Presented at the ISPOR Europe 2018; November 10–14, 2018; Barcelona, SpainPreviously presented at the text

Copies of this poster obtained through Quick Response (QR) Code or text message are for personal use only and may not be reproduced without written permission of the authors

Email:

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 12 24 36 48 60

Surv

ival

pro

bab

ility

Time (months)

Kaplan-Meier Sunitinib CM214

Log-normal

Log-logistic

Generalized Gamma

Spline 1 knot normal

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 12 24 36 48 60

Surv

ival

pro

bab

ility

Time (months)

Kaplan-Meier Sunitinib CM214

Log-normal

Generalized Gamma

Spline 1 knot – hazard

Spline 1 knot – odds

Background

● The extrapolation of survival outcomes is key to cost-effectiveness (CE) analyses in oncology, especially

when progression-free (PFS) and overall survival (OS) outcomes have limited trial follow-up.

● Different parametric models can achieve a comparable fit to the randomised clinical trial (RCT) data but

generate different long-term predictions, due to the variability of the tails of survival distributions.

● Guidance from the NICE DSU by Latimer [1] suggests that the choice of distribution should be based on

internal and external validation, the latter usually being undertaken by visual inspection to inform model

choice.

● The CheckMate 214 trial, a phase 3, randomized, open-label study of nivolumab combined with ipilimumab

versus sunitinib in subjects with previously untreated, advanced or metastatic RCC [2], was used to

illustrate the Guyot’s approach to include real-world data into extrapolations of survival outcomes. The

population of interest focused on first line metastatic RCC patients with intermediate to poor prognosis

based on IMDC criteria.

Objective● The aim of this study was to evaluate the impact of the use of the method developed by Guyot et al. to

include external data in statistical extrapolation models, using survival outcomes from Checkmate 214

in first line advanced or metastatic RCC patients with intermediate to poor prognosis.

Methods

● Since mean survival times are very sensitive to assumptions around what occurs after the trial follow-up (in

the tail of the curves), there is a need to improve the extrapolation of survival data for use in CE analyses.

This poster presents a method developed by Guyot et al. using external information to extrapolate survival

curves based on data from the Checkmate 214 study.

● External information may be observational data, other RCT data, expert opinion, enabling data

reconstruction through published Kaplan-Meier curves. The idea is to model both the RCT data and

external data simultaneously with related parameters.

Inclusion of real world data in statistical extrapolation models

● Guyot et al (2016)[3] developed a method to adjust extrapolations of RCT data using external information

with longer follow-up periods. The method is implemented in a Bayesian framework. This method was

implemented based on data from the Checkmate 214 study, using the following steps:

1. Identification of external data: a systematic literature review on observational studies was conducted to

identify relevant data comparable with characteristics of CheckMate 214 patients. It is indeed key to

ensure that patient characteristics are as close as possible to the trial population, thereby ensuring that

prognostic variables are not imbalanced.

2. Definition of constraints: two main assumptions were made on the survival of the comparator arm of

ChecKMate 214, sunitinib.

• Conditional survival constraint:

The underlying assumption is that the conditional survival in the clinical trial control arm is likely to

converge to that of the one observed in observational studies over the long term. This constraint was

applied to OS and PFS. The constraint was defined as follows:

𝐶𝑆𝑒𝑥𝑡 𝑖 + 1 𝑖) = 𝐶𝑆𝑐𝑜𝑛𝑡𝑟𝑜𝑙 𝑖 + 1 𝑖)

with 𝐶𝑆 𝑖 + 1 𝑖) =𝑆(𝑖+1)

𝑆(𝑖)

With a binomial likelihood, as each time-point is conditionally independent.

