an analysis of patient-reported outcomes in breast cancer ...an analysis of patient-reported...

An analysis of patient-reported outcomes in

breast cancer patients through

item-response theory pharmacometric modeling

Emilie Schindler1, Bei Wang2, Bert L. Lum2, Sandhya Girish2, Jin Y. Jin2,

Lena E. Friberg1, Mats O. Karlsson1

1 Department of Pharmaceutical Biosciences, Uppsala University, Sweden 2 Genentech Inc., South San Francisco, California, USA

Patient-reported outcomes (PROs)

• Reports of the status of health-related quality of life that come directly

from the patient

– Collected as questionnaires

– Inform on

• Physical status

• Disease-related symptoms

• Symptomatic adverse effects

• Therapeutic areas:

– Oncology

– Diabetes

– Respiratory

– CNS

– Auto-immune/auto-inflammatory

• Increasing interest from the regulatory agencies

2

Congratulations!

The tests were

negative, everything is

perfectly alright!

My whole life

is affected

I cannot

sleep

I cannot bend

over or exercise

I cannot eat and

drink whatever I

like

Example of PRO questionnaireFunctional Assessment of Cancer Therapy-Breast (FACT-B)

• 36-item questionnaire

• 5 aspects of well-being:

- Physical

- Social/Family

- Emotional

- Functional

- Breast-cancer related

• For each item: score 0 to 4

(ordered categorical data)

http://www.facit.org/

3

Analysis of multi-item PROsClassical statistical approach

Relies on the sum of scores

Easy to conduct

Individual items are not considered

Requires imputation of missing answers

Ignores correlations between items

Welslau et al. Cancer (2013)

Time to symptom worsening

(drop in 5 points)

4

Analysis of multi-item PROsItem response theory (IRT) approach

• IRT in a pharmacometric framework to describe composite scores

– Alzheimer’s disease1, Parkinson’s disease, multiple sclerosis,

schizophrenia, etc.

• Assumes that the score for each of the items depends on a latent

variable (not directly observable)

1Ueckert et al. Pharm Res (2014)

Physical well-being(Ψi)

I have a lack of energy

I have nausea

I am forced to spend time in bed

I have pain

f1(Ψi)

f2(Ψi) f3(Ψi)

f4(Ψi)

f5(Ψi)

f6(Ψi)

f7(Ψi)

I am bothered by side effects of treatment

I have trouble meeting the needs of my family I feel ill

Figure adapted from Ueckert et al. PAGE 21 (2012) Abstr 2318

5

Analysis of multi-item PROsIRT model structure

• Describes the probability of each score m for each item j as a function of:

─ Patient-specific well-being 𝜳𝒊

─ Item-specific parameters

─ 𝒂𝒋 : 1 per item

─ 𝒃𝒋𝒎: 4 per item

• Item characteristic curves

6

𝑷 𝒀𝒊𝒋 ≥ 𝒎 =𝟏

𝟏 + 𝒆−𝒂𝒋 𝜳𝒊−𝒃𝒋𝒎

180 parameters

for FACT-B

FACT-B dataEMILIA study

• Randomized, open-label, international phase III study

7

Adapted from Welslau et al. Cancer (2013)

HER2+ locally advanced or metastatic breast cancer

N=991

Prior taxane and trastuzumab

Progression on treatment for LABC/MBC, or within 6 months of

adjuvant treatment

Ado-trastuzumab emtansine (T-DM1)

(3.6 mg/kg IV q3w)

N=495

Capecitabine (1000 mg/m2 PO bid, days 1-14, q3w)

+Lapatinib (1250 mg/day PO qd)

N=496

1:1

FACT-B dataEMILIA study

• Randomized, open-label, international phase III study

8



N=991



adjuvant treatment


(3.6 mg/kg IV q3w)

N=495

2655 FACT-B, median 24 weeks



N=496

2192 FACT-B, median 18 weeks

1:1

1 2 3 4 5 6 7 8Treatment cycles i

FACT-B FACT-B FACT-B FACT-B FACT-B

Objectives

• To characterize FACT-B data in breast cancer patients following

treatment with T-DM1 using an IRT pharmacometric approach

• To investigate potential exposure-response relationships

• To compare the response of T-DM1 to the reference treatment

(capecitabine and lapatinib)

9

T-DM1 arm dataIRT model development

10


(N=991)



adjuvant treatment


(3.6 mg/kg IV q3w)

(N=495)



(N=496)

1:1

Model developmentA 3-step modeling approach

11

Patient 1 Patient i

Visit 1 Visit k

Patient 1 Patient i

Step 1: base IRT model– Item score data used as dependent variable

– Probability of a given score is a function of well-being Ψ and item-specific parameters

– Well-being Ψ predicted for each visit and patient

Step 2: longitudinal well-being model– Empirical Bayes estimates of Ψ from step 1 used

as dependent variable

– Effect of time, covariates and drug exposure on longitudinal well-being are investigated

