentially safe doses
Conclusions
KOL adaptive design presentation July 8, 2011

3. Dose escalation setting in Oncology 3
Primary objective: Estimate maximum tolerable dose (MTD) based on acceptable rate of dose-limiting toxicities (DLT)
Assume true DLT rate at MTD is in (0.16, 0.33)
Generally small number of patients resistant/refractory to other therapies : often 15 to 30
Adaptive setting: dose escalations depend on DLT data
One dose (often MTD) usually selected for dose expansion
Large uncertainty during and at the end of the trial
KOL adaptive design presentation July 8, 2011

4. Challenges and Design Requirements for Oncology Phase I Trials 4
Phase I Trial Challenges
Design Requirements
Untested drug in resistant patients
Escalating dose cohorts (3-6 patients)
Primary objective: determine MTD
Accurately estimate MTD
High toxicity potential: safety first
Robustly avoid toxic doses (“overdosing”)
Most responses occur 80%-120% of MTD *
Avoid subtherapeutic doses while controlling overdosing
Find best dose for dose expansion
Enroll more patients at acceptable**, active doses (flexible cohort sizes)
Complete trial in timely fashion
Use available information efficiently
(Joffe and Miller 2006 JCO)
* Joffe and Miller 2006 JCO
** acceptable: less than or equal to the MTD determined on study
KOL adaptive design presentation July 8, 2011

5. MTD Targeting and Safety 5
Statisticians have taken great care to show operating characteristics of designs under different dose response shapes (steep, shallow, etc.)
Show likelihood of finding true MTD, underdosing, overdosing, etc.
However, published on-study safety characteristics very important to clinicians and regulators
Number of patients exposed to excessively toxic doses in actual trials a concern
Need to do extensive data scenario testing (performance of model under explicit occurrences, e.g. x DLTs in 3 patients at 1st cohort) as well as long-run simulations
KOL adaptive design presentation July 8, 2011

6. Heterogeneity in Cancer Trials 6
There is often substantial heterogeneity in cancer trials
Rogatko et al (2004) show patient characteristics can compete with dose with regard to adverse events.
There can be marked treatment x marker interaction in terms of efficacy (e.g. cetuximab and panitumumab in KRAS wild-type vs. KRAS mutated colorectal cancer) (Amado et al (2008))
Predictive biomarker may require early diagnostic development
KOL adaptive design presentation July 8, 2011

7. Impact of Dose Chosen for Expansion 7
Dose selected for dose expansion generally becomes the recommended phase II dose (RP2D)
If MTD underestimated, so is RP2D.
If MTD overestimated, RP2D may be overestimated and MTD must be re-estimated if toxicity issues emerge
May choose dose lower than cycle 1 MTD as RP2D based on available clinical data
Carefully choose the RP2D during dose escalation
May need to enrich at safe and active doses near MTD (flexible cohort sizes)
KOL adaptive design presentation July 8, 2011

8. Flexible cohort sizes may be useful when: 8
PK is erratic, dose proportionality is questionable
> linear or < linear
High potential for chronic (long term) toxicity
Need ample evaluable patients for later cycles at dose cohort
Enrich to understand degree of activity
More patients in Phase II population
More patients with tumor samples
If predictive biomarker is a concern (e.g. need n=8 patients in a cohort to have 90% likelihood of at least 1 marker + and at least 1 marker – patient if prob (marker +) =0.25)
Chronic (long-term)
KOL adaptive design presentation July 8, 2011

9. Efficient use of available information – prior 9
Prior DLT information from previous Phase I studies may be available for
New Phase I study for that agent
New Phase Ib combination trial
Prior information about DLTs from one schedule may be available for new schedule of the same agent
Proposed DE design should efficiently use available prior information
KOL adaptive design presentation July 8, 2011