• General population survival constraint:

An additional constraint was set in order to ensure that long-term projections do not exceed survival in

the general population, using general population mortality rates. The likelihood of the external data is

𝑟~𝐵𝑖𝑛(𝑆𝐺𝑃 𝑡 , 𝑛) with 𝑟 the number alive at time 𝑡 and 𝑛 the number at risk at time 0. The constraint iswritten:

𝑆𝐺𝑃 𝑡 = 𝑆0,𝑅𝐶𝑇 𝑡 + 𝛽, 𝛽 > 0 ⇒ 𝑆𝐺𝑃 𝑡 > 𝑆0,𝑅𝐶𝑇(𝑡)

3. Estimation of the survival parameters: 100,000 Bayesian Markov Chain Monte Carlo (MCMC)

simulations using Winbugs were used to estimate survival parameters, based on the approach

presented by Guyot.

4. The results obtained based on the inclusion of two real world studies were compared to extrapolations

obtained without the use of the Guyot method.

Identification of relevant real-world data

● Two studies with similar patient characteristics to the Checkmate 214 study were identified from the SLR of

real-world studies providing long-term estimates of survival: Verma et al. (2011) [4] and Kubackova et al

(2015) [5]. These studies provided long-term estimates of survival for a maximum follow-up period of 107

months. Only Kubackova 2015 reported PFS estimates.

Conclusions

● NICE recommends both internal and external validations for extrapolations of long-term survival

outcomes. Guyot’s method is a useful tool for extrapolations of time-to-event data, as it statistically

adjusts long-term predictions rather than relying on a simple visual inspection of curves. Therefore,

the Guyot’s method appears to be a less subjective method of external validation.

● Our example showed that the inclusion of external evidence using Guyot’s method had an impact on

extrapolations. The Verma study, providing survival estimates over a longer time period, had a larger

impact on the results.

● The conclusions highlight the need to justify and test the choice of external data to adjust

extrapolations.

● Limitations

– Due to the limited external evidence available in the literature to validate the outputs, the MSKCC

risk category was also considered in the identification of relevant external data, which might be a

source of heterogeneity for survival estimates.

– As independent models were used to model PFS and OS, the extrapolation of the

nivolumab+ipilimumab arm was not adjusted based on external data, as applying real-world

evidence adjustments based on estimates from another drug class was not deemed clinically

credible. Long-term outcomes for nivolumab+ipilimumab were not available.

References[1] Latimer, N., NICE DSU Technical support document 14: survival analysis for economic evaluations alongside clinical trials - extrapolation with

patient-level data. 2011

[2] CA209214 CheckMate 214, CHECKpoint pathway and nivolumab clinical Trial Evaluation

[3] Guyot, P., et al., Extrapolation of survival curves from cancer trials using external information. Medical Decision Making, 2017. 37(4): p. 353-366.

[4] Verma, J., et al., Impact of tyrosine kinase inhibitors on the incidence of brain metastasis in metastatic renal cell carcinoma. Cancer, 2011. 117(21):

p. 4958-4965.

[5] Kubackova, K., et al., Comparison of two prognostic models in patients with metastatic renal cancer treated with sunitinib: a retrospective, registry-

based study. Targeted oncology, 2015. 10(4): p. 557-563.

Acknowledgments• The patients and families who made this trial possible

• The clinical study teams who participated in the trial

• Bristol-Myers Squibb (Princeton, NJ) and ONO Pharmaceutical Company Ltd. (Osaka, Japan)

• The study was supported by Bristol-Myers Squibb

• All authors contributed to and approved the presentation; writing and editorial assistance was

provided by Amaris, funded by Bristol-Myers Squibb

Extrapolation of survival curves using external information: implementation of Guyot’s method in previously untreated advanced or

metastatic renal cell carcinomaCawston H1, Genestier V1, Dale P2, Doan J3, Malcolm B2

1Amaris, Paris, France; 2Bristol-Myers Squibb, Uxbridge, United-Kingdom; 3Bristol-Myers Squibb, Princeton, NJ, USA

PRM228

Figure 2. Independent PFS – Extrapolations with Guyot’s method

Figure 3. Independent OS – Extrapolations with Guyot’s method

Extrapolations with Guyot’s method

● Extrapolation of PFS (Figure 2)

– The 1 knot-hazard spline presented the lowest AIC and BIC scores. This model was thus selected for

the base case using Guyot’s method.

