East-Escalate User Workshop - Cytel

East-Escalate User Workshop®

JSM 2015, Seattle

Hrishikesh Kulkarni, Charles Liu

[email protected]

Aug, 2015

Hrishikesh Kulkarni, Charles Liu East®Escalate Workshop Aug, 2015 1 / 68

Outline

1 Phase 1 Dose Finding

2 References


Phase 1 Dose Finding

Sequence of K fixed doses: d1, d2, ...dK

Each dose i has toxicity probability pi

Monotonicity assumption

Maximum Sample Size, cohort size, starting dose

Method for sequentially assigning doses to cohorts

Find MTD: highest dose with toxicity less than target pT


Dose Assignment Rules

In Phase 1 trial designs, why not assign patients equally at each dose?


Ethical Requirement

”The major difficulty in phase I trial design and conduct is the ethicalrequirement that the number of patients in the trial who experiencetoxicity must be limited.”


Phase I Trial Challenges

Under-dosing (risk of absence of any anti-tumour activity).

Over-dosing (risk of severe toxicity).

Minimize sample size (unproven agent).

Maximize information (on toxicity and PK profile).

Paoletti et al. (2006). European Journal of Cancer, 42(10), 1362–8.


Bayesian

Prior: uninformative, or published literature / experts

Likelihood: observed data

Posterior: updated evidence


Rule-based vs Model-based

Rule-based / Algorithmic (eg., 3+3) vs Model-based (CRM, BLRM)

mTPI is a mixture of both

Why model-based?

more flexible (eg., different cohort sizes)Bayesian statistical inferencesBut, more complex (need for user-friendly software!)


3+3 (Prevalence)

Over 98% of published Phase 1 trials (1991-2006) use variations of 3+3

(Rotgako et al., 2007)


3+3 (Prevalence)

Over 96% of published Phase 1 trials (2007-2008) use variations of 3+3

(LeTourneau et al., 2009)


East ESCALATE

http://www.youtube.com/watch?v=6txAE3eGOk8

Two modes: (1) Simulation; (2) Interim Monitoring


3 + 3 H (implicit pT = 1/3)

http://www.mdanderson.org/education-and-research/departments-programs-and-labs/departments-and-divisions/division-of-quantitative-sciences/lectures-and-seminars/peter-f.-thall-presentation.pdf


3 + 3 L (more common, implicit pT = 1/6)


Simulation Parameters


Response Generation Info


Create Multiple Profiles


Simulation Control Info


Output Summary


Subject-wise Dose Allocation


Dose-wise Summary


MTD Selected


Limitations of 3+3

Alessandro Matano, Novartis, http://www.smi-online.co.uk/pharmaceuticals/archive/4-2013/conference/adaptive-designs


Confidence Interval?


Target Toxicity?

Common misconception target toxicity is fixed (eg., 17%, or 33%).

He et al. (2006) showed via simulation that the expected toxicity level at theMTD for the 3+3 is between 19-22%.


Workshop Exercise (Simulation)

Adapted from Thall & Lee (2003):

”Patients with renal cell carcinoma (RCC) that was progressive after previoustreatment with interferon were eligible. Treatment consisted of fixed doses of5-FU and interferon, plus one of six doses of gemcitabine (GEM): 100, 200,300, 400, 500, or 600 mg/m2. Toxicity was defined as grade 3 or 4 diarrhea,mucositis, or hematological toxicity. A total of 36 patients were treated incohorts of size 3, with the first cohort given 200 mg/m2 of GEM.”

Enter relevant inputs into East






Output Details

For each scenario...

1 ...which dose is selected most often (and how often) as the MTD by 3+3?

2 ...how often does 3+3 fail to select an MTD?

3 ...what are the median and mean sample sizes?


modified Toxicity Probability Interval (mTPI)

mTPI is rule-based like 3+3 but Bayesian like CRM and BLRM

Challenges for model-based methods: complexity (esp for non-statisticians);sensitivity to priors



“...almost all phase I oncology trials conducted at Merck in past 2 years havebeen based on the mTPI design”



Probability of toxicity at each dose modeled by independent Beta distributions

Set of decision intervals specified (like in BLRM)

Dosing decisions determined by ’normalized’ posterior probability in eachinterval at the current dose di :

Escalate to di+1 if di is ’underdosing’Stay at di if ’proper dosing’De-escalate to di−1 if di is ’overdosing’

Compute UPM for each interval. The one with largest UPM implies thedecision.


mTPI Priors

“[W]e believe that for phase I trials with small sample sizes...the dependenceintroduced by prior models will have a strong influence on the operatingcharacteristics...The independent prior models performs quite well comparedto existing approaches.”


Equivalence Intervals

The Equivalence Interval (EI) is defined as [pT − ε1, pT + ε2]

pT − ε1 is the lowest toxicity probability that the physician would becomfortable using to treat future patients without dose escalation

pT + ε2 is the highest toxicity probability that the physician would becomfortable using to treat future patients without dose de-escalation


Unit Probability Mass

UPM(interval) = Post Pr(interval) / length(interval)


Trial Monitoring Table (for Clinicians)

E=Escalate, S=Stay, D=De-escalate, DU=De-escalate & Unacceptable


mTPI Dose Exclusion / Stopping Rule


Regulatory Guidelines

FDA Guidance (Clinical Considerations for Therapeutic Cancer Vaccines)

EMEA / CHMP Guideline on Clinical Trials in Small Populations


Continual Reassessment Method (CRM)

Bayesian model-based method (O’Quigley et al. 1990)

Uses all available information from doses to guide assignment

Inputs to specify:

target toxicity pT (e.g., 0.3)one-parameter (θ) dose-toxicity curveprior distribution for θprior mean probabilities at each dose (“skeleton”)

Next recommended dose: posterior toxicity probability closest to target


Modified CRM


Modified CRM

Differences:

Can start at lowest doseAllow multiple patients per cohortRestrict escalation to one dose (do not allow skipping when escalating)

“The unmodified CRM...produces only modest increases in accuracy overthe modified CRM, but at the price of greater toxicity, and, mostimportant, clinical acceptability.”


