TITE-CRM
The traditional design for
dose-escalation trials has poor statistical properties, and the obvious
fix--increasing the number of patients per dosing cohort--results in a
trial that takes too long to execute, especially when testing
chemoradiotherapies, which may have toxicities that take six months or
a year to appear. There are a number of alternative designs
for
Phase I trials, but TITE-CRM (Cheung & Chappell,
2000) is the
only one I know of that allows for continual enrollment of
subjects. The Radiation Oncology department has initiated
more
TITE-CRM trials than any other group in the world, and I have extended
the method to incorporate different parametric dose-toxicity models and
a more flexible nonparametric framework.
Pharmacoketic analysis using PROC NLMIXED
Statisticians are sometimes called upon to analyze
pharmacokinetic analyses, without access to the tool NONMEM, which is
widely used in the pharmacokinetics community. While SAS PROC
NLMIXED appears immediately appropriate for such analyses, I found that
there were a number of issues in even simple analyses that trip up the
statisitican who knows significantly less than he thinks he does
(nobody we know, of course).
Here is a technical report on UMCC 9941, a 12-patient PK study
on curcumin (a component of turmeric that has chemopreventitive
potential), and the SAS code that does the basic single compartment, first-order absorption and elimination PK analysis (but not
all the graphics, which made the program too large and intimidating).
It includes model-based calculations of AUC, Cmax, tmax and t1/2. If you are doing a PK analysis using NLMIXED for the first time, you may find some of this useful.
Bayesian
models for assay calibration
When several calibrated
assays are combined to form a scientifically or clinically meaningful
ratio, the derivation of
confidence and/or prediction intervals is messy, especially if the
calibration curves are nonlinear. Simultaneous estimation of
ratios of parameters and their prediction intervals can be automated
using a Bayesian model and Markov Chain Monte Carlo
estimation.
We will present a framework and an example of estimates of ratios of
polyamines from the Barrett's esophagus trial. This is still work
in progress, and there is no slide show yet. There is, however, a
slide show that describes the genesis of this work, using robust (M-)
estimators in a four-parameter logistic calibration model. This
material was also covered in the Statistics in Medicine paper.
Phase
I and Phase II trials
with continual improvement
Laboratory science moves faster
than clinical science.
Definitively proving that Treatment A is better than
Treatment B
requires a Phase III trial that takes several years, multiple clinical
centers and the associated bureaucracy, and by the time you're done,
Treatment F is all the vogue. How do you legitimately test
treatment innovations in a single center in fast lab time, rather than
glacial clinic time? The answer
is: by employing surrogate endpoints, but the devil is in the
details. We establish a framework for continual trials of
innovative chemoradiotherapies using surrogate endpoints.
Proteomics
to predict cancer
SELDI and MALDI are examples of high-throughput mass spectroscopy
technologies that show great promise for screening for cancer in
clinical or even naive populations, but there are significant technical
challenges, and other groups have made some ugly
missteps.
The Great
Lakes-New England Early Detection Research Network Clinical and
Epidemiology Center is conducting a study to see, if we do everything
right, can we use SELDI-TOF to identify colon cancers?
Early imaging
endpoints to assess
outcome and revise treatments
We would like to use
advanced imaging technologies, such as functional magnetic resonance
imaging (fMRI) and V/Q SPECT, to determine if patients are responding
while radiation treatment is being conducted, so that the treatment can
be changed accordingly. We are also interested in determining
if
radiation is opening the blood-tumor barrier, since it may be possible
to open the blood-tumor barrier while not opening the blood-brain
barrier, thus allowing selective administration of large molecules like
Gemcitabine to tumors, while protecting normal brain. Up to
now,
however, most clinical analyses of these technologies involve looking
at the images; Tim Johnson of the Department of Biostatistics &
I
want to make these analyses quantitative and objective by means of a
Markov Random Field model.
Hierarchical
Bayesian models
for biomarker outcomes
We imaged esophageal biopsies from
patients in the trial of DFMO for patients with Barrett's Esophagus to
obtain an index of Ki-67, a proliferation marker. Multiple
images
were available from each slide, multiple slides from each biopsy,
multiple biopsies from each procedure, and multiple procedures from
each subject, who were randomized to DFMO or placebo. I use a
deeply hierarchical Bayesian model to try and capture what is going on
. It isn't pretty.
Data
safety and monitoring
boards as education tools for statisticians
Most clinical
investigators view data safety and monitoring boards (DSMBs) as, at
best, a necessary evil. Many would also find them
superfluous for Phase I and II trials, but, since the Department of
Radiation Oncology is conducting Phase I dose-escalation trials on
chemoradiotherapies with potentially lethal side effects, we have
instituted monthly DSMB meetings for all our Phase I and Phase II
trials. While they may be inconvenient for the PIs, properly
conducted, regular DSMB meetings bring transparency and rigor to
decision-making, improve communication between clinicians, data
managers and statisticians, and are highly educational for
statisticians.