10. Efficient use of available information – emerging 10
Sometimes, multiple schedules or both single agents and combos are studied in parallel (but perhaps staggered) in the same DE trial
Should exploit structural information if possible
DLTs on MWF schedule  Increased likelihood of DLT for daily dosing at the same dose
DLTs on single agent  Increased likelihood of DLT for combination at the same single agent dose
Proposed DE design should efficiently use this emerging information
KOL adaptive design presentation July 8, 2011

11. Approaches/Designs 11 Model-based designs have advantages over algorithmic designs
Two main approaches
Algorithmic: fixed “data-only rules”, e.g. “3+3”
Model-based: statistical  accounts for uncertainty of true DLT rates
Algorithmic
Model-based
Applicability
Easy
More complex due to statistical component ( training)
Flexibility
Not very flexible
fixed cohort size
fixed doses
Flexible: allows for
different cohort sizes
intermediate doses
Extendability
Rather difficult
Easily extendable
2 or more treatment arms
combinations
Inference for true DLT rates
Observed DLT rates only
Full inference, uncertainty assessed for true DLT rates
Statistical requirements
None
“reasonable” model, “good” statistics

12. level: Enroll 3 patients
Traditional 3+3 design
New cohort at a new dose
level: Enroll 3 patients
DLT =0/3
DLT =1/3
DLT >1/3
Go to next higher dose level
or same dose if highest dose
level
Enroll 3 additional pts
at the same dose level
Go to next lower dose level
or declare MTD at next lower
dose level if 6 pts already tested (never re-escalate)
DLT =1/6
DLT >1/6
Go to next higher untested
dose level or
declare MTD otherwise
Go to next lower dose level
or declare MTD at next lower
dose level if 6 pts already tested (never re-escalate)
KOL adaptive design presentation July 8, 2011

13. Published performance of 3+3 design 13
Low probability of selecting true MTD (e.g. Thall and Lee. 2003)
High variability in MTD estimates (Goodman et al. 1995)
Poor targeting of MTD on study:
Low MTD: Can assign toxic doses to relatively large number of patients (Rogatko et al. 2007)
High MTD: Tends to declare MTD at dose levels below the true MTD
Behavior depends on number of cohorts before MTD – too many leads to underdosing, too few leads to overdosing (Chen et al. 2009)
Alternative approach needed to meet Oncology study design requirements
KOL adaptive design presentation July 8, 2011

14. Case Report with Model Based Design 14
Are model-based designs too aggressive?
Example: Muler et al. (JCO 2004)
Continual Reassessment Method (CRM)
One-parameter model was used.
MTD recommendation from CRM: 50mg!
Indeed an aggressive recommendation.
Poor model fit and ignores uncertainty about DLT rate
Is it justified? No!

15. CRM analysis for Muler et al 15

16. Our standard dose escalation design 16
Bayesian logistic regression with escalation with overdose control (EWOC) (since 2004) (Neuenschwander et al SIM)
Three key intervals:
Underdosing → Pr (true DLT rate < 0.16)
Targeted toxicity → Pr (true DLT rate is in (0.16, 0.33))
Overdosing→ Pr (true DLT rate >0.33)
EWOC criteria mandates that posterior probability of overdosing <0.25.
KOL adaptive design presentation July 8, 2011

17. BLR-EWOC applied to Muler et al data 17

18. Priors 18 Typical priors represent different types of information
Bivariate normal prior for (log(),log())  prior for DLT rates p1,p2,…
Uninformative Prior
wide 95%-intervals
(default prior)
Historical Prior
Data from historical trials (discounted due to between-trial variation!)
Mixture Prior
Different prior information (pre-clinical variation)
different prior weights

19. Dose recommen- dations Model based dose-DLT relationship
Clinically driven, statistically supported decisions
Historical Data
(prior info)
Decisions
Dose Escalation
Decision
DLT rates
p1, p2,...,pMTD,...
(uncertainty!)
Dose recommen- dations
Trial Data
0/3,0/3,1/3,...
Clinical
Expertise
Model based dose-DLT relationship
Responsible: Statistician Responsible: Investigators/Clinician Informing: Clinician (Prior, DLT) Informing: Statistician (risk)
Model Inference Decision/Policy

20. Summary of statistical component 20
Model Prior Expertise  
Input
Inference
Recommendations
Substantial uncertainty in MTD finding requires statistical component
Input: standard model (logistic regression) + prior
Inference: probabilistic quantification of DLT rates, a requirement that leads to informed recommendations/decisions
Dose Recommendations are based on the probability of
targeted toxicity
and overdosing. Overdose criterion is essential.

