International Biometric Society
SIMULATION METHODS FOR POWER ANALYSIS TO PREDICT TAMOXIFEN TREATMENT RESPONSE
FOR BREAST CANCER
USING DIFFUSE OPTICAL SPECTROSCOPIC IMAGING (DOSI)
Christine E. McLaren1, Wen-Pin Chen2, Thomas D. O’Sullivan3, Bruce J. Tromberg3
1Department
of Epidemiology, University of California, Irvine, CA
Family Comprehensive Cancer Center, University of California, Irvine Medical
Center, Orange, CA
3Laser Microbeam and Medical Program, Beckman Laser Institute and Medical
Clinic, University of California, Irvine, CA
2Chao
Introduction. Women with high percent mammographic density have 4-6 fold increased
cancer risk compared to women with lower density. There is also significant evidence that
tamoxifen treatment is more effective at reducing breast cancer risk when accompanied by a
reduction of breast density. Magnetic resonance imaging (MRI) is a safe and quantitative
technique for measuring breast density and volume, but its high cost precludes MRI from
widespread, frequent use in risk assessment, screening, or therapeutic monitoring. Diffuse
optical spectroscopic imaging (DOSI) is a promising alternative that may be useful in
monitoring changes in breast density. We describe a novel strategy for power and sample
size determination motivated by our design of a study to compare the reduction from
baseline in DOSI measures that may reflect changes in breast density in premenopausal
women receiving tamoxifen. Because this will be the first study of DOSI in tamoxifen
treatment, no preliminary information was available to provide critical estimates of expected
changes from baseline in DOSI measures for treated and control groups. Thus, we
developed a statistical simulation approach utilizing information from corresponding MRI
breast density and DOSI measures obtained from our study of 12 volunteers about to begin
neoadjuvant chemotherapy (O’Sullivan, et al., Breast Cancer Res 2013) and a separate
investigation of MRI assessment of breast density in 16 women before and after treatment
with tamoxifen (Chen, et al., Magn Reson Imaging 2011).
Methods. Step 1: For each of three independent DOSI measures, water, deoxyhemoglobin
(ctHHb), and lipid, separate linear regression models were formed with the DOSI measure at
baseline as the outcome variable and MRI density
at baseline as the predictor. There was a strong
linear correlation between the three DOSI
measures and MRI breast density prior to therapy
(water, r=0.843, p<0.001; ctHHB, r=0.785,
p=0.003; lipid, r=-0.707, p=0.010). Step 2: Based
on the study of Chen and colleagues (2011), ten
thousand pairs of simulated MRI breast density
measurements were generated from a bivariate
normal distribution with mean percent density
(standard deviation) of 22.1% (2.6%) before
tamoxifen treatment and 16.3% (3.3%) after
treatment, with a correlation coefficient between
pre- and post-treatment values of 0.9 (Fig. 1). Step
3: The regression models obtained in Step 1 were
applied to the simulated pre- and post-treatment
MRI values to obtain predicted mean DOSI values at baseline and after treatment and
residual variance estimates. These parameter estimates were then used in a second
simulation of a bivariate normal distribution with correlation coefficient of 0.9, to obtain
10,000 corresponding pairs of pre- and post-therapy DOSI values representing water, ctHHb,
and lipid. Pre- and post-treatment differences between values for each DOSI measure were
International Biometric Conference, Florence, ITALY, 6 – 11 July 2014
International Biometric Society
calculated and used to inform power and sample size calculations for a two-sample t-test of
the mean reduction from baseline in the tamoxifen-treated vs. control group with 80% power
and significance level 0.05. Based on results of a study of mammographic density, we
assumed that the reduction in DOSI measures in the control group would be half that of the
tamoxifen-treated group (Cusick, et al, J Natl Cancer Inst 2011).
Results. Expected mean changes in the treated group were 2.5%, 0.341µM, and -2.614%
for water, ctHHB, and lipid respectively, compared to 1.25%, 0.171µM, and -1.307%. To
demonstrate a mean reduction in the control group of half that of the treated group with
common standard deviations of 1.733%, 0.273µM, and 1.330% respectively, required sample
sizes per group are 32 subjects for water, 42 for ctHHb, and 18 for lipid. Using the largest
required number of subjects and assuming as much as a 10% dropout we recommend a
target sample size of 47 subjects per group. We conclude that simulation techniques can
inform power and sample size calculations when important preliminary information is not
available.
International Biometric Conference, Florence, ITALY, 6 – 11 July 2014