1. Dose Finding Designs Incorporating Patient Reported Outcomes
Shing M. Lee
Columbia University
Joint work with Bin Cheng and Xiaoqi Lu

2. Dose Finding Clinical Trials
Goal: To determine the maximum tolerated dose (MTD)
Outcome: Dose limiting toxicity (DLT) defined as a grade 3 or higher toxicity according to the NCI CTCAE
The NCI CTCAE is provided by clinicians
All methods to date only use clinician reported DLT outcomes

3. Patient Reported Outcomes
Increased interest in including patient perspective in clinical trials
Patients provide unique perspective
Literature suggests under-reporting of symptoms by clinicians
Lack of standardize instrument to collect patient reported outcomes
PRO-CTCAE

4. PRO-CTCAE
Recently, validated instrument for collecting patient reported symptoms
124 items covering 78 symptoms
Similar grading system to the NCI-CTCAE
Has been incorporated in dose finding trials as secondary outcomes

5. Advantages of Incorporating PRO-CTCAE
Incorporating patient perspective
May reveal unexpected symptoms
Doses selected can be tolerated by patients
Relevant for new anticancer treatments

6. Problem Formulation Let Yc be the clinician reported DLT
Let Yp be the patient reported DLT
Let d1,…, dk be the K doses of interest
Let θc be the DLT threshold based on clinicians and θp be the DLT threshold based on patients
Traditional MTD is defined as:
dc* = argmax (P(Yc = 1|d) ≤ θc)

7. Marginal Model Approach
Define MTD using patient data as:
dp* = argmax(P(Yp=1|d )≤ θp)
To incorporate both, we can define MTD as: d* = min(dc*, dp*)
To estimate this, we can fit two separate CRM*s
Estimate model parameters separately
Assign the next patient:
d [i+1] = argmax {Fc(d,beta1) ≤θc, Fp(d, beta2) ≤θp)}
d[n] -> d* almost surely as n -> ∞
* O’Quigley et al. Biometrics 1990

8. Joint Model Approach Redefine outcome as: Z = 0 if Yc = 0 and Yp = 0
Let θc be the threshold for P(Yc = 1) = P(Z = 2) and φ be the threshold for P(Yc = 1 or Yp = 1) = P(Z ≥ 1)
MTD is defined based on multiple constraints d* = min (dc*, dc,p*),
where dc,p*=argmax (P(Z ≥ 1 |d) ≤ φ)
We can apply the CRM-MC* which has been shown to be consistent
* Lee, et al. , Biostatistics 2011, Cheng and Lee JSPI 2015

9. Simulation Study Assume K=5, N=21 and 39, θc=0.25, θp=0.35, φ=0.50
Using empiric working models
Using two stage designs with MLE
We compare the performance of the marginal and the joint approach against the CRM

10. MTD = 3 and Coincides 0.05 0.25 0.45 0.55 P(Yc=1) 0.17 0.18 0.35 0.65
Percentage of Recommendation
Dose Level
0.05
0.25
0.45
0.55
P(Yc=1)
0.17
0.18
0.35
0.65
0.80
P(Yp=1)
0.20
0.50
0.90
P(Yc=1 or Yp=1)

11. MTD = 2 and Clinician 0.05 0.25 0.45 0.55 0.70 P(Yc=1) 0.10 0.15 0.35
Percentage of Recommendation
Dose Level
0.05
0.25
0.45
0.55
0.70
P(Yc=1)
0.10
0.15
0.35
0.60
0.75
P(Yp=1)
0.30
0.50
0.65
0.80
P(Yc=1 or Yp=1)

12. MTD = 3 and Patient 0.05 0.10 0.16 0.25 0.45 P(Yc=1) 0.20 0.35 0.65
Percentage of Recommendation
Dose Level
0.05
0.10
0.16
0.25
0.45
P(Yc=1)
0.20
0.35
0.65
0.80
P(Yp=1)
0.30
0.50
0.70
P(Yc=1 or Yp=1)

13. MTD = 5 and Coincides 0.01 0.02 0.05 0.10 0.25 P(Yc=1) 0.04 0.09 0.17
Percentage of Recommendation
Dose Level
0.01
0.02
0.05
0.10
0.25
P(Yc=1)
0.04
0.09
0.17
0.20
0.35
P(Yp=1)
0.50
P(Yc=1 or Yp=1)

14. MTD = 5 and Coincides (N=39)
Percentage of Recommendation
Dose Level
0.01
0.02
0.05
0.10
0.25
P(Yc=1)
0.04
0.09
0.17
0.20
0.35
P(Yp=1)
0.50
P(Yc=1 or Yp=1)

15. Conclusions
When the MTD coincides both marginal and joint models perform similar to the CRM, except when the MTD is at highest dose level
Both marginal and joint model achieve the goal of finding a dose that satisfies both constraints
The marginal model is more conservative than the joint model