Step 3: longitudinal IRT model– Longitudinal item score data used as dependent

variable

– Models from step 1 and 2 are combined

𝒑𝒓𝒐𝒃 𝒔𝒄𝒐𝒓𝒆 = 𝒇(𝚿, 𝐢𝐭𝐞𝐦 𝐩𝐫𝐦)

Visit 1 Visit k

Patient 1 Patient i

Visit 1 Visit k

𝒑𝒓𝒐𝒃 𝒔𝒄𝒐𝒓𝒆 = 𝒇(𝚿 𝐭, 𝒄𝒐𝒗𝒂𝒓𝒊𝒂𝒕𝒆𝒔 , 𝐢𝐭𝐞𝐦 𝐩𝐫𝐦)

Visit 1 Visit k

𝚿 𝐭 = 𝒇(𝒕, 𝒄𝒐𝒗𝒂𝒓𝒊𝒂𝒕𝒆𝒔)

Step 1: Base IRT modelDescribes the probability of each score for each item as a

function of a well-being variable Ψ

12

Latent variable

FACT-B item data

= dependent variable

Well being 𝚿

Score

Phys 1

Score

Phys 7…

Score

Soc 1

Score

Soc 7…

Score

Emo 1

Score

Emo 6…

Score

Func 1

Score

Func 7…

Score

Breast 1

Score

Breast 9…


function of five correlated well-being variables Ψ

13

𝚿𝐩𝐡𝐲𝐬𝐢𝐜𝐚𝐥 𝚿𝐬𝐨𝐜𝐢𝐚𝐥 𝚿𝐞𝐦𝐨𝐭𝐢𝐨𝐧𝐚𝐥 𝚿𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐚𝐥 𝚿𝐛𝐫𝐞𝐚𝐬𝐭 𝐜𝐚𝐧𝐜𝐞𝐫

Score

Phys 1

Score

Phys 7…

Score

Soc 1

Score

Soc 7…

Score

Emo 1

Score

Emo 6…

Score

Func 1

Score

Func 7…

Score

Breast 1

Score

Breast 9…

Latent variables

FACT-B item data


Step 1: Base IRT modelItem characteristic curves

14

So

cia

lP

hys

ica

lE

mo

tion

al

Fu

nctio

nal

Bre

as

t

Step 1: Base IRT modelItem characteristic curves

15

So

cia

lP

hys

ica

lE

mo

tion

al

Fu

nctio

nal

Bre

as

t

Step 1: Base IRT modelReassignment of breast cancer specific items

• Reassignment to other subscales

– Based on likelihood ratio tests

– Significantly improved the fit of the base IRT model (dOFV=-1138)

16

Breast cancer specific item Reassigned to

Short of breath Physical

Arms swollen or tender Physical

Self-conscious about the way I dress Emotional

Bothered by hair loss Emotional

Bothered by change of weight Emotional

Worry that family members get the same illness Emotional

Worry about effect of stress on illness Emotional

Feel like a woman Functional

Feel sexually attractive Functional


function of four correlated well-being variables Ψ

17

𝚿𝐩𝐡𝐲𝐬𝐢𝐜𝐚𝐥 𝚿𝐬𝐨𝐜𝐢𝐚𝐥 𝚿𝐞𝐦𝐨𝐭𝐢𝐨𝐧𝐚𝐥 𝚿𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐚𝐥

Score

Phys 1

Score

Phys 9

Score

Soc 1

Score

Soc 7

Score

Emo 1

Score

Emo 11

Score

Func 1

Score

Func 9… … … …

Latent variables

FACT-B item data


Step 2: Longitudinal well-being modelDescribes the time-course of well-being and investigates

the effect of covariates and drug exposure

18


Time Covariates Drug exposureExplanatory

variables

Latent variables


Step 2: Longitudinal well-being modelBase model: structure

• Best model: asymptotic function of time

Ψ 𝑡 = Ψ0 +Ψ𝑠𝑠 ⋅ 1 − 𝑒−ln 2𝑇1/2

⋅𝑡

19

Parameter Typical value

(RSE%)

Variability -

SD (RSE%)

𝚿𝐬𝐬, 𝐩𝐡𝐲𝐬𝐢𝐜𝐚𝐥0*

IIV:

0.57 (6)

ISV:

0.20 (20)

𝚿𝐬𝐬,𝐬𝐨𝐜𝐢𝐚𝐥-0.18 (19)

𝚿𝐬𝐬,𝐞𝐦𝐨𝐭𝐢𝐨𝐧𝐚𝐥0.32 (10)

𝚿𝐬𝐬,𝐟𝐮𝐧𝐜𝐭𝐢𝐨𝐧𝐚𝐥0*

T1/2 (days) 52 (21) 111% CV (11)