CRM Simulation Parameters


CRM Dose Toxicity Curves & Priors

Logistic: p(di ) = expc+diθ

1+expc+diθ, c fixed

Hyperbolic Tangent: p(di ) =(

tan(di )+12

)θ

Power: p(di ) = (pi )θ


CRM Stopping Rules


CRM Interim Monitoring


Bayesian Logistic Regression Model (BLRM)

Two-parameter logistic: logit(p(di )) = logα + β log(di/d∗)

d∗ is the “reference dose”: logit(p(d∗)) = logα.

The approach is recommending a range of doses instead of just one dose.


Uncertainty in toxicity rate

CRM relies on point estimate (posterior mean) , ignores uncertainty andcould be misleading.BLRM uses entire posterior distribution at each dose.eg, Two beta distributions with same posterior mean but very differentvariances. Pr(Overdosing), Pr(p > 0.6) = 0.168 vs 0.002

X-axis - probability of observing DLT


Escalation With Overdose Control (EWOC)

Choose dose that maximizes targeted toxicity probability, given notoverdosing.


Bayes Risk

Choose dose that minimizes posterior expected loss.


Prior Specification (direct vs indirect)

Enter directly bivariate normal for log(α) and log(β) OR

Enter “best guess” for p(d1) and MTD


Prior Specification (indirect)

1 Assuming logits of toxicity are linear, calculate prior probabilities of toxicity(predicted median) at each dose level

2 Assign a “minimally informative unimodal” Beta distribution at each doselevel (Neuenschwander et al., 2008 Appendix A)

3 Generate n sets of logits from Beta distributions, to obtain n estimates oflog(α) and log(β) using least squares

4 Use sample means, variance, correlation for bivariate normal


Posterior Sampling Methods


BLRM Stopping Rules


BLRM Interim Monitoring


Interval Probabilities by Dose


BLRM Final Inference


BLRM Example

First in-Human study on advanced solid tumors.Primary endpoint - frequency of DLTDoses - 1, 2 , 4, 8, 15, 30, 40.Weak Prior Informationd* set to 10BVN prior for log(α) and log(β)

means = (-0.693, 0), sd = (2, 1), corr = 0


BLRM Example (Contd..)


Workshop Exercise (Interim Monitoring)

Adapted from Neuenschwander (2008):

”This study is an open-label, multicenter, non-comparative, dose-escalationcancer trial designed to characterize the safety, tolerability, andpharmacokinetic profile of a drug and to determine its MTD. The pre-defineddoses were 1, 2.5, 5, 10, 15, 20, 25, 30, 40, 50, 75, 100, 150, 200 and 250mg.”

“Currently in our trials the dose recommendation relies on maximizing theprobability of target toxicity while controlling the probability of excessive orunacceptable toxicity at 25 per cent.”

Enter relevant inputs into East


Simulation Parameters

”The best matching bivariate normal prior for the two-parameter logisticmodel had parameters µ1 = 2.15, µ2 = 0.52, σ1 = 0.84, σ2 = 0.8, ρ = 0.2”






Simulation Stage Plots in Library


Interim Monitoring

“It was decided....not to skip dose levels during dose escalation.”

“The first cohort of patients was treated at 1 mg. No DLTs were observed forthe first four cohorts of patients, and the clinical team then decided to skiptwo dose levels...At 25 mg, two DLTs were seen in two patients. The datafrom the fifth cohort led to discussion about the dose for the next cohort.”

Enter these data into East. On the basis of information from the InterimMonitoring Dashboard, what dose would you recommend next?


Final Inference

“Eventually, the trial was continued with a dose of 20 mg, a total of ninepatients were enrolled at that dose, and two DLTs occurred. Then the trialwas stopped and 20 mg was declared as the MTD.

Enter and click “Final Inference”


Summary

Almost all trials to date have used rule-based methods

Rule-based methods (eg., 3+3) are easy to implement and simple to explain

Model-based methods (eg., CRM, BLRM) are more flexible and efficient, butperformance may depend on prior information

mTPI may be a useful compromise

East ESCALATE provides two modes:

1 Simulations for comparing and evaluating designs2 Interim Monitoring for executing designs and analyzing data


Work in Progress

Combination Designs

Enhancements


Outline

1 Phase 1 Dose Finding

2 References


References

Neuenschwander, Branson, Gsponer. Critical aspects of the Bayesian approach to phase I cancer trials.Statistics in Medicine 2008;27:2420 - 2439

Ji, Wang. Modified Toxicity Probability Interval Design: A Safer and More Reliable Method Than the 33 Design for Practical Phase I Trials. Journal of Clinical Oncology, 2013

Thall, Lee. Practical model-based dose-finding in phase I clinical trials: Methods based on toxicity. Int JGynecol Cancer 2003, 13, 251 - 261

He, Liu, Binkowitz, Quan. A model-based approach in the estimation of the maximum tolerated dose inphase I cancer clinical trials. Statistics in Medicine 2006; 25:2027 - 2042


East-Escalate User Workshop - Cytel

Documents