Translating Biomarker Discovery into Companion Diagnostics through Validation and Regulatory Consideration - Genomic Biomarkers for Pharmaceutical Development - page 168

numbers of subject samples that can be evaluated in a single run. If throughput is an impor-
tant factor to the device, then higher numbers of analytes may bring the feasible use of the
diagnostic into question. In addition, the numbers of analytes in an assay often determine the
price of an assay. This is due to the code stacking method of reimbursement applied to newly
approved devices. Manufacturers of the device typically will take existing current procedural
terminology (CPT) codes and use them multiple times to account for each reaction that is tak-
ing place. The use of putting the same codes down to an account multiple times for the total
cost of a procedure is termed code stacking. As described in a recent American Association of
Clinical Chemistry article:
'Unlike traditional assays, molecular tests do not have single, analyte-specific codes that labs can use to
bill Medicare or private payers. Rather, a list of codes that signify each procedure involved in performing the
assay are listed together. Each code describes a separate step or methodology performed to complete the test,
such as gene amplification, nucleic acid extraction, or nucleic acid probes. For some tests, these codes must
also be multiplied if the laboratory uses a step more than once to perform the test, especially in those tests
that look at multiple markers. As an example, if a PCR-based assay evaluates seven targets and three house-
keeping genes to determine a classifier, then codes for molecular extraction, nucleic acid transfer, and nucleic
acid PCR at a minimum would be stacked for each of the 10 reactions per sample. Changes to the coding
scheme have been made recently that are aimed at ending code stacking and separating molecular tests into
two major tiers, adding them to the physician fee schedule and away from the clinical laboratory fee sched-
ule. However, implementation of this coding method has yet to be fully adopted due to a number of complex
issues surrounding reduced reimbursement for clinical laboratories [17] .'
Regardless of the coding method being employed, it is important to understand the inter-
play between the number of analytes in a device and cost of the device.
7.4.2.10 Multiple Analyte Versus Few Analyte Classifiers
A very common concern in biomarker classifier development is the large number of ana-
lytes measured for a given patient. This issue has become almost commonplace in the last 15
years with the introduction of microarrays used to measure thousands of transcripts, to mul-
tiplex protein and cellular assays, to the more recent use of high-throughput sequencing. The
number of analytes or variables can far exceed millions for a given patient specimen. This
type of situation is known as the 'curse of dimensionality', where an estimator will converge
to the true value of a smooth function on a high-dimensional space very slowly [18,19] . Or
more specific to a biological context, this expression means that in order to obtain a good esti-
mate for a function of analytes to indicate a predefined expression pattern, a very large num-
ber of patient specimens would be required [20] . However, the ability to procure such large
numbers of specimens is often hindered by either cost or availability. There are factors that
can often reduce the initial set of analytes to a smaller set of useable components for a clas-
sifier, and various statistical methods have been developed to summarize multiple analytes
into a low-dimensional function that can be utilized to elucidate meaningful results.
For practical implementation, fewer analytes in a classifier can be argued as preferable to
a large panel of analytes. This concept follows from the issue of dimensionality explained
previously, where if one were to attempt to partition p patient specimens into two groups
(e.g., tumor vs. normal tissue) based on g genes, where g >> p , it is very likely that a clas-
sifier based on some set of genes would be capable of doing so into the appropriate two
groups with 100% accuracy. This result, however, would probably not be generalized enough