21. Combination of clinical and statistical expertise 21 Practical and logistical aspects
Additional
study data
(e.g. AE, labs, EKG,
PK, BM, Imaging
Trial Data mg
Historical Data
(prior info)
DLT rates
p1, p2,...,pMTD,...
(uncertainty!)
Decisions
Dose Escalation
Decision
Dose recommen- dations
Model based dose-DLT relationship
Clinical
Expertise
Protocol development
Study conduct
Preparation for the dose escalation conference (DETC)
Discussion/decision at the dose escalation conference (DETC)
Incorporating prior information
Model Specification Review design performance
Pts enrollment
Observation during each dose cohort
KOL adaptive design presentation July 8, 2011

22. Protocol development (1) 22
Model Specification - Incorporating prior information
Preclinical toxicity data (with possible difference among species/gender),
STD10 and/or HNSTD translated to human doses and respective start doses
Shape of dose-toxicity relationship – variations as single-agent
Previous clinical trials
Literature data related to compounds, combination partners, etc.
Relevance of study population

23. Protocol development (2) 23
Design Specification
Pre-define provisional dose escalation steps
Provisional doses decided on expected escalation scheme - typically indicate maximum one-step jump. Intermediate doses may be used on data-driven basis
Minimum cohort-size – typically 3.
Allow enrollment of additional subjects for dropouts or cohort expansion
Pre-define DLT criteria and appropriate toxicity intervals
Pre-define evaluable patients for DLT assessment
All patients with DLT are included
For patients with no DLT, they must have sufficient drug exposure and completed required safety assessment to be sure of “no” DLT, or they are excluded

24. Protocol development (3) 24
Stopping rules (“rules for declaring the MTD”)
At least x patients at the MTD level with at least y patients evaluated in total in the dose escalation phase or
At least z patients evaluated at a dose level with a high precision (model recommends the same dose as the highest dose that is not an overdose with at least q% posterior probability in the target toxicity interval.)

25. Protocol development (4) 25
Statistician test-runs the design (if required)
Decisions under various data scenarios (scenario testing)
e.g. what happens if we see 0, 1 or 2 DLT in the first, second or third cohort?
or - what escalations can be made if we see no DLT in first 6 cohorts?
Operating characteristics (simulation testing)
Performance of the design in terms of correct dose-determination, gain in efficiency under various assumed dose-toxicity relationships (truths)
Clinicians review design performance document
Appended to protocol for HA/IRB review

26. Study conduct 26 Patient enrollment / observation for each dose cohort
To assure patient safety during the conduct of the study a close interaction within clinical team is required
Clinician, statistician, clinical pharmacologist, etc
Investigators
Clinical trial leader provides regular updates on accrual:
For each cohort enroll subjects per minimum cohort-size, typically 3
May enroll additional subjects up to a pre-specified maximum
In the case of unexpected or severe toxicity all investigators will be informed immediately
The model will be updated in case the first 2 patients in a cohort experience DLT

27. Dose escalation teleconference (DETC) 27
DETC scheduled close to all subjects in cohort being “evaluable”
Statistician is informed how many DLT and evaluable subjects are expected at the DETC
Statistician performs analysis with number of patients with/without DLT from all cohorts
Prior to DETC key safety data, labs, VS, ECG, PK, PD, anti- tumor activity, particularly from current cohort as well as previous cohorts are shared with investigators
Real time data for discussion – not necessarily audited