– For the sunitinib arm, the predicted PFS without external data adjustment was 25% at 2 years, while it

was 17% with external data. At 5 years, the unadjusted estimate was 9%, while it was reduced to 2%

including external data (Table 2).

Table 1. Studies of interest from the SLR

Figure 1. OS and PFS variations

Results

Parametric extrapolations

● The models with the best fit based on AIC and BIC criteria were selected for this analysis, chosen from a

selection of parametric and spline survival distributions (Figure 1). Since external data was only identified

for the comparator arm sunitinib, independent models for each arm were selected. Hence, the

extrapolations of the nivolumab+ipilimumab arm were not impacted by the adjustments. Besides, since

nivolumab + ipilimumab has just recently been available, there was no real-world evidence (RWE) with

longer follow-up data.

● As presented in Figure 1, the unadjusted models generated different long-term predictions for PFS and

OS after the end of the CheckMate 214 follow-up period, although the differences of AIC and BIC were

sometimes small.

External

study

Treatment

received

Risk

system

used

Median

age % of men Location

N

(intermediate

to poor)

Follow-up period median

(min-max)

Verma, 2011Sunitinib or sorafenib

MSKCC 60 67.0% USA 30216.2 (0 – 107)

months

Kubackova, 2015

Sunitinib MSKCC 64 67.9% Czech Republic 4354.8 (0.1 – 39.7)

months

Kubackova, 2015

Sunitinib IMDC 64 67.9% Czech Republic 3864.8 (0.1 – 39.7)

months

CheckMate214

Nivolumab + ipilimumab or

sunitinibIMDC 62 73.0% Multi-countries 200 24 months

● OS extrapolation results (Figure 3)

– The log-normal model minimised AIC and BIC scores across trial arms.

– The inclusion of the Verma study using Guyot’s method had a greater impact than the Kubackova

study, mainly due to the longer follow-up, while only 3 data points were used for the Kubackova study.

– For the sunitinib arm, the predicted OS without external data adjustment based on the log-normal model

was 53% at 2 years, while it was 48% with external data. At 5 years, the unadjusted estimate was 27%,

while it was reduced to 21% including external data (Table 2).

Survival at 1

year

Survival at 2

years

Survival at 5

years

Survival at 10

years

Survival at 20

years

PFS

Base case sunitinib 41.03% 25.24% 9.63% 3.02% 0.54%

Kubackova adjusted (IMDC) sunitinib (%

difference with base case)

39.59% (-3.51) 17.33% (-31.34) 1.90% (-80.27) 0.06% (-98.01) 0.00% (-100)

OS

Base case sunitinib 72.13% 52.54% 26.42% 12.43% 4.68%

Kubackova adjusted (IMDC) sunitinib (%

difference with base case)

71.78% (-0.49) 51.67% (-1.66) 25.55% (-3.29) 11.47% (-7.72) 4.13% (-11.75)

Verma adjusted (MSKCC) sunitinib (% difference

with base case)70.38% (-2.43) 48.10% (-8.45) 20.57% (-22.14) 8.01% (-35.56) 2.34% (-50)

Table 2. Predicted progression-free and overall survival with and without adjustment

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 12 24 36 48 60 72

Base case nivolumab+ipilimumab (1 knot-hazard independent spline)

Base case sunitinib (1 knot-hazard independent spline)

Kubackova adjusted (IMDC) sunitinib

Kubackova (IMDC) EXT

End of trial follow-up

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

0 12 24 36 48 60 72 84

Base case nivolumab+ipilimumab (log-normal independent)Base case sunitinib (log-normal independent)Kubackova adjusted (IMDC) sunitinibVerma adjusted (MSKCC) sunitinibKubackova (IMDC) EXTVerma (MSKCC) EXTEnd of trial follow-up

Overall survival Progression free survival

Email: helene.cawston@amaris.com

To request a copy of this poster:

Scan QR codevia a barcode

reader applicationBy requesting this content, you agree to receive a one-time communication

using automated technology. Msg. & data rates may apply. Links are valid

for 30 days after the congress presentation date.

Scientific Content On-demand