– Ψ0: baseline well-being

– Ψ𝑠𝑠: steady-state well-being

• Ψ𝑠𝑠 = 0: stable

• Ψ𝑠𝑠 > 0: improvement

• Ψ𝑠𝑠 < 0: worsening

– Additive inter-individual (IIV) and

inter-subscale (ISV) variability on Ψ𝑠𝑠

– 𝑇1/2 common to all subscales

* fixed

Step 2: Longitudinal well-being modelBase model: visual predictive checks

20

Step 2: Longitudinal well-being modelCovariate model: investigated relations

• Covariates tested additively on Ψ0 and Ψ𝑠𝑠

21

Demographics

• Age

• Region

• Ethnicity

• Race

Disease status

• Baseline tumor burden

• Baseline ECOG

• Site of disease involvement

• Hormone receptor status

• Liver, bone, lung, brain metastases

• Measurable disease

Prior therapies

& T-DM1 exposure

• Treatment line

• T-DM1 AUCcycle1

• T-DM1 Cmin,cycle1

ECOG: Eastern Cooperative Oncology Group functional status.

Step 2: Longitudinal well-being modelCovariate model: significant relations

• Race (Asian vs non-Asian) affects baseline social and functional

well-being

22

Demographics

• Age

• Region

• Ethnicity

• Race

• Ψ0,social*

• Ψ0,functional*

Disease status

• Baseline tumor burden

• Baseline ECOG

• Site of disease involvement

• Hormone receptor status

• Liver, bone, lung, brain metastases

• Measurable disease

Prior therapies

& T-DM1 exposure

• Treatment line

• T-DM1 AUCcycle1

• T-DM1 Cmin,cycle1

No exposure-response

ECOG: Eastern Cooperative Oncology Group functional status. * p<0.001

Step 3: Longitudinal IRT modelCombining models from step 1 and 2

23


TimeCovariates

(Race)Drug exposure

Explanatory variables

Latent variables

Score

Phys 1

Score

Phys 9

Score

Soc 1

Score

Soc 7

Score

Emo 1

Score

Emo 11

Score

Func 1

Score

Func 9… … … …

FACT-B item data


Step 3: Longitudinal IRT modelVisual predictive checks of the items’ average score

24

So

cia

lP

hys

ica

lE

mo

tion

al

Fu

nctio

nal

Step 3: Longitudinal IRT modelAsian patients typically have lower baseline

social and functional well-being

25

Social Functional

Reference arm dataIRT model applied and refined

26


(N=991)



adjuvant treatment


(3.6 mg/kg IV q3w)

(N=495)


+

Lapatinib (1250 mg/day PO qd)

(N=496)

1:1

Step 1: Base IRT model

• The base IRT model developed with T-DM1 data was

applied to the reference arm data

• Empirical Bayes estimates of Ψ were obtained and

used in Step 2

27

Step 2: longitudinal well-being modelUsing parameter estimates obtained with T-DM1 arm model

28

Step 2: longitudinal well-being model Re-estimating model parameters

29

Step 3: longitudinal IRT modelT-DM1 vs reference

30

Parameter T-DM1 arm Reference arm

Ψss,physical 0* -0.341

Ψss, social -0.0987 -0.115

Ψss,emotional 0.428 0.338

Ψss,functional 0* -0.0243

T1/2 (days) 77.4 28.7

Influence of Asian on Ψ0,social -0.536 -0.482

Influence of Asian on Ψ0,functional -0.195 -0.250

Ψ 𝑡 = Ψ0 +Ψ𝑠𝑠 ⋅ 1 − 𝑒−ln 2𝑇1/2

⋅𝑡

* fixed

Step 3: longitudinal IRT modelT-DM1 vs reference

31

So

cia

lP

hys

ica

lE

mo

tion

al

Fu

nctio

nal

ConclusionsEMILIA study

• The IRT pharmacometric framework well-characterized FACT-B

item-level data in both T-DM1 and capecitabine/lapatinib treated

patients

• No T-DM1 exposure-response relationships were identified for

FACT-B

• Differences between Asian and non-Asian may exist in baseline

social and functional well-being

• T-DM1 arm showed similar or better typical well-being than

reference arm for all physical, social, emotional and functional

aspects

32

ConclusionsMethodological achievements

• First application of IRT pharmacometric modeling to PRO data

• Handles missing values without imputation

• Accounts for correlations between items

• Describes the multi-dimensional nature of the questionnaire

• Provides insight into the questionnaire structure

• Can easily be extended to binary data and count data

33

Acknowledgements

• All T-DM1 team members at Genentech

• All colleagues at Uppsala University

34

Major reported adverse eventsEMILIA study

• Clinician-rated adverse events

36

Adverse events Capecitabine + Lapatinib T-DM1

Any grade Diarrhea (80%)

Hand-foot syndrome (58%)

Nausea (45%)

Vomiting (29%)

Fatigue (28%)

Nausea (39%)

Fatigue (35%)

Thrombocytopenia (28%)

Diarrhea (23%)

Elevated AST (22%)

Grade 3 or above Diarrhea (21%)

Hand-foot syndrome (16%)

Thrombocytopenia (13%)

Elevated AST (4%)


an analysis of patient-reported outcomes in breast cancer ...an analysis of patient-reported...

Documents