28. Dose escalation teleconference (DETC) 28
Discussion with investigators during the DETC
Investigators and sponsor review all available data (DLT, AE, labs, VS, ECG, PK, PD, efficacy) particularly from current cohort as well as previous cohorts
Agree on total number of DLTs and evaluable subjects for current cohort
Statistician informs participants of the highest dose level one may escalate to per statistical analysis and protocol restrictions

29. Dose escalation decision 29
Participants decide if synthesis of relevant clinical data justifies a dose escalation and to which dose (highest supported by the Bayesian analysis and protocol or intermediate)
Even though BLR-EWOC recommends dose escalation, team may enroll more at current dose to learn more from PK/PD, potential safety issues (later toxicities, lower grade toxicities, etc.)
Decisions are documented via minutes and communicated to all participants.

30. Summary 30 Patient safety is the primary objective
Statistical approach quantifies knowledge about DLT data only
Statistical inference is used as one component of a decision-making framework
Provides upper bound for potential doses based on uncertainty statements
To reduce risk of overdose  obtain more information at lower doses
Logistical application of our approach can be protocol/drug specific
Maximum escalation steps, minimum and maximum cohort sizes, stopping rules are pre-specified
Studies require active review of ongoing study data by Novartis and investigators
KOL adaptive design presentation July 8, 2011

31. Current state of Oncology Phase I trials 31
Rogatko et al (2007)
Investigated about 1200 Phase I Oncology trials
Only about 1.6% used innovative designs (most used 3+3)
In the past 3-4 years, the number has increased to 3-4%
This is disappointing. Reasons are:
Phase I has (for too long) been non-statistical
20 years of using the CRM has not changed this
Large scale implementation of innovative (Bayesian ) designs require a lot of effort
Guidance / support from key stakeholders is needed
Improper dose/regimen/patient population identified as a leading cause of failure of Phase III trials
KOL adaptive design presentation July 8, 2011

32. Acknowledgements 32 Many thanks to my Novartis Oncology BDM colleagues
Beat Neuenschwander
Stuart Bailey
Jyotirmoy Dey
Kannan Natarajan
KOL adaptive design presentation July 8, 2011

33. References
Amado, Wolf, Peeters, Van Cutsem et al (2008) Wild Type KRAS is required for panitumumab efficacy in patients with metastaic colorectal cancer Journal of Clinical Oncology, 26:
Babb, Rogatko, Zacks (1998). Cancer Phase I clinical trials: efficient dose escalation with overdose control . Statistics in Medicine, 17:
Bailey, Neuenschwander, Laird, Branson (2009). A Bayesian case study in oncology phase I combination dose-finding using logistic regression with covariates. Journal of Biopharmaceutical Statistics, 19:
Chen, Krailo, Sun, Azen (2009). Range and trend of the expected toxicity level (ETL) in standard A+B designs: A report from the children’s oncology group. Contemporary Clinical Trials, 30:
Goodman,Zahurak, Piantadosi (1995). Some practical improvements in the continual reassessment method for Phase I studies. Statistics in Medicine, 14:

34. References
Joffe, Miller (2006). Rethinking risk-benefit assessment for Phase I cancer trials. Journal of Clinical Oncology, 24:
Neuenschwander, Branson, Gsponer (2008) Critical aspects of the Bayesian approach to Phase I cancer trials. Statistics in Medicine, 27:
Rogatko, Babb, Wang, Slifker, Hudes (2004) Patient characteristics compete with dose as predictors of acute treatment toxicity in early phase clinical trials . Clinical Cancer Research 10:
Rogatko, Schroeneck, Jonas, Tighioart, Khuri, Porter (2007). Translation of innovative designs into Phase I trials. Journal of Clinical Oncology, 25:
Thall, Lee (2003) Practical model-based dose-finding in phase I clinical trials: methods based on toxicity. Int J Gynecol Cancer 13:
Thall, Millikan, Mueller, Lee (2003) Dose-finding with two agents in phase I oncology trials. Biometrics 59: