Biology Reference
In-Depth Information
Validation with training and validation sets, independent
sets, and samples from multicenter prospective studies
The most accepted statistical approach for the validation of biomarkers is
to randomly divide patients into training and validation sets (usually 2:1).
The statistical predictive model is developed with the training set and sub-
sequently tested in the validation set, representing therefore a blinded mea-
sure of biomarker performance. This statistical approach was performed in
our first biomarker panel
[12]
. Because of potential center effects, the bio-
markers must also be tested in independent sets from other centers
[22]
.
Finally, biomarkers must be validated in multicenter prospective samples,
such as what we recently did for predicting responses to treatment and
mortality post-treatment by measuring GVHD biomarker concentrations
from samples prospectively obtained on the initiation of treatment, day 14,
and day 28 in a multicenter, randomized, four-arm phase II clinical trial for
newly diagnosed aGVHD
[41]
.
Prognostic markers and risk stratification
Prognostic markers are used to predict an individual's risk for a future
event, such as 3-year post-HSCT death not due to relapse (3-year NRM). In
this context, the key issue is to identify subjects at high and low risk for the
event and to quantify the information for the biomarker that is pertinent to
such a prediction
[12]
.
459
Predictive model
If sufficient proteins are validated as diagnostic biomarkers, the next logical
step is to ask whether these biomarkers can be used to predict the future
development of GVHD before clinical signs become manifest, NRM, or OS.
In this case, we would expect that the biomarkers will correlate with sub-
clinical disease. This possibility is likely, given that plasma tumor necrosis
factor receptor 1 (TNFR1) levels on day 7 post-allogeneic transplantation
correlate with both the eventual severity and the incidence of aGVHD and
OS
[42]
. A common objective in such an observational study in which lon-
gitudinal biomarkers may be highly associated with a time to event, such as
GVHD occurrence, NRM, or OS, is to characterize the relationship between
the longitudinal marker and the time to event. Biomarkers will be systemati-
cally measured at the time points proposed above [day 14 (7 to 21 days prior
to the onset of aGVHD) and day 21 post-HSCT]. Next, a predictive model
using a Cox proportional hazard analysis with time-dependent covariates
must be developed. Time to aGVHD development is a time-to-event out-
come; protein biomarkers measured on days 14 and 21 post-HSCT are also
time-dependent covariates; and other clinical and demographic variables
are time-independent covariates. Two strategies can be used to model time-
dependent biomarkers. One strategy is a simple linear interpolation and
extrapolation of biomarker measurement to days other than days 14, 21,
and 28. The second strategy is to model biomarkers using either a linear or
a quadratic trend. Although these two strategies pose different Cox regres-
sion analyses with time-dependent covariates, they can be implemented
into the likelihood model framework proposed by Wulfsohn and Tsiatis
[43]
. Indeed, considerable interest has recently focused on so-called joint
Search WWH ::
Custom Search